The Alzheimer's Disease Prediction Of Longitudinal Evolution (TADPOLE) Challenge: Results after 1 Year Follow-up

Razvan V. Marinescu1,2Orcid, Neil P. Oxtoby1Orcid, Alexandra L. Young1,3, Esther E. Bron4Orcid, Arthur W. Toga5Orcid, Michael W. Weiner6Orcid, Frederik Barkhof1,7,8Orcid, Nick C. Fox8Orcid, Arman Eshaghi9,1Orcid, Tina Toni10, Marcin Salaterski10, Veronika Lunina10, Manon Ansart11, Stanley Durrleman11, Pascal Lu11, Samuel Iddi12,13, Dan Li12, Wesley K. Thompson14, Michael C. Donohue12, Aviv Nahon15, Yarden Levy15, Dan Halbersberg15, Mariya Cohen15, Huiling Liao16, Tengfei Li16, Kaixian Yu16, Hongtu Zhu16, José G. Tamez-Peña17, Aya Ismail18, Timothy Wood18, Hector Corrada Bravo18, Minh Nguyen19, Nanbo Sun19, Jiashi Feng19, B.T. Thomas Yeo19, Gang Chen20, Ke Qi21, Shiyang Chen21,22, Deqiang Qiu21,22, Ionut Buciuman23, Alex Kelner23, Raluca Pop23, Denisa Rimocea23, Mostafa M. Ghazi24,25,26,1, Mads Nielsen24,25,26, Sebastien Ourselin27,1, Lauge Sørensen24,25,26, Vikram Venkatraghavan4, Keli Liu28, Christina Rabe28, Paul Manser28, Steven M. Hill29, James Howlett29, Zhiyue Huang29, Steven Kiddle29, Sach Mukherjee30, Anaïs Rouanet29, Bernd Taschler30, Brian D. M. Tom29, Simon R. White29, Noel Faux31, Suman Sedai31, Javier de Velasco Oriol17, Edgar E. V. Clemente17, Karol Estrada32,33, Leon Aksman1, Andre Altmann1, Cynthia M. Stonnington34, Yalin Wang35, Jianfeng Wu35, Vivek Devadas36, Clementine Fourrier11, Lars Lau Raket37,38, Aristeidis Sotiras39, Guray Erus39, Jimit Doshi39, Christos Davatzikos39, Jacob Vogel40, Andrew Doyle40, Angela Tam40, Alex Diaz-Papkovich40, Emmanuel Jammeh41, Igor Koval11, Paul Moore42, Terry J. Lyons42, John Gallacher43, Jussi Tohka44, Robert Ciszek44, Bruno Jedynak45, Kruti Pandya45, Murat Bilgel46, William Engels45, Joseph Cole45, Polina Golland2, Stefan Klein4Orcid, Daniel C. Alexander1Orcid, The EuroPOND Consortium 10, The Alzheimer's Disease Neuroimaging Initiative 10
1: Centre for Medical Image Computing, University College London, UK, 2: Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, USA, 3: Department of Neuroimaging, Institute of Psychiatry, Psychology and Neuroscience, King's College London, London, UK, 4: Biomedical Imaging Group Rotterdam, Department of Radiology and Nuclear Medicine, Erasmus MC, Netherlands, 5: Laboratory of NeuroImaging, University of Southern California, USA, 6: Center for Imaging of Neurodegenerative Diseases, University of California San Francisco, USA, 7: Department of Radiology and Nuclear Medicine, VU Medical Centre, Netherlands, 8: Dementia Research Centre and the UK Dementia Research Institute, UCL Queen Square Institute of Neurology, UK, 9: Queen Square Multiple Sclerosis Centre, UCL Queen Square Institute of Neurology, UK, 10: Author not affiliated with any research institution, 11: Institut du Cerveau et de la Moelle épinière, Paris, France, 12: Alzheimer's Therapeutic Research Institute, University of Southern California, USA, 13: Department of Statistics and Actuarial Science, University of Ghana, Ghana, 14: Department of Family Medicine and Public Health, University of California San Diego, USA, 15: Ben Gurion University of the Negev, Beersheba, Israel, 16: The University of Texas Health Science Center at Houston, Houston, USA, 17: Instituto Tecnológico y de Estudios Superiores de Monterrey, Monterrey, Mexico, 18: University of Maryland, College Park, USA, 19: National University of Singapore, Singapore, Singapore, 20: Medical College of Wisconsin, Milwaukee, USA, 21: Emory University, Atlanta, USA, 22: Georgia Institute of Technology, Atlanta, USA, 23: Vasile Lucaciu National College, Baia Mare, Romania, 24: Biomediq A/S, Denmark, 25: Cerebriu A/S, Denmark, 26: University of Copenhagen, Denmark, 27: School of Biomedical Engineering and Imaging Sciences, King's College London, UK, 28: Genentech, USA, 29: MRC Biostatistics Unit, University of Cambridge, UK, 30: German Center for Neurodegenerative Diseases, Bonn, Germany, 31: IBM Research Australia, Melbourne, Australia, 32: Brandeis University, Waltham, USA, 33: Department of Statistical Genetics, Biomarin, San Rafael, USA, 34: Mayo Clinic, Scottsdale, AZ, USA, 35: School of Computing, Informatics and Decision Systems Engineering, Arizona State University, Tempe, USA, 36: Banner Alzheimer's Institute, Phoenix, USA, 37: H. Lundbeck A/S, Denmark, 38: Clinical Memory Research Unit, Department of Clinical Sciences Malmo, Lund University, Lund, Sweden, 39: Center for Biomedical Image Computing and Analytics, University of Pennsylvania, 40: McGill University, Montreal, Canada, 41: University of Plymouth, UK, 42: Mathematical Institute, University of Oxford, UK, 43: Department of Psychiatry, University of Oxford, UK, 44: A.I. Virtanen Institute for Molecular Sciences, University of Eastern Finland, Finland, 45: Portland State University, Portland, USA, 46: Laboratory of Behavioral Neuroscience, National Institute on Aging, National Institutes of Health, Baltimore, MD, USA
Publication date: 2021/12/31
https://doi.org/10.59275/j.melba.2021-2dcc
PDF · Source code · TADPOLE-SHARE · Video · Website · arXiv

Abstract

Accurate prediction of progression in subjects at risk of Alzheimer's disease is crucial for enrolling the right subjects in clinical trials. However, a prospective comparison of state-of-the-art algorithms for predicting disease onset and progression is currently lacking. We present the findings of "The Alzheimer's Disease Prediction Of Longitudinal Evolution" (TADPOLE) Challenge, which compared the performance of 92 algorithms from 33 international teams at predicting the future trajectory of 219 individuals at risk of Alzheimer's disease. Challenge participants were required to make a prediction, for each month of a 5-year future time period, of three key outcomes: clinical diagnosis, Alzheimer's Disease Assessment Scale Cognitive Subdomain (ADAS-Cog13), and total volume of the ventricles. The methods used by challenge participants included multivariate linear regression, machine learning methods such as support vector machines and deep neural networks, as well as disease progression models. No single submission was best at predicting all three outcomes. For clinical diagnosis and ventricle volume prediction, the best algorithms strongly outperform simple baselines in predictive ability. However, for ADAS-Cog13 no single submitted prediction method was significantly better than random guesswork. Two ensemble methods based on taking the mean and median over all predictions, obtained top scores on almost all tasks. Better than average performance at diagnosis prediction was generally associated with the additional inclusion of features from cerebrospinal fluid (CSF) samples and diffusion tensor imaging (DTI). On the other hand, better performance at ventricle volume prediction was associated with inclusion of summary statistics, such as the slope or maxima/minima of patient-specific biomarkers. On a limited, cross-sectional subset of the data emulating clinical trials, performance of the best algorithms at predicting clinical diagnosis decreased only slightly (2 percentage points) compared to the full longitudinal dataset. The submission system remains open via the website https://tadpole.grand-challenge.org, while TADPOLE SHARE (https://tadpole-share.github.io/) collates code for submissions. TADPOLE's unique results suggest that current prediction algorithms provide sufficient accuracy to exploit biomarkers related to clinical diagnosis and ventricle volume, for cohort refinement in clinical trials for Alzheimer's disease. However, results call into question the usage of cognitive test scores for patient selection and as a primary endpoint in clinical trials.

Keywords

statistical modelling · machine learning · benchmark · alzheimer's disease prediction

Bibtex @article{melba:2021:019:marinescu, title = "The Alzheimer's Disease Prediction Of Longitudinal Evolution (TADPOLE) Challenge: Results after 1 Year Follow-up", author = "Marinescu, Razvan V. and Oxtoby, Neil P. and Young, Alexandra L. and Bron, Esther E. and Toga, Arthur W. and Weiner, Michael W. and Barkhof, Frederik and Fox, Nick C. and Eshaghi, Arman and Toni, Tina and Salaterski, Marcin and Lunina, Veronika and Ansart, Manon and Durrleman, Stanley and Lu, Pascal and Iddi, Samuel and Li, Dan and Thompson, Wesley K. and Donohue, Michael C. and Nahon, Aviv and Levy, Yarden and Halbersberg, Dan and Cohen, Mariya and Liao, Huiling and Li, Tengfei and Yu, Kaixian and Zhu, Hongtu and Tamez-Peña, José G. and Ismail, Aya and Wood, Timothy and Bravo, Hector Corrada and Nguyen, Minh and Sun, Nanbo and Feng, Jiashi and Yeo, B.T. Thomas and Chen, Gang and Qi, Ke and Chen, Shiyang and Qiu, Deqiang and Buciuman, Ionut and Kelner, Alex and Pop, Raluca and Rimocea, Denisa and Ghazi, Mostafa M. and Nielsen, Mads and Ourselin, Sebastien and Sørensen, Lauge and Venkatraghavan, Vikram and Liu, Keli and Rabe, Christina and Manser, Paul and Hill, Steven M. and Howlett, James and Huang, Zhiyue and Kiddle, Steven and Mukherjee, Sach and Rouanet, Anaïs and Taschler, Bernd and Tom, Brian D. M. and White, Simon R. and Faux, Noel and Sedai, Suman and de Velasco Oriol, Javier and Clemente, Edgar E. V. and Estrada, Karol and Aksman, Leon and Altmann, Andre and Stonnington, Cynthia M. and Wang, Yalin and Wu, Jianfeng and Devadas, Vivek and Fourrier, Clementine and Raket, Lars Lau and Sotiras, Aristeidis and Erus, Guray and Doshi, Jimit and Davatzikos, Christos and Vogel, Jacob and Doyle, Andrew and Tam, Angela and Diaz-Papkovich, Alex and Jammeh, Emmanuel and Koval, Igor and Moore, Paul and Lyons, Terry J. and Gallacher, John and Tohka, Jussi and Ciszek, Robert and Jedynak, Bruno and Pandya, Kruti and Bilgel, Murat and Engels, William and Cole, Joseph and Golland, Polina and Klein, Stefan and Alexander, Daniel C. and , The EuroPOND Consortium and , The Alzheimer's Disease Neuroimaging Initiative", journal = "Machine Learning for Biomedical Imaging", volume = "1", issue = "December 2021 issue", year = "2021", pages = "1--60", issn = "2766-905X", doi = "https://doi.org/10.59275/j.melba.2021-2dcc", url = "https://melba-journal.org/2021:019" }
RISTY - JOUR AU - Marinescu, Razvan V. AU - Oxtoby, Neil P. AU - Young, Alexandra L. AU - Bron, Esther E. AU - Toga, Arthur W. AU - Weiner, Michael W. AU - Barkhof, Frederik AU - Fox, Nick C. AU - Eshaghi, Arman AU - Toni, Tina AU - Salaterski, Marcin AU - Lunina, Veronika AU - Ansart, Manon AU - Durrleman, Stanley AU - Lu, Pascal AU - Iddi, Samuel AU - Li, Dan AU - Thompson, Wesley K. AU - Donohue, Michael C. AU - Nahon, Aviv AU - Levy, Yarden AU - Halbersberg, Dan AU - Cohen, Mariya AU - Liao, Huiling AU - Li, Tengfei AU - Yu, Kaixian AU - Zhu, Hongtu AU - Tamez-Peña, José G. AU - Ismail, Aya AU - Wood, Timothy AU - Bravo, Hector Corrada AU - Nguyen, Minh AU - Sun, Nanbo AU - Feng, Jiashi AU - Yeo, B.T. Thomas AU - Chen, Gang AU - Qi, Ke AU - Chen, Shiyang AU - Qiu, Deqiang AU - Buciuman, Ionut AU - Kelner, Alex AU - Pop, Raluca AU - Rimocea, Denisa AU - Ghazi, Mostafa M. AU - Nielsen, Mads AU - Ourselin, Sebastien AU - Sørensen, Lauge AU - Venkatraghavan, Vikram AU - Liu, Keli AU - Rabe, Christina AU - Manser, Paul AU - Hill, Steven M. AU - Howlett, James AU - Huang, Zhiyue AU - Kiddle, Steven AU - Mukherjee, Sach AU - Rouanet, Anaïs AU - Taschler, Bernd AU - Tom, Brian D. M. AU - White, Simon R. AU - Faux, Noel AU - Sedai, Suman AU - de Velasco Oriol, Javier AU - Clemente, Edgar E. V. AU - Estrada, Karol AU - Aksman, Leon AU - Altmann, Andre AU - Stonnington, Cynthia M. AU - Wang, Yalin AU - Wu, Jianfeng AU - Devadas, Vivek AU - Fourrier, Clementine AU - Raket, Lars Lau AU - Sotiras, Aristeidis AU - Erus, Guray AU - Doshi, Jimit AU - Davatzikos, Christos AU - Vogel, Jacob AU - Doyle, Andrew AU - Tam, Angela AU - Diaz-Papkovich, Alex AU - Jammeh, Emmanuel AU - Koval, Igor AU - Moore, Paul AU - Lyons, Terry J. AU - Gallacher, John AU - Tohka, Jussi AU - Ciszek, Robert AU - Jedynak, Bruno AU - Pandya, Kruti AU - Bilgel, Murat AU - Engels, William AU - Cole, Joseph AU - Golland, Polina AU - Klein, Stefan AU - Alexander, Daniel C. AU - , The EuroPOND Consortium AU - , The Alzheimer's Disease Neuroimaging Initiative PY - 2021 TI - The Alzheimer's Disease Prediction Of Longitudinal Evolution (TADPOLE) Challenge: Results after 1 Year Follow-up T2 - Machine Learning for Biomedical Imaging VL - 1 IS - December 2021 issue SP - 1 EP - 60 SN - 2766-905X DO - https://doi.org/10.59275/j.melba.2021-2dcc UR - https://melba-journal.org/2021:019 ER -

2021:019 cover


Disclaimer: the following html version has been automatically generated and the PDF remains the reference version. Feedback can be sent directly to publishing-editor@melba-journal.org

Abstract

Accurate prediction of progression in subjects at risk of Alzheimer’s disease is crucial for enrolling the right subjects in clinical trials. However, a prospective comparison of state-of-the-art algorithms for predicting disease onset and progression is currently lacking. We present the findings of The Alzheimer’s Disease Prediction Of Longitudinal Evolution (TADPOLE) Challenge, which compared the performance of 92 algorithms from 33 international teams at predicting the future trajectory of 219 individuals at risk of Alzheimer’s disease. Challenge participants were required to make a prediction, for each month of a 5-year future time period, of three key outcomes: clinical diagnosis, Alzheimer’s Disease Assessment Scale Cognitive Subdomain (ADAS-Cog13), and total volume of the ventricles. No single submission was best at predicting all three outcomes. For clinical diagnosis and ventricle volume prediction, the best algorithms strongly outperform simple baselines in predictive ability. However, for ADAS-Cog13 no single submitted prediction method was significantly better than random guessing. On a limited, cross-sectional subset of the data emulating clinical trials, performance of best algorithms at predicting clinical diagnosis decreased only slightly (3% error increase) compared to the full longitudinal dataset. Two ensemble methods based on taking the mean and median over all predictions, obtained top scores on almost all tasks. Better than average performance at diagnosis prediction was generally associated with the additional inclusion of features from cerebrospinal fluid (CSF) samples and diffusion tensor imaging (DTI). On the other hand, better performance at ventricle volume prediction was associated with inclusion of summary statistics, such as patient-specific biomarker trends. The submission system remains open via the website https://tadpole.grand-challenge.org, while code for submissions is being collated by TADPOLE SHARE: https://tadpole-share.github.io/. Our work suggests that current prediction algorithms are accurate for biomarkers related to clinical diagnosis and ventricle volume, opening up the possibility of cohort refinement in clinical trials for Alzheimer’s disease.

1 Introduction

Accurate prediction of the onset of Alzheimer’s disease (AD) and its longitudinal progression is important for care planning and for patient selection in clinical trials. Current opinion holds that early detection will be critical for the successful administration of disease modifying treatments during presymptomatic phases of the disease prior to widespread brain damage, e.g. when pathological amyloid and tau start to accumulate mehta2017trials . Moreover, accurate prediction of the progression of at-risk subjects will help select homogenous patient groups for clinical trials, thus reducing variability in outcome measures that can obscure positive effects on patients at the right stage to benefit.

Several mathematical and computational methods have been developed to predict the onset and progression of AD. Traditional approaches leverage statistical regression to model relationships between target variables (e.g. clinical diagnosis or cognitive/imaging markers) with other known markers scahill2002mapping ; sabuncu2011dynamics or measures derived from these markers such as the rate of cognitive decline doody2010erratum . More recent approaches involve supervised machine learning techniques such as support vector machines, random forests and artificial neural networks. These approaches have been used to discriminate AD patients from cognitively normal individuals kloppel2008automatic ; zhang2011multimodal , and for discriminating at-risk individuals who convert to AD in a certain time frame from those who do not young2013accurate ; mattila2011disease . The emerging approach of disease progression modelling aims to reconstruct biomarker trajectories or other disease signatures across the disease progression timeline, without relying on clinical diagnoses or estimates of time to symptom onset. Examples include models built on a set of scalar biomarkers to produce discrete fonteijn2012event ; young2014data or continuous jedynak2012computational ; donohue2014estimating ; lorenzi2017probabilistic ; oxtoby2018data ; schiratti2017bayesian biomarker trajectories; spatio-temporal models that focus on evolving image structure bilgel2016multivariate ; marinescu2019dive , potentially conditioned by non-imaging variables koval2018simulating ; and models that emulate putative disease mechanisms to estimate trajectories of change raj2012network ; iturria2016early ; zhou2012predicting . All these models show promise for predicting AD biomarker progression at group and individual levels. However, previous evaluations within individual publications provide limited information because: (1) they use different data sets or subsets of the same dataset, different processing pipelines, and different evaluation metrics and (2) over-training can occur due to heavy use of popular training datasets. Currently, the field lacks a comprehensive comparison of the capabilities of these methods on standardised tasks relevant to real-world applications.

Community challenges have consistently proved effective in moving forward the state of the art in technology to address specific data-analysis problems by providing platforms for unbiased comparative evaluation and incentives to maximise performance on key tasks maier2018rankings . In medical image analysis, for example, such challenges have provided important benchmarks in tasks such as registration murphy2011evaluation and segmentation menze2014multimodal , and revealed fundamental insights about the problem studied, for example in structural brain-connectivity mapping maier2017challenge . Previous challenges in AD include the CADDementia challenge bron2015standardized , which aimed to identify clinical diagnosis from MRI scans. A similar challenge, the International challenge for automated prediction of MCI from MRI data castiglioni2018machine , asked participants to predict diagnosis and conversion status from extracted MRI features of subjects from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) study weiner2017recent . Yet another challenge, The Alzheimer’s Disease Big Data DREAM Challenge allen2016crowdsourced , asked participants to predict cognitive decline from genetic and MRI data. These challenges have however several limitations: (i) they did not evaluate the ability of algorithms to predict biomarkers at future timepoints (with the exception of one sub-challenge of DREAM), which is important for patient stratification in clinical trials; (ii) the test data was available to organisers when the competitions were launched, leaving room for potential biases in the design of the challenges; (iii) the training data was drawn from a limited set of modalities.

The Alzheimer’s Disease Prediction Of Longitudinal Evolution (TADPOLE) Challenge (https://tadpole.grand-challenge.org) aims to identify the data, features and approaches that are the most predictive of future progression of subjects at risk of AD. In contrast to previous challenges, our challenge is designed to inform clinical trials through identification of patients most likely to benefit from an effective treatment, i.e., those at early stages of disease who are likely to progress over the short-to-medium term (1-5 years). The challenge focuses on forecasting the trajectories of three key features: clinical status, cognitive decline, and neurodegeneration (brain atrophy), over a five-year timescale. It uses “rollover” subjects from the ADNI study weiner2017recent for whom a history of measurements (imaging, psychology, demographics, genetics) is available, and who are expected to continue in the study, providing future measurements for testing. TADPOLE participants were required to predict future measurements from these individuals and submit their predictions before a given submission deadline. Since the test data did not exist at the time of forecast submissions, the challenge provides a performance comparison substantially less susceptible to many forms of potential bias than previous studies and challenges. The design choices were published marinescu2018tadpole before the test set was acquired and analysed. TADPOLE also goes beyond previous challenges by drawing on a vast set of multimodal measurements from ADNI which might support prediction of AD progression.

This article presents the results of the TADPOLE Challenge and documents its key findings. We summarise the challenge design and present the results of the 92 prediction algorithms contributed by 33 participating teams worldwide, evaluated after an 18-month follow-up period. We discuss the results obtained by TADPOLE participants, which represent the current state-of-the-art in Alzheimer’s disease prediction. To understand what key characteristics of algorithms were important for good predictions, we also report results on which input data features were most informative, and which feature selection strategies, data imputation methods and classes of algorithms were most effective.

2 Results

2.1 Predictions

TADPOLE Challenge asked participants to forecast three key biomarkers: (1) clinical diagnosis, which can be either cognitively normal (CN), mild cognitive impairment (MCI), or probable AD; (2) Alzheimer’s Disease Assessment Scale Cognitive Subdomain (ADAS-Cog13) score; and (3) ventricle volume (divided by intra-cranial volume) from MRI. The exact time of future data acquisitions for any given individual was unknown at forecast time, so participants submitted month-by-month predictions for every individual. Predictions of clinical status comprise relative likelihoods of each option (CN, MCI, and AD) for each individual at each month. Predictions of ADAS-Cog13 and ventricle volume comprise a best-guess estimate as well as a 50% confidence interval for each individual at each month. Full details on challenge design are given in the TADPOLE white paper marinescu2018tadpole .

2.2 Algorithms

SubmissionFeature selectionNumber of featuresMissing data imputationDiagnosis prediction modelADAS/Vent. prediction modelTraining timePrediction time (one subject)
AlgosForGood   manual16+5*forward-fillingAalen modellinear regression1 min.1 sec.
Apocalypse   manual16population averageSVMlinear regression40 min.3 min.
ARAMIS-Pascal   manual20population averageAalen model-16 sec.0.02 sec.
ATRI-Biostat-JMM   automatic15random forestrandom forestlinear mixed effects model2 days1 sec.
ATRI-Biostat-LTJMM   automatic15random forestrandom forestDPM2 days1 sec.
ATRI-Biostat-MA   automatic15random forestrandom forestDPM + linear mixed effects model2 days1 sec.
BGU-LSTM   automatic67nonefeed-forward NNLSTM1 daymillisec.
BGU-RF/ BGU-RFFIX   automatic\approx 67+1340*nonesemi-temporal RFsemi-temporal RFa few min.millisec.
BIGS2   automaticallIterative Thresholded SVDRFlinear regression2.2 sec.0.001 sec.
Billabong (all)   manual15-16linear regressionlinear scalenon-parametric SM7 hours0.13 sec.
BORREGOSTECMTY   automatic\approx100 + 400*nearest-neighbourregression ensembleensemble of regression + hazard models18 hours0.001 sec.
BravoLab   automatic25hot deckLSTMLSTM1 houra few sec.
CBIL   manual21linear interpolationLSTMLSTM1 hourone min.
Chen-MCW   manual9nonelinear regressionDPM4 hours<< 1 hour
CN2L-NeuralNetwork   automaticallforward-fillingRNNRNN24 hoursa few sec.
CN2L-RandomForest   manual>>200forward-fillingRFRF15 min.<< 1 min.
CN2L-Average   automaticallforward-fillingRNN/RFRNN/RF24 hours<< 1 min.
CyberBrains   manual5population averagelinear regressionlinear regression20 sec.20 sec.
DIKU (all)   semi-automatic18noneBayesian classifier/LDA + DPMDPM290 sec.0.025 sec.
DIVE   manual13noneKDE+DPMDPM20 min.0.06 sec.
EMC1   automatic250nearest neighbourDPM + 2D spline + SVMDPM + 2D spline80 min.a few sec.
EMC-EB   automatic200-338nearest-neighbourSVM classifierSVM regressor20 sec.a few sec.
FortuneTellerFish-Control   manual19nearest neighbourmulticlass ECOC SVMlinear mixed effects model1 min.<< 1 sec.
FortuneTellerFish-SuStaIn   manual19nearest neighbourmulticlass ECOC SVM + DPMlinear mixed effects model + DPM5 hours<< 1 sec.
Frog   automatic\approx 70+420*nonegradient boostinggradient boosting1 hour-
GlassFrog-LCMEM-HDR   semi-automaticallforward-fill/nearest-neigh.multi-state modelDPM + regression15 min.2 min.
GlassFrog-SM   manual7linear modelmulti-state modelparametric SM93 sec.0.1 sec.
GlassFrog-Average   semi-automaticallforward-fill/linearmulti-state modelDPM + SM + regression15 min.2 min.
IBM-OZ-Res   manualOct-15filled with zerostochastic gradient boostingstochastic gradient boosting20 min.0.1 sec.
ITESMCEM   manual48mean of previous valuesRFLASSO + Bayesian ridge regression20 min.0.3 sec.
lmaUCL (all)   manual5regressionmulti-task learningmulti-task learning2 hoursmillisec.
Mayo-BAI-ASU   manual15population averagelinear mixed effects modellinear mixed effects model20 min.1.3 sec.
Orange   manual17noneclinician’s decision treeclinician’s decision treenone0.2 sec.
Rocket   manual6median of diagnostic grouplinear mixed effects modelDPM5 min.0.3 sec.
SBIA   manual30-70dropped visits with missing dataSVM + density estimatorlinear mixed effects model1 min.a few sec.
SPMC-Plymouth (all)   automatic20noneunknown-unknown1 min.
SmallHeads-NeuralNetwork   automatic376nearest neighbourdeep fully -connected NNdeep fully -connected NN40 min.0.06 sec.
SmallHeads-LinMixedEffects   automaticunknownnearest neighbour-linear mixed effects model25 min.0.13 sec.
Sunshine (all)   semi-automatic6population averageSVMlinear model30 min.<< 1 min.
Threedays   manual16noneRF-1 min.3 sec.
Tohka-Ciszek-SMNSR   manual\approx 32nearest neighbour-SMNSRseveral hoursa few sec.
Tohka-Ciszek-RandomForestLin   manual32mean patient valueRFlinear modela few min.a few sec.
VikingAI (all)   manual10noneDPM + ordered logit modelDPM10 hours8 sec.
BenchmaskLastVisit   None3noneconstant modelconstant model7 sec.millisec.
BenchmarkMixedEffects   None3noneGaussian modellinear mixed effects model30 sec.0.003 sec.
BenchmarkMixedEffects-APOE   None4noneGaussian modellinear mixed effects model30 sec.0.003 sec.
BenchmarkSVM   manual6mean of previous valuesSVMsupport vector regressor (SVR)20 sec.0.001 sec.
Table 1: Summary of prediction methods used in the TADPOLE submissions. Keywords: SVM – Support Vector Machine, RF – random forest, LSTM – long short-term memory network, NN – neural network, RNN – recurrent neural network, SMNSR – Sparse Multimodal Neighbourhood Search Regression, DPM – disease progression model, KDE – kernel density estimation, LDA – linear discriminant analysis, SM – slope model, ECOC – error-correcting output codes, SVD – singular value decomposition (*) Augmented features, or summary statistics, such as trends, slope, min/max, moments, generally derived patient-wise using longitudinal data. Color tags denote prediction method category:   regression/proportional hazards model,   random forest,   neural networks,   disease progression model,   machine learning (other),   benchmark,   other. The left-side box denotes the category for diagnosis prediction method, while the right-side box denotes the category for ADAS/Ventricle prediction method.

We had a total of 33 participating teams, who submitted a total of 58 predictions from the longitudinal prediction set (D2), 34 predictions from the cross-sectional prediction set (D3), and 6 predictions from custom prediction sets (see Online Methods section 5.1 for description of D2/D3 datasets). A total of 8 D2/D3 submissions from 6 teams did not have predictions for all three target variables, so we computed the performance metrics for only the submitted target variables. Another 3 submissions lacked confidence intervals for either ADAS-Cog13 or ventricle volume, which we imputed using default low-width confidence ranges of 2 for ADAS-Cog13 and 0.002 for Ventricles normalised by intracranial volume (ICV).

Table 1 summarises the methods used in the submissions in terms of feature selection, handling of missing data, predictive models for clinical diagnosis and ADAS/Ventricles biomarkers, as well as training and prediction times. A detailed description of each method is in Online Methods Section 5.4. In particular, some entries constructed augmented features (i.e. summary statistics), which are extra features such as slope, min/max or moments that are derived from existing features.

In addition to the forecasts submitted by participants, we also evaluated four benchmark methods, which were made available to participants during the submission phase of the challenge: (i) BenchmaskLastVisit uses the measurement of each target from the last available clinical visit as the forecast, (ii) BenchmarkMixedEffects uses a mixed effects model with age as predictor variable for ADAS and Ventricle predictions, and Gaussian likelihood model for diagnosis prediction, (iii) BenchmarkMixedEffectsAPOE is as (ii) but adds APOE status as a covariate and (iv) BenchmarkSVM uses an out-of-the-box support vector machine (SVM) classifier and regressor (SVR) to provide forecasts. More details on these methods can be found in Online Methods section 5.4. We also evaluated two ensemble methods based on taking the mean (ConsensusMean) and median (ConsensusMedian) of the forecasted variables over all submissions.

To control for potentially spurious strong performance arising from multiple comparisons, we also evaluated 100 random predictions by adding Gaussian noise to the forecasts of the simplest benchmark model (BenchmarkLastVisit). In the subsequent results tables we will show, for each performance metric, only the best score obtained by any of these 100 random predictions (RandomisedBest) – See end of Online Methods section 5.4 for more information on RandomisedBest.

2.3 Forecasts from the longitudinal prediction set (D2)

OverallDiagnosisADAS-Cog13Ventricles (% ICV)
SubmissionRankRankMAUCBCARankMAEWESCPARankMAEWESCPA
ConsensusMedian   --0.9250.857-5.125.010.28-0.380.330.09
Frog   110.9310.84944.854.740.44100.450.330.47
ConsensusMean   --0.9200.835-3.753.540.00-0.480.450.13
EMC1-Std   280.8980.81123-246.055.400.451-20.410.290.43
VikingAI-Sigmoid   3160.8750.76075.205.110.0211-120.450.350.20
EMC1-Custom   4110.8920.79823-246.055.400.451-20.410.290.43
CBIL   590.8970.803155.665.650.37130.460.460.09
Apocalypse   670.9020.827145.575.570.50200.520.520.50
GlassFrog-Average   74-60.9020.82585.265.270.26290.680.600.33
GlassFrog-SM   84-60.9020.825175.775.920.20210.520.330.20
BORREGOTECMTY   9190.8660.808205.905.820.3950.430.370.40
BenchmarkMixedEffects   --0.8460.706-4.194.190.31-0.560.560.50
EMC-EB   1030.9070.805396.756.660.5090.450.400.48
lmaUCL-Covariates   11-12220.8520.760276.286.290.2830.420.410.11
CN2L-Average   11-12270.8430.79295.315.310.35160.490.490.33
VikingAI-Logistic   13200.8650.754216.025.910.2611-120.450.350.20
lmaUCL-Std   14210.8590.781286.306.330.2640.420.410.09
RandomisedBest   --0.8000.803-4.524.520.27-0.460.450.33
CN2L-RandomForest   15-16100.8960.792165.735.730.42310.710.710.41
FortuneTellerFish-SuStaIn   15-16400.8060.68534.814.810.21140.490.490.18
CN2L-NeuralNetwork   17410.7830.717105.365.360.3470.440.440.27
BenchmarkMixedEffectsAPOE   18350.8220.74924.754.750.36230.570.570.40
Tohka-Ciszek-RandomForestLin   19170.8750.796226.036.030.15220.560.560.37
BGU-LSTM   20120.8830.779256.096.120.39250.600.600.23
DIKU-GeneralisedLog-Custom   21130.8780.79011-125.405.400.2638-391.051.050.05
DIKU-GeneralisedLog-Std   22140.8770.79011-125.405.400.2638-391.051.050.05
CyberBrains   23340.8230.74765.165.160.24260.620.620.12
AlgosForGood   24240.8470.810135.465.110.13300.693.310.19
lmaUCL-halfD1   25260.8450.753386.536.510.3160.440.420.13
BGU-RF   26280.8380.67329-306.336.100.3517-180.500.380.26
Mayo-BAI-ASU   27520.6910.62454.984.980.32190.520.520.40
BGU-RFFIX   28320.8310.67329-306.336.100.3517-180.500.380.26
FortuneTellerFish-Control   29310.8340.69214.704.700.22501.381.380.50
GlassFrog-LCMEM-HDR   304-60.9020.825316.346.210.47511.661.590.41
SBIA   31430.7760.721437.107.380.4080.440.310.13
Chen-MCW-Stratify   32230.8480.78336-376.486.240.2336-371.011.000.11
Rocket   33540.6800.519185.815.710.34280.640.640.29
BenchmarkSVM   34-35300.8360.764406.826.820.42320.860.840.50
Chen-MCW-Std   34-35290.8360.77836-376.486.240.2336-371.011.000.11
DIKU-ModifiedMri-Custom   3636-370.8070.67032-356.446.440.2734-350.920.920.01
DIKU-ModifiedMri-Std   3738-390.8060.67032-356.446.440.2734-350.920.920.01
DIVE   38510.7080.568427.107.100.34150.490.490.13
ITESMCEM   39530.6800.657266.266.260.35330.920.920.43
BenchmarkLastVisit   4044-450.7740.792417.057.050.45270.630.610.47
Sunshine-Conservative   41250.8450.81644-457.907.900.5043-441.121.120.50
BravoLab   42460.7710.682478.228.220.49240.580.580.41
DIKU-ModifiedLog-Custom   4336-370.8070.67032-356.446.440.2747-481.171.170.06
DIKU-ModifiedLog-Std   4438-390.8060.67032-356.446.440.2747-481.171.170.06
Sunshine-Std   45330.8250.77144-457.907.900.5043-441.121.120.50
Billabong-UniAV45   46490.7200.61648-499.228.820.2941-421.090.990.45
Billabong-Uni   47500.7180.62248-499.228.820.2941-421.090.990.45
ATRI-Biostat-JMM   48420.7790.7105112.8869.620.35541.955.120.33
Billabong-Multi   49560.5410.5565527.0119.900.46401.071.070.45
ATRI-Biostat-MA   50470.7410.6715212.8811.320.19531.845.270.23
BIGS2   51580.4550.4885011.6214.650.50491.201.120.07
Billabong-MultiAV45   52570.5270.5305628.4521.220.47451.131.070.47
ATRI-Biostat-LTJMM   53550.6360.5635416.0774.650.33521.805.010.26
Threedays   -20.9210.823--------
ARAMIS-Pascal   -150.8760.850--------
IBM-OZ-Res   -180.8680.766----461.151.150.50
Orange   -44-450.7740.792--------
SMALLHEADS-NeuralNet   -480.7370.6055313.8713.870.41----
SMALLHEADS-LinMixedEffects   ----468.097.940.04----
Tohka-Ciszek-SMNSR   ----195.875.870.14----
Table 2: Ranked scores for all TADPOLE submissions and benchmarks using the longitudinal prediction data set (D2). Best scores in each category are bolded. Missing numerical entries indicate that submissions did not include forecasts for the corresponding target variable. The “Diagnosis” ranking uses multiclass area under the receiver operating characteristic curve (MAUC), those of ADAS-Cog13 and Ventricles use mean absolute error (MAE). The overall ranking on the left uses the sum of the ranks from the three target variables. The table also lists the secondary metrics: BCA – balanced classification accuracy, WES – weighted error score, CPA – coverage probability accuracy. See Online Methods section 5.2 for details on performance metrics.

Table 2 compiles all metrics for all TADPOLE submitted forecasts, as well as benchmarks and ensemble forecasts, from the longitudinal D2 prediction set. For details on datasets D2 and D3, see Online Methods section 5.1, while for details on performance metrics see Online Methods section 5.2. Box-plots showing the distribution of scores, computed on 50 bootstraps of the test set, are shown in Supplementary Fig. 2, while the distribution of ranks is shown in Supplementary Figs. 911. Among the benchmark methods, BenchmarkMixedEffectsAPOE had the best overall rank of 18, obtaining rank 35 on clinical diagnosis prediction, rank 2 on ADAS-Cog13 and rank 23 on Ventricle volume prediction. Removing the APOE status as covariate proved to significantly increase the predictive performance (BenchmarkMixedEffects), although we do not show ranks for this entry as it was found during the evaluation phase. Among participant methods, the submission with the best overall rank was Frog, obtaining rank 1 for prediction of clinical diagnosis, rank 4 for ADAS-Cog13 and rank 10 for Ventricle volume prediction.

For clinical diagnosis, the best submitted forecasts (team Frog) scored better than all benchmark methods, reducing the error of the best benchmark methods by 58% for the multiclass area under the receiver operating characteristic curve (MAUC) and by 38% for balanced classification accuracy (BCA). Here, the best benchmarks obtained a MAUC of 0.846 (BenchmarkMixedEffects) and a BCA of 0.792 (BenchmarkLastVisit). Among participant methods, Frog had the best MAUC score of 0.931, significantly better than all entries other than Threedays according to the bootstrap test (p-value = 0.24, see Supplementary section B.1 for details on significance testing). Supplementary Figure 9 further shows the variability in performance ranking over bootstrap samples and highlights that the top two entries consistently remain at the top of the ranking. In terms of BCA, ARAMIS-Pascal had the best score of 0.850. Moreover, ensemble methods (ConsensusMedian) achieved the second best MAUC score of 0.925 and the best BCA score of 0.857. In contrast, the best randomised prediction (RandomisedBest) achieved a much lower MAUC of 0.800 and a BCA of 0.803, suggesting entries below these scores did not perform significantly better than random guessing according to the bootstrap test (p-value = 0.01). MAUC and BCA performance metrics had a relatively high correlation across all submissions (r𝑟r = 0.88, Supplementary Fig. 4).

For Ventricle volume, the best submitted forecasts among participants (team EMC1) also scored considerably better than all benchmark methods, reducing the error of the best benchmark methods by almost one third (29%) for mean absolute error (MAE) and around one half (51%) for weighted error score (WES). Here, the best benchmark method (BenchmarkMixedEffects) had an overall Ventricle MAE and WES of 0.56. Among participant submissions, EMC1-Std/-Custom had the best MAE of 0.41 (% ICV), significantly lower than all entries other than lmaUCL-Covariates/-Std/-half-D1, BORREGOTECMTY and SBIA according to the Wilcoxon signed-rank test (see Supplementary section B.2) – this is also confirmed in Supplementary Fig. 11 by the variability in performance ranking over bootstrap samples. Team EMC1 also had the best Ventricle WES of 0.29, while DIKU-ModifiedMri-Custom/-Std had the best Ventricle coverage probability accuracy (CPA) of 0.01. Ensemble methods (ConsensusMean) achieved the best Ventricle MAE of 0.38. In contrast, the best randomised prediction (RandomisedBest) achieved a higher MAE of 0.46, WES of 0.45 and CPA of 0.33. MAE and WES scores showed high correlation (r𝑟r = 0.99, Supplementary Fig. 4) and were often of equal value for many submissions (n=24𝑛24n=24), as teams set equal weights for all subjects analysed. CPA did not correlate (r0.01𝑟0.01r\approx-0.01, Supplementary Fig. 4) with either MAE or WES.

For ADAS-Cog13, the best submitted forecasts did not score significantly better than the simple benchmarks. Here, the simple BenchmarkMixedEffects model obtained the second-best MAE of 4.19, which was significantly lower than all other submitted entries according to the Wilcoxon signed-rank test. BenchmarkMixedEffects also had the best ADAS-Cog13 WES of 4.19, while VikingAI-Sigmoid had the best ADAS-Cog13 CPA of 0.02. Among participants’ submissions, FortuneTellerFish-Control ranked first in ADAS-Cog13 prediction with a MAE of 4.70, which is 11% higher than the error of the best benchmark. Moreover, all participants’ forecasts scored worse than the best randomised prediction (RandomisedBest), which here achieved a MAE of 4.52 and WES of 4.52. Nevertheless, the ensemble method ConsensusMean obtained the best ADAS scores for MAE (3.75), WES (3.54) and CPA (0.0), which along with BenchmarkMixedEffects were the only entries that performed significantly better than random guessing (p-value = 0.01). The MAE and WES scores for ADAS-Cog13 had relatively high correlation (r=0.97𝑟0.97r=0.97, Supplementary Fig. 4) and were often of equal value for many submissions (n=25𝑛25n=25). CPA had a weak but significant correlation with MAE (r=0.37𝑟0.37r=0.37, p-value << 0.02) and WES (r=0.35𝑟0.35r=0.35, p-value << 0.02).

2.4 Forecasts from the cross-sectional prediction set (D3) and custom prediction sets

OverallDiagnosisADAS-Cog13Ventricles (% ICV)
SubmissionRankRankMAUCBCARankMAEWESCPARankMAEWESCPA
ConsensusMean   --0.9170.821-4.584.340.12-0.730.720.09
ConsensusMedian   --0.9050.817-5.445.370.19-0.710.650.10
GlassFrog-Average   12-40.8970.82655.865.570.2530.680.550.24
GlassFrog-LCMEM-HDR   22-40.8970.82696.576.560.3410.480.380.24
GlassFrog-SM   32-40.8970.82645.775.770.1990.820.550.07
Tohka-Ciszek-RandomForestLin   4110.8650.78624.924.920.10100.830.830.35
RandomisedBest   --0.8110.783-4.544.500.26-0.920.500.00
lmaUCL-Std   5-912-140.8540.69816-186.956.930.055-70.810.810.22
lmaUCL-Covariates   5-912-140.8540.69816-186.956.930.055-70.810.810.22
lmaUCL-halfD1   5-912-140.8540.69816-186.956.930.055-70.810.810.22
Rocket   5-9100.8650.77135.275.140.39231.061.060.27
VikingAI-Logistic   5-980.8760.76865.945.910.22221.041.010.18
EMC1-Std   10300.7050.56776.296.190.4740.800.620.48
BenchmarkMixedEffects   --0.8390.728-4.234.230.34-1.131.130.50
SBIA   11280.7790.782106.636.430.4080.820.750.18
BGU-LSTM   12-145-70.8770.77613-156.756.170.3926-281.110.790.17
BGU-RFFIX   12-145-70.8770.77613-156.756.170.3926-281.110.790.17
BGU-RF   12-145-70.8770.77613-156.756.170.3926-281.110.790.17
BravoLab   15180.8130.730288.028.020.4720.640.640.42
BORREGOTECMTY   16-17150.8520.74886.445.860.46301.141.020.49
CyberBrains   16-17170.8300.75514.724.720.21351.541.540.50
ATRI-Biostat-MA   18190.7990.772267.396.630.04110.930.970.10
DIKU-GeneralisedLog-Std   19-20200.7980.68420-216.996.990.1716-170.950.950.05
EMC-EB   19-2090.8690.765277.717.910.50211.031.070.49
DIKU-GeneralisedLog-Custom   21210.7980.68120-216.996.990.1716-170.950.950.05
DIKU-ModifiedLog-Std   22-2322-230.7980.68822-257.107.100.1712-150.950.950.05
DIKU-ModifiedMri-Std   22-2322-230.7980.68822-257.107.100.1712-150.950.950.05
DIKU-ModifiedMri-Custom   24-2524-250.7980.69122-257.107.100.1712-150.950.950.05
DIKU-ModifiedLog-Custom   24-2524-250.7980.69122-257.107.100.1712-150.950.950.05
Billabong-Uni   26310.7040.62611-126.696.690.3819-200.980.980.48
Billabong-UniAV45   27320.7030.62011-126.696.690.3819-200.980.980.48
ATRI-Biostat-JMM   28260.7940.781298.458.120.34180.971.450.37
CBIL   29160.8470.7803310.9911.650.49291.121.120.39
BenchmarkLastVisit   30270.7850.771196.977.070.42331.170.640.11
Billabong-MultiAV45   31330.6820.60330-319.309.300.4324-251.091.090.49
Billabong-Multi   32340.6810.60530-319.309.300.4324-251.091.090.49
ATRI-Biostat-LTJMM   33290.7320.6753412.7463.980.37321.171.070.40
BenchmarkSVM   34360.4940.4903210.0110.010.42311.151.180.50
DIVE   35350.5120.4983516.6616.740.41341.421.420.34
IBM-OZ-Res   -10.9050.830----361.771.770.50
Table 3: Ranked prediction scores for all TADPOLE submissions that used the cross-sectional prediction data set (D3). Best scores in each category are bolded. Missing numerical entries indicate that submissions did not include predictions for the corresponding target variable. The “Diagnosis” ranking uses multiclass area under the receiver operating characteristic curve (MAUC), those of ADAS-Cog13 and Ventricles use mean absolute error (MAE). The overall ranking on the left uses the sum of the ranks from the three target variables. The table also lists the secondary metrics: BCA – balanced classification accuracy, WES – weighted error score, CPA – coverage probability accuracy.

Table 3 shows the ranking of the forecasts from the cross-sectional D3 prediction set. Box-plots showing the distribution of scores, computed on 50 bootstraps of the test set, are shown in Supplementary Fig 3, while the distribution of ranks is shown in Supplementary Figs. 1214. Due to the lack of longitudinal data, most submissions had lower performance compared to their equivalents from the D2 longitudinal prediction set. Among submitted forecasts, GlassFrog-Average had the best overall rank, as well as rank 2-4 on diagnosis prediction, rank 5 on ADAS-Cog13 prediction and rank 3 on ventricle prediction.

For clinical diagnosis prediction on D3, the best prediction among TADPOLE participants (team IBM-OZ-Res) scored considerably better than all benchmark methods, reducing the error of the best benchmark method by 40% for MAUC and by 25% for BCA, and achieving error rates comparable to the best predictions from the longitudinal prediction set D2. The best benchmark methods obtained a MAUC of 0.839 (BenchmarkMixedEffects) and a BCA of 0.771 (BenchmarkLastVisit). Among participant methods, IBM-OZ-Res had the best MAUC score of 0.905, significantly better than all entries other than GlassFrog-SM/-Average/-LCMEM-HDR, BGU-RF/-RFFIX/-LSTM, VikingAI-Logistic, EMC-EB, Rocket and Tohka-Ciszek-RandomForestLin according to the bootstrap hypothesis test (same methodology as in D2). This is further confirmed in Supplementary Fig. 12 by the variability of ranks under boostrap samples of the dataset, as these teams often remain at the top of the ranking. IBM-OZ-Res also had the best BCA score of 0.830 among participants. Among ensemble methods, ConsensusMean obtained the best Diagnosis MAUC of 0.917. In contrast, the best randomised prediction (RandomisedBest) obtained an MAUC of 0.811 and a BCA of 0.783. MAUC and BCA performance metrics had a relatively high correlation across all submissions (r=0.9𝑟0.9r=0.9, Supplementary Fig. 5).

For Ventricle volume prediction on D3, the best prediction (GlassFrog-LCMEM-HDR) scored considerably better than all benchmark methods, reducing the error of the best benchmark methods by 58% for MAE and 41% for WES, and achieving error rates comparable to the best predictions of D2. Here, the best benchmark methods had an overall Ventricle MAE of 1.13 (BenchmarkMixedEffects) and WES of 0.64 (BenchmarkLastVisit). Among participant submissions, GlassFrog-LCMEM-HDR had the best MAE of 0.48, significantly lower than all other submitted entries according to the Wilcoxon signed-rank test – this is also confirmed in Supplementary Fig. 14 by the rank distribution under dataset boostraps. GlassFrog-LCMEM-HDR also had the best Ventricle WES of 0.38, while submissions by team DIKU had the best Ventricle CPA of 0.05. Among ensemble methods, ConsensusMedian obtained a Ventricle MAE of 0.71 (4th best) and WES of 0.65 (7th best). In contrast, the best randomised prediction (RandomisedBest) obtained a Ventricle MAE of 0.92, WES of 0.50 and CPA of 0. As in D2, MAE and WES scores in D3 for Ventricles had very high correlation (r=0.99𝑟0.99r=0.99, Supplementary Fig. 5), while CPA showed weak correlation with MAE (r=0.24𝑟0.24r=0.24, p-value = 0.17) and WES (r=0.37𝑟0.37r=0.37, p-value <0.032absent0.032<0.032).

For ADAS-Cog13 on D3, the predictions submitted by participants again did not perform better than the best benchmark methods. BenchmarkMixedEffects had the best MAE of 4.23, which was significantly lower than all entries by other challenge participants. Moreover, the MAE of 4.23 was only marginally worse than the equivalent error (4.19) by the same model on D2. BenchmarkMixedEffects also had the best ADAS-Cog13 WES of 4.23, while ATRI-Biostat-MA had the best ADAS-Cog13 CPA of 0.04. Among participants’ submissions, CyberBrains ranked first in ADAS-Cog13 prediction with a MAE of 4.72, an error 11% higher than the best benchmark. Among ensemble methods, ConsensusMean obtained an ADAS-Cog13 MAE of 4.58, WES of 4.34, better than all participants’ entries. As in D2, the best randomised predictions (RandomisedBest) obtained an ADAS-Cog13 MAE of 4.54 (2nd best) and WES of 4.50 (3rd best). As in D2, MAE and WES scores for ADAS-Cog13 had high correlation (r𝑟r = 0.97, Supplementary Fig. 5), while CPA showed weak, non-significant correlation with MAE (r𝑟r = 0.34, p-value \approx 0.052) or WES (r𝑟r = 0.33, p-value \approx 0.057).

Results on the custom prediction sets are presented in Supplementary Table 6.

2.5 Algorithm characteristics associated with increased performance

To understand what characteristics of algorithms could have yielded higher performance, we show in Figure 1 associations from a general linear model between predictive performance and feature selection methods, different types of features, methods for data imputation, and methods for forecasting of target variables. For each type of feature/method and each target variable (clinical diagnosis, ADAS-Cog13 and Ventricles), we show the distribution of estimated coefficients from a general linear model, derived from the approximated inverse Hessian matrix at the maximum likelihood estimator (see Online Methods section 5.3). From this analysis we removed outliers, defined as submissions with ADAS MAE higher than 10 and Ventricle MAE higher than 1.15 (%ICV). For all plots, distributions to the right of the gray dashed vertical line denote increased performance compared to baseline (i.e. when those characteristics are not used).

For feature selection, Figure 1 shows that methods with manual selection of features tend to be associated with better predictive performance in ADAS-Cog13 and Ventricles. In terms of feature types, CSF and DTI features were generally associated with an increase in predictive performance for clinical diagnosis, while augmented features were associated with performance improvements for ventricle prediction. In terms of data imputation methods, while some differences can be observed, no clear conclusions can be drawn. In terms of prediction models, the only positive association that indicates increased performance is in the neural networks for ventricle prediction. However, given the small number of methods tested (<<50) and the large number of degrees of freedom, these results should be interpreted with care.

Refer to caption
Figure 1: Associations between the prediction of clinical diagnosis, ADAS-Cog13 and Ventricle volume and different strategies of (top) feature selection, (upper-middle) types of features, (lower-middle) data imputation strategies and (bottom) prediction methods for the target variables. For each type of feature/method (rows) and each target variable (columns), we show the distribution of estimated coefficients from a general linear model. Positive coefficients, where distributions lie to the right of the dashed vertical line, indicate better performance than baseline (vertical dashed line). For ADAS-Cog13 and Ventricle prediction, we flipped the sign of the coefficients, to consistently show better performance to the right of the vertical line.

3 Discussion

In this work, we presented the results of the TADPOLE Challenge. The results of the challenge provide important insights into the current state of the field, and how well current algorithms can predict progression of AD diagnoses and markers of disease progression both from rich longitudinal data sets and, comparatively, from sparser cross-sectional data sets typical of a clinical trial scenario. The challenge further highlights the algorithms, features and data-handling strategies that tend to lead to improved forecasts. In the following sections we discuss the key conclusions that we draw from our study and highlight important limitations.

3.1 TADPOLE pushed forward performance on AD clinical diagnosis prediction

When comparing to previous state-of-the-art results in the literature, the best TADPOLE methods show similar or higher performance in AD diagnostic classification while also tackling a harder problem than most previous studies of predicting future, rather than estimating current, classification. A comparison of 15 studies presented by moradi2015machine reported lower performance (maximum AUC of 0.902 vs 0.931 obtained by the best TADPOLE method) for the simpler two-class classification problem of separating MCI-stable from MCI-converters in ADNI. A more recent method by long2017prediction reported a maximum AUC of 0.932 and accuracy of 0.88 at the same MCI-stable vs -converter classification task. However, a) TADPOLE’s discrimination of CN-converters from CN-stable subjects is harder as disease signal is weaker at such early stages, and b) the predictive performance drops in three-class problems like TADPOLE compared to two-class. Furthermore, the best out of 19 algorithms in the CADDementia Challenge bron2015standardized obtained an MAUC of 0.78.

We are unaware of previous studies forecasting future ventricle volume or ADAS-Cog13, so TADPOLE sets a new benchmark state-of-the-art performance on these important prediction tasks.

3.2 No one-size-fits-all prediction algorithm

The results on the longitudinal D2 prediction set suggest no clear winner on predicting all target variables – no single method performed well on all tasks. While Frog had the best overall submission with the lowest sum of ranks, for each performance metric individually we had different winners: Frog (clinical diagnosis MAUC of 0.931), ARAMIS-Pascal (clinical diagnosis BCA of 0.850), BenchmarkMixedEffects (ADAS-Cog13 MAE and WES of 4.19), VikingAI-Sigmoid (ADAS-Cog13 CPA of 0.02), EMC1-Std/EMC1-Custom (ventricle MAE of 0.41 and WES or 0.29), and DIKU-ModifiedMri-Std/-Custom (ventricle CPA of 0.01). Moreover, on the cross-sectional D3 prediction set, the methods by Glass-Frog had the best performance. Associations of method-type with increased performance in Fig. 1 confirm no clear increase in performance for any types of prediction methods (with the exception of neural networks for ventricle volume prediction). This raises an important future challenge to algorithm designers to develop methods able to perform well on multiple forecasting tasks and also in situations with limited data, such as D3.

3.3 Ensemble methods perform strongly

Consistently strong results from ensemble methods (ConsensusMean/ConsensusMedian outperformed all others on most tasks) might suggest that different methods over-estimate future measurements for all subjects while others under-estimate them, likely due to the underlying assumptions they make. This is confirmed by plots of the difference between true and estimated measures (Supplementary Figures 68), where most methods systematically under- or over-estimate in all subjects. However, even if methods were completely unbiased, averaging over all methods could also help predictions by reducing the variance in the estimated target variables.

3.4 Predictability of ADAS-Cog13 scores

ADAS-Cog13 scores were more difficult to forecast than clinical diagnosis or ventricle volume. The only single method able to forecast ADAS-Cog13 better than informed random guessing (RandomisedBest) was the BenchmarkMixedEffects, a simple mixed effects model with no covariates and age as a regressor. The difficulty could be due to variability in administering the tests or practice effects. A useful target performance level comes from the 4 points change generally used to identify responders to a drug treatment grochowalski2016examining . With the exception of the ensemble method, all submitted forecasts failed to produce mean error below 4, highlighting the substantial challenge of estimating change in ADAS-Cog13 over the 1.4 year interval – although over longer time periods, non-trivial forecasts are likely to improve in comparison to RandomisedBest, which is independent of time period. Nevertheless, for the longitudinal D2 prediction set, the MAE in ADAS-Cog13 from ConsensusMean was 3.75, which restores hope in forecasting cognitive score trajectories even over relatively short timescales.

3.5 Prediction errors from limited cross-sectional dataset mimicking clinical trials are similar to those from longitudinal dataset

For clinical diagnosis, the best performance on the limited, cross-sectional D3 prediction set was similar to the best performance on the D2 longitudinal prediction set: 0.917 vs 0.931 for MAUC (p-value = 0.14), representing a 3% error increase for D3 compared to D2. Slightly larger and significant differences were observed for ADAS MAE (3.75 vs 4.23, p-value << 0.01) and Ventricle MAE (0.38 vs 0.48, p-value << 0.01). It should be noted that Ventricle predictions for D3 were extremely difficult, given that only 25% of subjects to be forecasted had MRI data in D3. This suggests that, for clinical diagnosis, current forecast algorithms are reasonably robust to lack of longitudinal data and missing inputs, while for ADAS and Ventricle volume prediction, some degree of performance is lost. Future work is also required to determine the optimal balance of input data quality and quantity versus cost of acquisition.

3.6 DTI and CSF features appear informative for clinical diagnosis prediction, augmented features appear informative for ventricle prediction

DTI and CSF features are most associated with increases in clinical diagnosis forecast performance. CSF, in particular, is well established as an early marker of AD jack2010hypothetical and likely to help predictions for early-stage subjects, while DTI, measuring microstructure damage, may be informative for middle-stage subjects. On the other hand, for prediction of ventricle volume, augmented features had the highest association with increases in prediction performance. Future work is required to confirm the added value of these features and others in a more systematic way.

3.7 Challenge design and limitations

TADPOLE Challenge has several limitations that future editions of the challenge may consider addressing. One limitation is the reliability of the three target variables: clinical diagnosis, ADAS-Cog13 and Ventricle volume. First of all, clinical diagnosis has only moderate agreement with gold-standard neuropathological post-mortem diagnosis. In particular, one study beach2012accuracy has shown that a clinical diagnosis of probable AD has sensitivity between 70.9% and 87.3% and specificity between 44.3% and 70.8%. With the advent of post-mortem confirmation in ADNI, future challenges might address this by evaluating the algorithms on subjects with pathological confirmation. Similarly, ADAS-Cog13 is known to suffer from low reliability across consecutive visits grochowalski2016examining , and TADPOLE algorithms fail to forecast it reliably. However, this might be related to the short time-window (1.4 years), and more accurate predictions might be possible over longer time-windows, when there is more significant cognitive decline. Ventricle volume measurements depend on MRI scanner factors such as field strength, manufacturer and pulse sequences han2006reliability , although these effects have been removed to some extent by ADNI through data preprocessing and protocol harmonization. TADPOLE Challenge also assumes all subjects either remain stable or convert to Alzheimer’s disease, whereas in practice some of them might develop other types of neurodegenerative diseases.

For performance evaluation, we elected to use very simple yet reliable metrics as the primary performance scores: the multiclass area under the curve (mAUC) for the clinical categorical variable and the mean absolute error (MAE) for the two numerical variables. While the mAUC accounts for decision confidence, the MAE does not, which means that the confidence intervals submitted by participants do not contribute to the rankings computed in Tables 2 and 3. While the weighted error score (WES) takes confidence intervals into account, we consider it susceptible to “hacking”, e.g. participants might assign high confidence to only one or two data points and thereby skew the score to ignore most of the predictions – in practice, we did not observe this behaviour in any submission. For clinical relevance, we believe that confidence intervals are an extremely important part of such predictions and urge future studies to consider performance metrics that require and take account of participant-calculated confidence measures.

TADPOLE has other limitations related to the algorithms’ comparability and generalisability. First of all, the evaluation and training were both done on data collected by ADNI – in future work, we plan to assess how the models will generalise on different datasets. Another limitation is that we can only compare full methods submissions and not different types of features, and strategies for data imputation and prediction used within the full method. While we tried to evaluate the effect of these characteristics in Figure 1, in practice the numbers were small and hence most effects did not reach statistical significance. Moreover, the challenge format does not provide an exhaustive comparison of all combinations of data processing, predictive model, features, etc., so does not lead to firm conclusions on the best combinations but rather provides hypotheses for future testing. In future work, we plan to test inclusion of features and strategies for data imputation and prediction independently, by changing one such characteristic at a time.

Another limitation is that the number of controls and MCI converters in the D4 test set is low (9 MCI converters and 9 control converters). However, these numbers will increase over time as ADNI acquires more data, and we plan to re-run the evaluation at a later stage with the additional data acquired after April 2019. A subsequent evaluation will also enable us to evaluate the TADPOLE methods on longer time-horizons, over which the effects of putative drugs would be higher.

4 Conclusion

In this work we presented the results of the TADPOLE Challenge. The results of the challenge provide important insights into the current state of the art in AD forecasting, such as performance levels achievable with current data and technology as well as specific algorithms, features and data-handling strategies that support the best forecasts. The developments and outcomes of TADPOLE Challenge can aid refinement of cohorts and endpoint assessment for clinical trials, and can support accurate prognostic information in clinical settings. The challenge website (https://tadpole.grand-challenge.org) will stay open for submissions, which can be added to our current ranking. The open test set remains available on the ADNI LONI website and also allows individual participants to evaluate future submissions. Through TADPOLE-SHARE https://tadpole-share.github.io/, we further plan to implement many TADPOLE methods in a common framework, to be made publicly available. TADPOLE provides a standard benchmark for evaluation of future AD prediction algorithms.

5 Online Methods – Challenge design and prediction algorithms

5.1 Data

The challenge uses data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) weiner2017recent . Specifically, the TADPOLE Challenge made four key data sets available to the challenge participants:

  • D1: The TADPOLE standard training set draws on longitudinal data from the entire ADNI history. The data set contains measurements for every individual that has provided data to ADNI in at least two separate visits (different dates) across three phases of the study: ADNI1, ADNI GO, and ADNI2.

  • D2: The TADPOLE longitudinal prediction set contains as much available data as we could gather from the ADNI rollover individuals for whom challenge participants are asked to provide predictions. D2 includes data from all available time-points for these individuals. It defines the set of individuals for which participants are required to provide forecasts.

  • D3: The TADPOLE cross-sectional prediction set contains a single (most recent) time point and a limited set of variables from each rollover individual in D2. Although we expect worse predictions from this data set than D2, D3 represents the information typically available when selecting a cohort for a clinical trial.

  • D4: The TADPOLE test set contains visits from ADNI rollover subjects that occurred after 1 Jan 2018 and contain at least one of the three outcome measures: diagnostic status, ADAS-Cog13 score, or ventricle volume.

While participants were free to use any training datasets they wished, we provided the D1-D3 datasets in order to remove the need for participants to pre-process the data themselves, and also to be able to evaluate the performance of different algorithms on the same standardised datasets. Participants that used custom training data sets were asked also to submit results using the standard training data sets to enable direct performance comparison. We also included the D3 cross-sectional prediction set in order to simulate a clinical trial scenario. For information on how we created the D1-D4 datasets, see Supplementary section A. The software code used to generate the standard datasets is openly available on Github: https://github.com/noxtoby/TADPOLE.

Table 4 shows the demographic breakdown of each TADPOLE data set as well as the proportion of biomarker data available in each dataset. Many entries are missing data, especially for certain biomarkers derived from exams performed on only subsets of subjects, such as tau imaging (AV1451). D1 and D2 also included demographic data typically available in ADNI (e.g. education, marital status) as well as standard genetic markers (e.g. Alipoprotein E – APOE epsilon 4 status).

Demographics
D1D2D3D4
Overall number of subjects1667896896219
ControlsNumber (% all subjects)508 (30.5%)369 (41.2%)299 (33.4%)94 (42.9%)
Visits per subject8.3 ±plus-or-minus\pm 4.58.5 ±plus-or-minus\pm 4.91.0 ±plus-or-minus\pm 0.01.0 ±plus-or-minus\pm 0.2
Age74.3 ±plus-or-minus\pm 5.873.6 ±plus-or-minus\pm 5.772.3 ±plus-or-minus\pm 6.278.4 ±plus-or-minus\pm 7.0
Gender (% male)48.6%47.2%43.5%47.9%
MMSE29.1 ±plus-or-minus\pm 1.129.0 ±plus-or-minus\pm 1.228.9 ±plus-or-minus\pm 1.429.1 ±plus-or-minus\pm 1.1
Converters*189--
MCINumber (% all subjects)841 (50.4%)458 (51.1%)269 (30.0%)90 (41.1%)
Visits per subject8.2 ±plus-or-minus\pm 3.79.1 ±plus-or-minus\pm 3.61.0 ±plus-or-minus\pm 0.01.1 ±plus-or-minus\pm 0.3
Age73.0 ±plus-or-minus\pm 7.571.6 ±plus-or-minus\pm 7.271.9 ±plus-or-minus\pm 7.179.4 ±plus-or-minus\pm 7.0
Gender (% male)59.3%56.3%58.0%64.4%
MMSE27.6 ±plus-or-minus\pm 1.828.0 ±plus-or-minus\pm 1.727.6 ±plus-or-minus\pm 2.228.1 ±plus-or-minus\pm 2.1
Converters*11737-9
ADNumber (% all subjects)318 (19.1%)69 (7.7%)136 (15.2%)29 (13.2%)
Visits per subject4.9 ±plus-or-minus\pm 1.65.2 ±plus-or-minus\pm 2.61.0 ±plus-or-minus\pm 0.01.1 ±plus-or-minus\pm 0.3
Age74.8 ±plus-or-minus\pm 7.775.1 ±plus-or-minus\pm 8.472.8 ±plus-or-minus\pm 7.182.2 ±plus-or-minus\pm 7.6
Gender (% male)55.3%68.1%55.9%51.7%
MMSE23.3 ±plus-or-minus\pm 2.023.1 ±plus-or-minus\pm 2.020.5 ±plus-or-minus\pm 5.919.4 ±plus-or-minus\pm 7.2
Converters*---9
Number of clinical visits for all subjects with data available (% of total visits)
D1D2D3D4
Cognitive8862 (69.9%)5218 (68.1%)753 (84.0%)223 (95.3%)
MRI7884 (62.2%)4497 (58.7%)224 (25.0%)150 (64.1%)
FDG-PET2119 (16.7%)1544 (20.2%)--
AV452098 (16.6%)1758 (23.0%)--
AV145189 (0.7%)89 (1.2%)--
DTI779 (6.1%)636 (8.3%)--
CSF2347 (18.5%)1458 (19.0%)--
Table 4: Summary of TADPOLE datasets D1-D4. () Diagnosis at first visit with available data. For D3 and D4, 192 and 6 subjects respectively did not have a diagnosis at any clinical visit, so numbers don’t add up to 100%. (*) For D4, converters are ADNI3 subjects who are MCI, but were previously CN, or who are AD, but were previously CN or MCI in their last visit in ADNI2. For D1, D2 and D3, converters are CN or MCI at their earliest available visit, who progress to a later classification of MCI/AD within 1.4 years (same duration as D4)

.

5.2 Forecast Evaluation

For evaluation of clinical status predictions, we used similar metrics to those that proved effective in the CADDementia challenge bron2015standardized : (i) the multiclass area under the receiver operating characteristic curve (MAUC) and (ii) the overall balanced classification accuracy (BCA). For ADAS-Cog13 and ventricle volume, we used three metrics: (i) mean absolute error (MAE), weighted error score (WES) and coverage probability accuracy (CPA). BCA and MAE focus purely on prediction accuracy ignoring confidence, MAUC and WES account for accuracy and confidence, while CPA assesses the confidence interval only. The formulas for each performance metric are summarised in Table 5. See the TADPOLE white paper marinescu2018tadpole for further rationale for choosing these performance metrics. In order to characterise the distribution of these metric scores, we compute scores based on 50 bootstraps with replacement on the test dataset.

FormulaDefinitions
mAUC=1L(L1)i=2Lj=1iA^(ci|cj)+A^(cj|ci)𝑚𝐴𝑈𝐶1𝐿𝐿1superscriptsubscript𝑖2𝐿superscriptsubscript𝑗1𝑖^𝐴conditionalsubscript𝑐𝑖subscript𝑐𝑗^𝐴conditionalsubscript𝑐𝑗subscript𝑐𝑖mAUC=\frac{1}{L(L-1)}\sum_{i=2}^{L}\sum_{j=1}^{i}\hat{A}(c_{i}|c_{j})+\hat{A}(c_{j}|c_{i})where A^(ci|cj)=Sini(ni+1)/2ninj^𝐴conditionalsubscript𝑐𝑖subscript𝑐𝑗subscript𝑆𝑖subscript𝑛𝑖subscript𝑛𝑖12subscript𝑛𝑖subscript𝑛𝑗\hat{A}(c_{i}|c_{j})=\frac{S_{i}-n_{i}(n_{i}+1)/2}{n_{i}n_{j}}nisubscript𝑛𝑖n_{i}, njsubscript𝑛𝑗n_{j} – number of points from class i𝑖i and j𝑗j. Sijsubscript𝑆𝑖𝑗S_{ij} – the sum of the ranks of the class i𝑖i test points, after ranking all the class i𝑖i and j𝑗j data points in increasing likelihood of belonging to class i𝑖i. L𝐿L – number of classes. cisubscript𝑐𝑖c_{i} – class i𝑖i.
BCA=12Li=1L[TPiTPi+FNi+TNiTNi+FPi]𝐵𝐶𝐴12𝐿superscriptsubscript𝑖1𝐿delimited-[]𝑇subscript𝑃𝑖𝑇subscript𝑃𝑖𝐹subscript𝑁𝑖𝑇subscript𝑁𝑖𝑇subscript𝑁𝑖𝐹subscript𝑃𝑖BCA=\frac{1}{2L}\sum_{i=1}^{L}\left[\frac{TP_{i}}{TP_{i}+FN_{i}}+\frac{TN_{i}}{TN_{i}+FP_{i}}\right]TPi𝑇subscript𝑃𝑖TP_{i}, FPi𝐹subscript𝑃𝑖FP_{i}, TNi𝑇subscript𝑁𝑖TN_{i}, FNi𝐹subscript𝑁𝑖FN_{i} – the number of true positives, false positives, true negatives and false negatives for class i𝑖i. L𝐿L – number of classes
MAE=1Ni=1N|M~iMi|𝑀𝐴𝐸1𝑁superscriptsubscript𝑖1𝑁subscript~𝑀𝑖subscript𝑀𝑖MAE=\frac{1}{N}\sum_{i=1}^{N}\left|{\tilde{M}_{i}-M_{i}}\right|Misubscript𝑀𝑖M_{i} is the actual value in individual i𝑖i in future data. M~isubscript~𝑀𝑖\tilde{M}_{i} is the participant’s best guess at Misubscript𝑀𝑖M_{i} and N𝑁N is the number of data points
WES=i=1NC~i|M~iMi|i=1NC~i𝑊𝐸𝑆superscriptsubscript𝑖1𝑁subscript~𝐶𝑖subscript~𝑀𝑖subscript𝑀𝑖superscriptsubscript𝑖1𝑁subscript~𝐶𝑖WES=\frac{\sum_{i=1}^{N}\tilde{C}_{i}\left|\tilde{M}_{i}-M_{i}\right|}{\sum_{i=1}^{N}\tilde{C}_{i}}Misubscript𝑀𝑖M_{i}, M~isubscript~𝑀𝑖\tilde{M}_{i} and N𝑁N defined as above. C~i=(C+C)1subscript~𝐶𝑖superscriptsubscript𝐶subscript𝐶1\tilde{C}_{i}=(C_{+}-C_{-})^{-1}, where [C,C+]subscript𝐶subscript𝐶\left[C_{-},C_{+}\right] is the 50% confidence interval
CPA=|ACP0.5|𝐶𝑃𝐴𝐴𝐶𝑃0.5CPA=|ACP-0.5|actual coverage probability (ACP) - the proportion of measurements that fall within the 50% confidence interval.
Table 5: TADPOLE performance metric formulas and definitions for the terms.

5.3 Statistical Analysis of Method Attributes with Performance

To identify which features and types of algorithms enable good predictions, we annotated each TADPOLE submission with a set of 21 attributes related to (i) feature selection (manual/automatic and large vs. small number of features), (ii) feature types (e.g. “uses Amyloid PET”), (iii) strategy for data imputation (e.g. “patient-wise forward-fill”) and (iv) prediction method (e.g. “neural network”) for clinical diagnosis and ADAS/Ventricles separately. To understand which of these annotations were associated with increased performance, we applied a general linear model kiebel2007general , Y=Xβ+ϵ𝑌𝑋𝛽italic-ϵY=X\beta+\epsilon, where Y𝑌Y is the performance metric (e.g. diagnosis MAUC), X𝑋X is the nr_submissions x 21 design matrix of binary annotations, and β𝛽\beta show the contributions of each of the 21 attributes towards achieving the performance measure Y𝑌Y.

5.4 Prediction Algorithms

Team: AlgosForGood    (Members: Tina Toni, Marcin Salaterski, Veronika Lunina, Institution: N/A)
Feature selection: Manual + Automatic: Manual selection of uncorrelated variables from correlation matrix and automatic selection of variables that have highest cumulative hazard rates in survival regression from MCI to AD.
Selected features: Demographics (age, education, gender, race, marital status), cognitive tests (ADAS-Cog13, RAVLT immediate, RAVLT forgetting, CDRSOB, ADAS11, FDG), Ventricles, AV45, ICV, APOE4.
Missing data: Fill-in using last available value from corresponding patient
Confounder correction: none
Method category: Statistical Regression / Proportional hazards model
Prediction method:

  • Diagnosis: Aalen additive regression constructing a cumulative hazard rate for progressing to AD.

  • ADAS-Cog13: regression using change in ventricles/ICV as predictive variable, stratified by last known diagnosis.

  • Ventricles: regression over month, with several pre-processing steps: 1. Enforced monotonicity by accumulating maximum value, 2. For APOE positive patients used only last three visits due to non-linearity 3. Stratified by diagnosis

Team: Apocalypse    (Members: Manon Ansart, Stanley Durrleman, Institution: Institut du Cerveau et de la Moelle épinière, ICM, Paris, France)
Feature selection: Manual – important features were identified by looking at the correlations with the diagnosis. Personal knowledge of the disease was also used to complement those results and select relevant features. Different feature sets were compared using cross-validation.
Selected features: Cognitive features (ADAS-Cog13, MMSE, RAVLT immediate, FAQ, CDRSOB), MRI features (WholeBrain, Entorhinal, Fusiform, MidTemp, Ventricles, Entorhinal, Hippocampus), APOE4, education, age, clinical diagnosis
Missing data: Filled in using the mean feature value
Confounder correction: none
Method category: Machine learning / Regression
Prediction method: Linear regression is used to first predict the future of a set of features (MMSE, ADAS-Cog13, CDRSOB, RAVLT, Ventricles) at the prediction dates. Afterwards, an SVM is used to predict the current diagnosis for each prediction date, based on the forecasted features as well as other features which are constant for the subject (APOE4, education, age at last known visit).


Team: ARAMIS-Pascal    (Members: Pascal Lu, Institution: Institut du Cerveau et de la Moelle épinière, ICM, Paris, France)
Feature selection: Manual, based on known biomarkers from the literature.
Selected features: APOE4, cognitive tests (CDRSB, ADAS11, ADAS-Cog13, MMSE, RAVLT immediate, FAQ), volumetric MRI (hippocampus, ventricles, whole brain, entorhinal, fusiform, middle temporal, ICV), whole brain FDG, CSF biomarkers (amyloid-beta, tau, phosphorylated tau), education and age.
Missing data: Imputed using the average biomarker values across the population.
Confounder correction: none
Method category: Statistical Regression / Proportional hazards model
Prediction method: For diagnosis prediction, the Aalen model for survival analysis was used to predict the conversion from MCI to AD, which returns the probability of a subject remaining MCI as a function of time. The method assumes cognitively normal and dementia subjects will not convert and thus will remain constant. The method did not predict ADAS-Cog13 or Ventricles.
Publication link: https://hal.inria.fr/tel-02433613/document


Team: ATRI_Biostat (Members: Samuel Iddi1,2, Dan Li1, Wesley K. Thompson3 and Michael C. Donohue1. Institutions: 1Alzheimer’s Therapeutic Research Institute, USC, USA; 2Department of Statistics and Actuarial Science, University of Ghana, Ghana; 3Department of Family Medicine and Public Health, University of California, USA)
Feature selection: Automatic - features were ranked by their importance in classifying diagnostic status using a random forest algorithm. All cognitive tests, imaging biomarkers, demographic information and APOE status were considered as potential features
Selected features: ADAS-Cog13, EcogTotal, CDRSOB, FAQ, MOCA, MMSE, RAVLT immediate, Ventricles/ICV, Entorhinal, Hippocampus/ICV and FDG Pet. Age, gender and APOE status were included as covariates. The interaction between diagnosis at first available visit and years since first visit was also considered.
Missing data: Imputed using the MissForest Algorithm, based on a non-parametric random forest methodology stekhoven2011missforest . The algorithm was chosen based on its ability to handle mixed-type outcomes, complex interactions and non-linear relationships between variables.
Confounder correction: APOE status, last known clinical status, age and gender.
Method category: Machine learning and data-driven disease progression models
Prediction method: The method applied different types of mixed-effects models to forecast ADAS-Cog13 and Ventricles, and then used a Random Forest classifier to predict the clinical diagnosis from the forecasted continuous scores.

  • JMM    – Joint Mixed-effect Modelling with subject-specific intercept and slope.

  • LTJMM    – Latent Time Joint Mixed-effect Modelling with subject-specific intercept, slope and time-shift

  • MA    – Model average of the two models above, as well as a third model where random intercepts are shared across outcomes.

Confidence Intervals: The 50% prediction intervals for ADAS-Cog13 and Ventricles were obtained by taking the 25th and 75th percentile of the posterior predicted samples.
Publication link: https://braininformatics.springeropen.com/articles/10.1186/s40708-019-0099-0


Team: BGU (Members: Aviv Nahon, Yarden Levy, Dan Halbersberg, Mariya Cohen, Institution: Ben Gurion University of the Negev, Beersheba, Israel)
Feature selection: Automatic – used the following algorithm: 1. Find the two variables with highest correlation (Spearman for continuous variables and Mutual information for discrete variables). 2. Compute the correlation of each variable with the target variables separately and remove the variable with the lower correlation. 3. If there are still pairs of variables with a correlation of more than 80%, repeat from step 1.
Selected features: Cognitive tests (CDRSOB, MMSE, RAVLT, MOCA, all Ecog), MRI biomarkers (Freesurfer cross-sectional and longitudinal), FDG- PET (hypometabolic convergence index), AV45 PET (Ventricles, Corpus Callosum, Hippocampus), White-matter hypointensities’ volume, CSF biomarkers (amyloid-beta, tau, phosphorylated tau). For each continuous variable, an additional set of 20 augmented features was used, representing changes and trends in variables (e.g. mean, standard deviation, trend mean, trend standard deviation, minimum, mean minus global mean, baseline value, last observed value). This resulted in 233 features, which were used for prediction.
Missing data: Random forest can deal automatically with missing data. LSTM network used indicator that was set to zero for missing data.
Confounder correction: None
Method category: Machine learning
Prediction method:

  • BGU-LSTM   : This model consisted of two integrated neural networks: an LSTM network for modelling continuous variables and a feed-forward neural network for the static variables.

  • BGU-RF   : A semi-temporal Random Forest was used which contained the augmented features.

  • BGU-RFFIX   : Same as BGU-RF, but with small correction for the prediction of diagnosis: whenever the model predicted AD with probability higher than 80%, the probability of CN was changed to zero and vice versa.

Team: BIGS2    (Members: Huiling Liao, Tengfei Li, Kaixian Yu, Hongtu Zhu, Yue Wang, Binxin Zhao, Institution: University of Texas, Houston, USA)
Feature selection: Automatic – used auto-encoder to extract aggregated features.
Selected features: All continuous features in D1/D2, which represented the input for the autoencoder. Apart from the autoencoder-extracted features, other features used for the classifier were demographic information, APOE status, whole brain biomarkers from MRI (volume) and PET (FDG, PIB and AV45), and MMSE.
Missing data: SoftImpute method mazumder2010spectral was used for imputing missing data. The complete dataset was then used as input to the autoencoder.
Confounder correction: none
Method category: Regression and Machine Learning
Prediction method: Linear models were used to predict ADAS-Cog13 and Ventricle scores independently. For prediction of clinical diagnosis, a random forest was used based on the autoencoder-extracted features and the other selected features.


Team: Billabong    (Members: Neil Oxtoby, Institution: University College London, UK)
Feature selection: Manual, using knowledge from literature
Selected features: MRI biomarkers normalised by ICV (ventricles, hippocampus, whole brain, entorhinal, fusiform, middle temporal), FDG, AV45, CSF biomarkers (amyloid beta, tau, phosphorylated tau) and cognitive tests (ADAS-Cog13, MMSE, MOCA, RAVLT immediate). Separate submissions (Billabong-UniAV45, Billabong-MultiAV45) were made which also included AV45, that was initially excluded due to noise.
Missing data: Only imputed during staging via linear regression against age. The method can deal with missing data during training.
Confounder correction: None
Method category: Data-driven disease progression model
Prediction method: For each selected feature independently, a data-driven longitudinal trajectory was estimated using a differential equation model based on Gaussian Process Regression oxtoby2018data . Subjects were staged using either a multivariate or univariate approach:

  • Billabong-Uni: Univariate staging which estimates disease stage for each target variable independently.

  • Billabong-Multi: Multivariate staging that combines all selected features, producing an average disease stage.

For the prediction of clinical diagnosis, the historical ADNI diagnoses were mapped to a linear scale using partially-overlapping squared-exponential distribution functions. The linear scale and the three distributions were used to forecast the future diagnoses.
Custom prediction set: Predictions were made also for a custom dataset, which was similar to D3 but missing data was filled in using the last available biomarker data.
Confidence Intervals: The 25th and 75th percentiles of the GPR posterior were each integrated into a trajectory to obtain 50% confidence (credible) intervals for the forecasts of ADAS-Cog13 and Ventricles/ICV.
Publication link: https://doi.org/10.1093/brain/awy050


Team: BORREGOSTECMTY    (Members: José Gerardo Tamez-Peña, Institution: Tecnologico de Monterrey, Monterrey, Mexico)
Feature selection: Automatic, using bootstrapped stage-wise selection.
Selected features: Main cognitive tests (excluding subtypes), MRI biomarkers, APOE status, demographic information (age, gender, education) and diagnosis status. Augmented features were further constructed from the MRI set: the cubic root of all volumes, the square root of all surface areas, the compactness, the coefficient of variation, as well as the mean value and absolute difference between the left and right measurements.
Missing data: Imputed using nearest-neighbourhood strategy based on L1 norm.
Confounder correction: Gender and intracranial volume (ICV) adjustments relative to controls.
Method category: Regression (ensemble of statistical models)
Prediction method: ADAS-Cog13 and Ventricles were predicted using an ensemble of 50 linear regression models, one set for each diagnostic category. The best models were selected using Bootstrap Stage-Wise Model Selection, using statistical fitness pencina2008evaluating tests to evaluate models and features to use within the models. All selected models were then averaged in a final prediction using bagging. For prediction, the last known diagnosis of the subject was used to select the category of models for forecasting.
For the prediction of clinical diagnosis, a two-stage approach was used based on prognosis and time-to-event estimation. The prognosis approach used an ensemble of 50 regression models to estimate the future diagnosis, while the time-to-event method used an ensemble of 25 models to estimate the square root of the time it took for a subject to convert to MCI or AD. These approaches were performed independently for CN-to-MCI, MCI-to-AD and CN-to-AD conversion.
Confidence Intervals: The 50% confidence intervals for ADAS-Cog13 and Ventricle volume were estimated by extracting the interquartile range of the 50 regression estimates.
Repository link: https://github.com/joseTamezPena/TADPOLE


Team: BravoLab    (Members: Aya Ismail, Timothy Wood, Hector Corrada Bravo, Institution: University of Maryland, USA)
Feature selection: Automatic, using a random forest to select features with highest cross-entropy or GINI impurity reduction.
Selected features:

  • Ventricle prediction: MRI volumes of ventricular sub-regions (Freesurfer cross-sectional and longitudinal)

  • ADAS-Cog13 prediction: RAVLT, Diagnosis, MMSE, CDRSOB

  • Diagnosis prediction: ADAS-Cog13, ADAS11, MMSE, CSRSOB

Missing data: Imputation using Hot Deck andridge2010review was done only for data missing at random.
Confounder correction: None
Method category: Machine learning
Prediction method: A long-short term memory network (LSTM) with target replication was trained independently for each category: Diagnosis, ADA13 and Ventricles. All existing data was used for the first forecast, after which the output of the last prediction was used as input for the next prediction, along with other features that remain constant over time. Since subjects had a different number of visits and available biomarker data, the network was adapted to accept inputs of variable length. For predictions, the network used a weighted mean absolute error as a loss function. In addition, for the prediction of the clinical diagnosis, a soft-max function was used to get the final prediction.



Team: CBIL    (Members: Minh Nguyen, Nanbo Sun, Jiashi Feng, Thomas Yeo, Institution: National University of Singapore, Singapore)
Feature selection: Manual, based on model performance on D1 subset.
Selected features: Cognitive tests (CDRSOB, ADAS11, ADAS-Cog13, MMSE, RAVLT immediate, learning, forgetting and percent forgetting, MOCA, FAQ), MRI biomarkers (entorhinal, fusiform, hippocampus, ICV, middle temporal, ventricles, whole brain), whole brain AV45 and FDG, CSF biomarkers (amyloid-beta, tau, phosphorylated tau).
Missing data: Imputation using interpolation.
Confounder correction: None
Method category: Machine learning and data-driven disease progression model
Prediction method: Recurrent neural network adapted for variable duration between time-points. A special loss function was designed, which ensured forecasts at timepoints close together are more correlated than those at timepoints further apart.
Confidence Intervals: hardcoded values
Publication link: https://www.biorxiv.org/content/10.1101/755058v1
Repository link: https://github.com/ThomasYeoLab/CBIG/tree/master/stable_projects/predict_phenotypes/Nguyen2020_RNNAD


Team: Chen-MCW    (Members: Gang Chen, Institution: Medical College of Wisconsin, Milwaukee, USA)
Feature selection: Manual
Selected features: ADAS-Cog13, MMSE, MRI volumes (hippocampus, whole brain, entorhinal, fusiform and middle temporal), APOE status, gender and education.
Missing data: No imputation performed.
Confounder correction: None
Method category: Regression and data-driven disease progression model
Prediction method: Prediction of ADAS-Cog13 and Ventricles was made using linear regression using age, APOE status, gender and education as covariates. Different models were estimated for CN, MCI and AD subjects. For diagnosis prediction, an AD risk stage was calculated based on the Event-based probability (EBP) model chen2016staging . Prediction of clinical diagnosis was then made based on two approaches:

  • Chen-MCW-Std: Predict diagnosis based on AD stage as well as APOE4, gender and education using a Cox proportional hazards model.

  • Chen-MCW-Stratify: As above, but the model was further stratified based on AD risk stages, into low risk and high-risk.

Team: CN2L (Members: Ke Qi1, Shiyang Chen1,2, Deqiang Qiu1,2, Institutions: 1Emory University, 2Georgia Institute of Technology)
Feature selection: Automatic
Selected features: For the neural network, all features in D1/D2 are used; For the random forest, the main cognitive tests, MRI biomarkers (cross-sectional only), FDG, AV45, AV1451, DTI, CSF, APOE, demographics and clinical diagnosis are used.
Missing data: Imputation in a forward filled manner (i.e. using last available value).
Confounder correction: None
Method category: Machine learning
Prediction method:

  • CN2L-NeuralNetwork    : 3-layer recurrent neural network, 1024 units/layer. Dropout layers (dropout rate:0.1) were added to output and state connections to prevent overfitting. Adam method was used for training. Validation was performed using a leave-last-time-point-out approach.

  • CN2L-RandomForest   : Random forest method was used, where features of small importance for diagnosis prediction were filtered out. For the prediction of clinical diagnosis, an ensemble of 200 trees was trained on a class-balanced bootstrap sample of the training set. For the prediction of ADAS-Cog13 and Ventricles, an ensemble of 100 trees was used. Different predictions are made for different previous visits of a patient, and the final prediction is taken as the average of all predictions.

  • CN2L-Average   : The average of the above two methods.

Confidence Intervals: Confidence intervals are estimated based on probabilities output of the model.
Publication link: https://cds.ismrm.org/protected/18MPresentations/abstracts/3668.html Chen et al., ISMRM, 2018 chen2018is

Team: CyberBrains    (Members: Ionut Buciuman, Alex Kelner, Raluca Pop, Denisa Rimocea, Kruk Zsolt, Institution: Vasile Lucaciu College, Baia Mare, Romania)
Feature selection: Manual
Selected features: MRI volumes (Ventricles, middle temporal), ADAS-Cog13, APOE status
Missing data: For subjects with no ventricle measurements, authors computed an average value based on ADAS-Cog13 tests. This was used especially for D3 predictions.
Confounder correction: None
Method category: Regression
Prediction method: Fit a linear model of monthly difference in ventricle volume, as a function of ventricle volume, stratified by clinical diagnosis and ventricle volumes smaller and larger than 140,000 mm3. A similar model is applied for ADAS-Cog13 prediction, but stratified by APOE status and middle temporal volume smaller or greater than 16,000 mm3. Prediction of clinical diagnosis also used a linear model that was stratified based on the ADAS-Cog13, for ADAS-Cog13 ranges between 10 and 45. For ADAS-Cog13 greater than 45 and smaller than 10, pre-defined values were used for the probabilities of each diagnosis.



Team: DIKU    (Members: Mostafa Mehdipour Ghazi1,2,3,5, Mads Nielsen1,2,3, Akshay Pai1,2,3, Marc Modat4,5, M. Jorge Cardoso4,5, Sebastien Ourselin4,5, Lauge Søitalic-ø\orensen1,2,3; Institutions: 1Biomediq A/S, 2Cerebriu A/S, 3University of Copenhagen, Denmark, 4King’s College London, UK, 5University College London, UK)
Feature selection: Semi-automatic; linear discriminant analysis (LDA) was applied to select the top most-informative biomarkers, and ventricular volume and a few other MRI measures were subsequently manually added.
Selected features: Cognitive tests (CDR-SB, ADAS-11, ADAS-13, MMSE, FAQ, MOCA, RAVLT-Immediate, RAVLT-Learning, RAVLT-Percent-Forgetting), CSF measures (amyloid-beta, phosphorylated tau), MRI volumetric measures divided by ICV (ventricles, hippocampus, whole brain, entorhinal, fusiform, middle temporal).
Missing data: Method automatically deals with missing data.
Confounder correction: Linear transformation of age as part of the algorithm.
Method category: Data-driven disease progression model.
Prediction method: For predicting ADAS-13 and Ventricles, a data-driven disease progression model was used, which estimated a parametric trajectory for each selected feature over a common disease progression axis reflecting an estimated latent disease progression score (DPS). The chosen parametric function was generalised logistic function (Richard’s curve), and the DPS was a linear transformation of the age of subjects representing the subject-specific time shift and progression speed. The constrained fitting was performed alternating between estimation of subject-specific DPS transformations and global biomarker trajectories using L2-norm loss functions. The authors made three submissions:

  • DIKU-GeneralisedLog-Std: constrained, generalised logistic function for the trajectory model;

  • DIKU-ModifiedLog-Std: constrained sigmoid function for the trajectory model;

  • DIKU-ModifiedMri-Std: as above, but separately fitting MRI biomarkers for Ventricles prediction.

The above models were trained on D1 data only. Authors also made predictions from a custom training set (D1+D2 together), named DIKU-***-Custom. Clinical diagnosis was predicted based on the DPS scores using both a Bayesian classifier with likelihoods modeling using Gaussian mixture models, as well as an ensemble of LDAs. The final prediction was obtained through bagging of the two classifiers’ predictions. The whole method and a robust extension developed post-TADPOLE is described in ghazi2019robust .
Confidence Intervals: They were obtained by using bootstrapping via Monte Carlo resampling and evaluating the model performance assuming a Gaussian distribution.
Publication link: https://arxiv.org/abs/1908.05338



Team: DIVE    (Members: Razvan Marinescu, Institution: University College London, UK, Massachusetts Institute of Technology, USA)
Feature selection: Manual
Selected features: FDG, AV45, CDRSOB, ADAS-Cog13, MRI volumes (ventricles, hippocampus, whole brain, entorhinal, middle temporal), CSF (amyloid-beta, tau, phosphorylated tau)
Missing data: Method automatically deals with missing data
Confounder correction: None
Method category: Data-driven disease progression model
Prediction method:
For predicting the ADAS-Cog13 and Ventricle volume, the “Data-Driven Inference of Vertexwise Evolution” (DIVE) algorithm was used marinescu2019dive , which clusters the input biomarkers based on how similar their progression is over the disease time-course. While the original DIVE method was a spatio-temporal model, for TADPOLE it was applied on extracted features directly. The model estimates a parametric, sigmoidal trajectory of the biomarkers, which are a function of subjects’ disease progression scores (DPS), representing a linear transformation of their age. Subject-specific parameters included the latent time-shift and progression speed, as well as an intercept. For the prediction of clinical diagnosis, the posterior probability of each class was computed given the future DPS scores using non-parametric Kernel Density Estimators (KDE), fitted on the DPS scores for each diagnostic class independently. Code for the model is available online:
Confidence Intervals: The model estimates a variance parameter under a gaussian noise model, which was scaled accordingly to obtain the 50% confidence intervals for ADAS-Cog13 and Ventricle volume.
Publication link: https://www.sciencedirect.com/science/article/pii/S1053811919301491
Repository link: https://github.com/mrazvan22/dive


Team: EMC1    (Members: Vikram Venkatraghavan, Esther Bron, Stefan Klein, Institution: Erasmus MC, The Netherlands)
Feature selection: Automatic – Only the subjects who had converted to AD were used for feature selection. Features with the largest changes over time after correcting for age, gender, education and ICV were selected
Selected features: 250 features from the set of FDG, AV45, DTI, MRI (cross-sectional Freesurfer volumes), Arterial Spin Labelling (ASL) MRI, CSF and cognitive tests.
Missing data: Imputed using nearest-neighbour interpolation. For D2, visits with missing diagnosis were excluded. For the D3 subjects with no known diagnosis, this was estimated using a nearest-neighbour search based on disease severity
Confounder correction: Corrected for age, gender, education and ICV using linear regression based on data from controls.
Method category: Data-driven disease progression model and machine learning
Prediction method: Authors hypothesize that aging and progression of AD are the primary causes for the change in biomarker values with time and that these changes eventually lead to a change in clinical status. To predict biomarker values at future timepoints, the rate of AD progression is estimated in each subject. This is followed by estimating the interactions of aging and AD progression in the progression of different biomarkers. Lastly, authors use the biomarkers estimated at the future timepoint to predict the change in clinical status. These steps are elaborated below:
Rate of Progression of AD: To assess the severity of AD, we estimated the sequence in which the selected features became abnormal in AD using a Discriminative Event-Based Model venkatraghavan2019disease and used it to estimate the disease severity at all the timepoints for each subject. A linear mixed effect model was fit to estimate the rate of change of disease severity for different subjects. This model was used for predicting the disease severity at all the future timepoints.
Interactions of aging and AD progression: For predicting the biomarker values at the future timepoint, we fit linear mixed effect models for each biomarker considering interactions between the estimated disease severity and age, with gender and ICV as additional covariates. This model was used to forecast the future values of all 250 selected features, including ADAS-Cog13 scores and Ventricle volumes.
Predicting the change in clinical status: For the diagnosis prediction, the forecasted values of the biomarkers and the last known clinical diagnosis of the subject were used as inputs for a soft-margin SVM classifier with a radial basis function kernel. Two separate submissions were made:

  • EMC1-Std (ID 1): ASL based features were excluded in this model

  • EMC1-Custom (ID 2): ASL based features were included in this model

Confidence Intervals: Standard errors of the predicted values of Ventricles and ADAS-Cog-13 were estimated by repeating the prediction procedure, including the estimation of disease severity, for 10 repetitions of bootstrap sampling. These standard errors were used to define the confidence intervals.
Publication link: https://doi.org/10.1016/j.neuroimage.2018.11.024
Repository link: https://github.com/88vikram/TADPOLE_submission_with_debm


Team: EMC-EB    (Members: Esther E. Bron, Vikram Venkatraghavan, Stefan Klein, Institution: Erasmus MC, The Netherlands)
Feature selection: Automatic – For the D2 prediction, features were selected that had the largest change over time in subjects who converted to AD using corrections for age, gender, education and ICV, i.e. the same approach as EMC1. For the D3 prediction, features with less than 50% missing data were selected.
Selected features: 200 (D2 prediction) and 338 (D3 prediction) from the set of clinical diagnosis, cognitive tests, MRI volumes (Freesurfer cross-sectional), FDG PET, DTI measures (FA, MD, RD, AD) and CSF features.
Missing data: Imputation using nearest-neighbour interpolation based on the subject’s earlier timepoints. If not possible, imputation by the mean of training set was used. Visits with no clinical diagnosis were excluded for classifier training.
Confounder correction: None
Method category: Machine learning
Prediction method: All predictions were based on SVMs. For diagnosis and ADAS-Cog13 prediction, respectively a classifier with balanced class weights and a regressor were trained to predict the target measures at the next visit. These predictions do not explicitly take account of time, but assume that the times between two visits are roughly equal. For Ventricle prediction, a regressor was trained to predict the change of ventricle volume per year. Ventricle volumes were normalized using ICV at baseline (ICV at current time point for D3). Using the predicted change, the normalized ventricle volume at each future visit was computed. For all predictions, authors used a radial basis function (RBF) kernel SVM, of which the C-parameter was set to C=0.5𝐶0.5C=0.5 and gamma to the reciprocal of the number of features. All features were normalized to zero mean and unit standard deviation.
Confidence Intervals: Bootstrap resampling (n=100𝑛100n=100) of the training set.
Repository link: https://github.com/tadpole-share/tadpole-algorithms/tree/master/tadpole_algorithms/models/ecmeb


Team: FortuneTellerFish (Members: Alexandra Young, Institutions: University College London, UK, King’s College London, UK)
Feature selection: Manual
Selected features: Age at assessment, age at scan, APOE4 status, education, gender, MRI volumes (ventricles, hippocampus, whole brain, entorhinal, fusiform, middle temporal, all major lobes, insula, and basal ganglia). The probability of being amyloid positive, obtained from joint mixture modelling of CSF amyloid-beta and global AV45, was also included as a feature. Two key features, disease subtype and stage, were derived from the Subtype and Stage Inference (SuStaIn) model based on the MRI features young2018uncovering .
Missing data: Imputed by averaging over the k-nearest neighbours with k=5𝑘5k=5
Confounder correction: Brain volumes were corrected for age, intracranial volume and field strength using linear regression. Parameters for the linear regression were estimated based on amyloid-negative controls.
Method category: Data-driven disease progression models + statistical regression
Prediction method: For ADAS-Cog13 and Ventricle prediction, a linear mixed effects model was used which used a different set of fixed/random effects, depending on the submission:

  • FortuneTellerFish-Control   : For Ventricles, fixed effects were age at scan and gender. For ADAS-Cog13 and MMSE the fixed effects were age at scan, education, APOE status and amyloid positivity. For all target measures, there was one random effect per subject.

  • FortuneTellerFish-SuStaIn   : two additional fixed effects from the SuStaIn model: subtype and stage.

For the prediction of clinical diagnosis, a multiclass error-correcting output codes (ECOC) classifier based on SVMs was trained with the following inputs: age at assessment, age at scan, APOE status, amyloid positivity, gender, education, SuStaIn subtype and stage, ADAS-Cog13, MMSE and ventricle volume. For diagnosis prediction at future timepoints, the forecasted values for ADAS-Cog13, MMSE and Ventricle volume were used as input to the classifier.


Team: Frog    (Members: Keli Liu, Christina Rabe, Paul Manser Institution: Genentech, USA)
Feature selection: Automatic using the Xgboost package chen2016xgboost
Selected features: Cognitive tests (ADAS-Cog13, CDRSB, MMSE, RAVLT), clinical diagnosis, MRI measurements, FDG PET measurements, APOE status and CSF measurements. For each longitudinal measurement (e.g. test scores and MRI), the following transformations were computed and used to augment the original feature set: most recent measurement, time since most recent measurement, the historical highest (lowest) measurement, time since the historical highest and lowest measurement, and the most recent change in measurement.
Missing data: Xgboost package automatically deals with missing data through inference based on reduction of training loss.
Confounder correction: None
Method category: Statistical prediction using regression
Prediction method: Flexible models and features were chosen automatically using gradient boosting (Xgboost package). Different models were trained for the following forecast windows: 0-8 months, 9-15, 16-27, 28-39, 40-60, >>60 (given windows are for clinical status prediction, slightly different windows used for ADAS-Cog13 and ventricular volume prediction) . Variable importance scores from Xgboost suggest that MRI features play a bigger role in models trained for longer forecast windows.
Confidence Intervals: Standard deviation for prediction error was estimated based on cross validation (on training set). Normality of prediction errors was then assumed to construct prediction intervals based on estimated standard deviation.



Team: GlassFrog-LCMEM-HDR    (Members: Steven Hill1, James Howlett1, Robin Huang1, Steven Kiddle1, Sach Mukherjee2, Anaïs Rouanet1, Bernd Taschler2, Brian Tom1, Simon White1, Institutions: 1MRC Biostatistics Unit, University of Cambridge, UK; 2German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany)
Feature selection:

  • MSM: Automatic, by selecting features which passed a likelihood ratio test when compared against a model with no covariates.

  • LCMEM: Manual

  • HDR: Automatic, selected via sparse Lasso regression.

Selected features:

  • MSM: D2 - gender, age, education, ADAS-Cog13, diagnosis, MMSE, CDRSOB, APOE status, first 5 principal components from imaging, amyloid positivity, tau level. D3 - gender, age, education, ADAS-Cog13, diagnosis, MMSE, Ventricles/ICV

  • LCMEM: ADAS-Cog13, gender, education, age at baseline

  • HDR: all features were provided to the method, excluding some features with many missing values

Missing data:

  • MSM: Filling with last known value, or nearest neighbour if feature was never observed.

  • LCMEM: Complete case analysis (assumption that missing data are missing at random)

  • HDR: Imputation using within-subject interpolation and nearest neighbour matching

Confounder correction: None
Method category: Combination of statistical regression and data-driven disease progression models
Prediction method: The prediction of clinical diagnosis was done using a Multi-State Model (MSM). Multi-state models (MSMs) kalbfleisch1985analysis are continuous-time Markov chain models, here with states corresponding to CN, MCI, AD, and transition rates estimated from the data using covariates selected as described above. The model accounts for noise in the historical diagnostic labels. Predictions for a given forecast month were made using the last observed disease state and associated covariates.

Prediction of ADAS-Cog13 was done using a Latent class mixed effects model (LCMEM). The model used four latent classes, where class membership probability was modelled via a multinomial logistic. For each latent class a specific linear mixed effects model defined a Gaussian latent process, with class-specific fixed effects, random effects for intercept, slope and square slope, and Gaussian noise. Age at baseline, gender and education were also included as covariates. Finally, a Beta cumulative distribution link function was used (and estimated simultaneously) between ADAS-Cog 13 and the latent process, to account for the departure from the Gaussian assumption on the outcome. The number of latent classes was optimised with the Bayesian Information Criterion (BIC).

Prediction of Ventricle volume was done using high-dimensional regression (HDR) and disease state-specific slope models: Subject-specific slopes were obtained by a combination of Lasso regression and shrinkage towards disease-state-specific shrinkage targets. Conversion times, from one disease state to another, were forecasted using the MSM model.
Confidence Intervals: LCMEM: the 50% confidence intervals for ADAS-Cog13 were obtained using a bootstrap approach. HDR: Confidence intervals were set as percentages of the predicted values.



Team: GlassFrog-SM    (members and affiliations as above)
Feature selection: Manual
Selected features: ADAS-Cog13, Ventricles/ICV, age at visit, APOE status, education, diagnosis
Missing data: For training, complete case analysis was performed. For prediction, imputation was performed for missing outcomes using a linear model with age, education, diagnosis and APOE status as covariates.
Confounder correction: None
Method category: Combination of statistical regression and data-driven disease progression models
Prediction method: The prediction of clinical diagnosis was using MSM models as in GlassFrog-LCMEM-HDR. The prediction for ADAS-Cog13 and Ventricles used a Slope Model (SM), which used a quadratic function to model the slope of the outcome variable as a function of the current outcome value and covariates. Covariates used were age at visit, education and APOE status.
Confidence Intervals: SM: Confidence intervals were set as percentages of the predicted values. The percentages used were manually selected and depended on the missingness of covariates for each individual.


Team: GlassFrog-Average    (members and affiliations as above)
Prediction method: The prediction of clinical diagnosis was using MSM models as in GlassFrog-LCMEM-HDR. For ADAS-Cog13 and Ventricles an ensemble approach was used that averaged the predictions from three methods: LCMEM, HDR and SM, as described above. Confidence interval bounds were also averaged.

Team: IBM-OZ-Res (Members: Noel Faux, Suman Sedai, Institution: IBM Research Australia, Melbourne, Australia)
Feature selection: using boosting regression
Selected features: Ventricle volume, AV45, FDG PET, cognitive tests, clinical diagnosis, age
Missing data: Imputed with zero; observations with missing ventricle volume are dropped.
Confounder correction: None
Method category: Machine Learning
Prediction method: A stochastic gradient boosting regression machine (GBM) was used to predict Ventricle Volume, with a Huber loss function. To reduce overfitting, a shrinkage mechanism was adopted, where the response of each tree is reduced by a factor of 0.01. Independent predictions were made for each individual visit, and averaged when a subject had more than one visit. For the prediction of clinical status, a similar GBM model was adopted, but with a multinomial deviance loss function.


Team: ITESMCEM    (Members: Javier de Velasco Oriol1, Edgar Emmanuel Vallejo Clemente1, Karol Estrada2. Institution: 1Instituto Tecnológico y de Estudios Superiores de Monterrey, 2Brandeis University)
Feature selection: Manual
Selected features: Demographics, MRI volumes, FDG PET and all cognitive tests
Missing data: Imputation using the mean of previous values of that patient, otherwise mean across all patients.
Confounder correction: None
Method category: Machine learning
Prediction method: ADAS-Cog13 is predicted with a Lasso model with α=0.1𝛼0.1\alpha=0.1. Ventricles are predicted using a Bayesian ridge regression. Clinical diagnosis is predicted using two different Random Forest models, one which selected between CN and either MCI or AD and the second which in turn predicted between MCI and AD for those selected as non-CN. The predictions for all target variables made use of a “transition model”, which predicted the next timepoint given the current one, until all 60 monthly predictions were made. The transition model was implemented using a total of 29 Lasso models.
Confidence Intervals: They were calculated by sampling the test samples multiple times, evaluating the performance of the model with those samples and analyzing the CIs supossing a Gaussian distribution and the corresponding t-distribution.


Team: lmaUCL    (Members: Leon Aksman, Institution: University College London, UK)
Feature selection: Manual
Selected features: Diagnosis, gender, education, APOE4 status and MMSE.
Missing data: Imputation using regression over ventricles and demographics (age, gender, education)
Confounder correction: None
Method category: Statistical regression and machine learning
Prediction method: For ADAS-Cog13 and Ventricles, a multi-task learning model was used with similar trajectories across subjects. The regression model was a linear model over age, but dependencies between different subjects were modelled through a special prior structure over the coefficients of the linear model. The prior structure has hyperparameters that control for the amount of coupling across subjects, and are optimised through empirical Bayes. Clinical diagnosis was predicted using the ADAS-Cog13 trajectory estimates plus a simple estimate of the mean and standard deviation of ADAS-Cog13 in each diagnostic group (AD/MCI/CN). Each ADAS-Cog13 prediction was then assigned a probability of belonging to each group. Three different submissions were made:

  • lmaUCL-Std: used only last available diagnosis as covariate

  • lmaUCL-Covariates: as above, but also used gender, education, APOE status and MMSE as covariates

  • lmaUCL-halfD1: trained only on half of the D1 dataset, but allowed for longer training time of 10 hours.

Confidence Intervals: Both the multi-task learning and clinical diagnosis models are probabilistic, providing estimates of predictive mean and standard deviation assuming a normal distribution. These were converted to confidence intervals using the inverse normal CDF.
Publication link: https://doi.org/10.1002/hbm.24682
Repository link: https://github.com/LeonAksman/bayes-mtl-traj

Team: Mayo-BAI-ASU    (Members: Cynthia M. Stonnington1, Yalin Wang2, Jianfeng Wu2, Vivek Devadas3, Institution: 1Mayo Clinic, Scottsdale, AZ, USA, 2School of Computing, Informatics and Decision Systems Engineering, Arizona State University, USA, 3Banner Alzheimer’s Institute, Phoenix, AZ, USA)
Feature selection: Manual, from clinical experience
Selected features: Age, PET(AV45, AV1451, FDG), hippocampal volume/ICV, ventricle volume/ICV, diagnosis, cognitive tests (ADAS11, ADAS-Cog13, MMSE, RAVLT, MOCA, Ecog), amyloid-beta, tau, phosphorylated tau, APOE status. All features except age were z-score normalised.
Missing data: Imputation with zero.
Confounder correction: None
Method category: Statistical regression
Prediction method: ADAS-Cog13 and Ventricles were forecasted using a linear mixed effects model, using all features as fixed effects and one random effect per subject (intercept). Training used all visits, but the forecasts only used the last visit. Clinical diagnosis was predicted with a similar model, by converting to CN/MCI/AD to a categorical variable (1/2/3).


Team: Orange    (Members: Clementine Fourrier, Institution: Institut du Cerveau et de la Moelle épinière, ICM, Paris, France)
Feature selection: Manual, through knowledge from literature
Selected features: demographics (age, education, gender), cognitive tests (ADAS11, CDSRB, MMSE), imaging(AV45 PET, FDG PET, hippocampus size, cortical thickness) and molecular markers (phosphorylated tau to amyloid-beta ratio, total tau, CMRgl, HCI)
Method category: Decision tree of a clinician
Prediction method: This method is based on the decision tree of a clinician. It looks at the latest available visit for a patient, and based on the value of the selected features, it predicts a duration to conversion. The duration to conversion depends on the initial clinical diagnosis and the other available data. Depending on the initial diagnosis, the algorithm assumes the patient will convert within a certain time period, and this time period is modulated by the available data about the patient. In that regard, the algorithm does not need to account for missing values. The ADAS-Cog13 and Ventricle measures at each month are computed assuming a linear evolution between the current time point and the conversion date.


Team: Rocket    (Members: Lars Lau Raket, Institutions: H. Lundbeck A/S, Denmark; Clinical Memory Research Unit, Department of Clinical Sciences Malmö, Lund University, Lund, Sweden )
Feature selection: Manual
Selected features: ADAS-Cog13, baseline diagnosis, and APOE4 carrier status.
Missing data: APOE4: Imputed using median number of alleles per baseline diagnostic group. Missing ADAS-Cog13 scores were imputed using multivariate imputation by chained equations based on age, sex, diagnosis and cognitive tests.
Confounder correction: None
Method category: Statistical regression and data-driven disease progression modelling
Prediction method: Prediction of ADAS-Cog13 is done through a latent-time non-linear mixed-effects model, where the trajectory is parameterised using an exponential function. The time shift is built from a fixed effect shift relative to time since baseline for different diagnostic groups (e.g. AD patients will be shifted to be later in the course of cognitive decline) and a random effect shift for each subject. The model also includes APOE status as a fixed effect that modifies rate of decline. This disease progression modeling methodology along with several extensions is presented in raket2019disease . Prediction of Ventricles/ICV uses a linear mixed-effects model using an integrated B-spline basis (5 knots + intercept) in predicted ADAS-Cog13 disease time. A random intercept is included per subject. Prediction of clinical diagnosis is based on kernel density estimation of the states (CN/MCI/AD) across the disease time from the ADAS-Cog13 model.
Confidence Intervals: Prediction intervals conditioned on the predicted disease time of the subject were derived based on the estimated variance-covariance matrix of the model. Because of the monotone nature of cognitive decline, and to heuristically compensate for the conditioning on disease time, the upper limit of the prediction interval was multiplied by 1.5.
Publication link: https://doi.org/10.1101/2019.12.13.19014860

Team: SBIA    (Members: Aristeidis Sotiras, Guray Erus, Jimit Doshi, Christos Davatzikos, Institution: Center for Biomedical Image Computing and Analytics, University of Pennsylvania)
Feature selection: Manual
Selected features: demographics, cognitive tests, diagnosis, MRI features (Freesurfer cross-sectional). Imaging indices (SPARE-AD and SPARE-MCI) that summarise brain atrophy patterns were estimated through support vector machines with linear kernels. Another index representing brain age (SPARE-BA) was estimated using a regressor model applied to imaging features.
Missing data: Features and time-points with missing data were not included
Confounder correction: Age, gender, APOE4, education, and SPARE scores were used as covariates in the linear mixed effects models. Also, the regression model was applied separately on different diagnosis groups.
Method category: Statistical regression and machine learning
Prediction method: A linear mixed effects model was used to forecast the SPARE indices for future timepoints. For diagnosis predictions, authors used class probability distribution estimations based on the forecasted SPARE-AD score. For the prediction of ADAS-Cog13 and Ventricles, linear mixed effects models were used, with age, gender and SPARE scores as covariates.


Team: SmallHeads – BigBrains (Members: Jacob Vogel, Andrew Doyle, Angela Tam, Alex Diaz-Papkovich, Institution: McGill University, Montreal, Canada)
Feature selection: Automatic
Selected features:

  • SmallHeads-NeuralNetwork   : All features in the TADPOLE spreadsheet were considered, as long as they had less than 50% missing data. This resulted in a final set of 376 features. Features were normalised to zero mean and unit variance.

  • SmallHeads-LinMixedEffects   : All features were normalised to zero mean and unit variance. A LASSO feature selection algorithm with ADAS-Cog13 and “Y” variable was used to select the best features, which were required to have a weight greater than 0.001. 10-fold cross-validation was used to estimate the best LASSO parameters.

Missing data: Imputed using 5-nearest neighbour method, using Euclidean distance (FancyImpute 0.0.4)
Confounder correction: None
Method category: Machine learning
Prediction method:

  • SmallHeads-NeuralNetwork: A Deep fully connected neural network was trained to predict the future diagnosis using the selected features and time until future timepoint as input. Network has 5 fully-connected layers with Leaky ReLU activations. Each layer has 512, 512, 1024, 1024 and 256 neurons, with softmax layer at output. Training used P(0.5) dropout using the Adam optimiser, based on a class-unweighted categorical cross-entropy loss function.

  • SmallHeads-LinMixedEffects: Only ADAS-Cog13 was predicted with a linear mixed effects model, using months since baseline as an interaction term. A random (subject — time) effect was also added, allowing variable subject-specific slopes over time.

Repository link: https://github.com/SmallHeads/tadpole


Team: SPMC-Plymouth    (Members: Emmanuel Jammeh, Institution: University of Plymouth, UK)
Feature selection: Automatic – Authors used the WEKA (https://www.cs.waikato.ac.nz/ml/weka/) machine learning tool.
Selected features: Age, gender, education, ApoE4, CDRSB, ADAS11, MMSE, RAVLT, Moca, Ecog, Hippocampus, WholeBrain, Entorhinal, MidTemp, FDG, AV45, PIB, ABETA, TAU, PTAU
Missing data:
Confounder correction:
None
Method category: Machine learning
Prediction method: A machine learning classifier based on k-nearest neighbours, Naive Bayes, Random Forest and SVM was used to predict the clinical diagnosis. ADAS-Cog13 and Ventricle volume were not predicted. Two predictions were made, SPMC-Plymouth1 and SPMC-Plymouth2, but the authors could not be contacted to provide details on the differences between the two.

Team: Sunshine    (Members: Igor Koval, Stanley Durrleman, Institution: Institut du Cerveau et de la Moelle épinière, ICM, Paris, France)
Feature selection: Semi-automatic – An initial set of 60 features was selected by a clinical expert. Out of this set, the features that had more than 30% missing data were removed. A final subset of features was chosen based on cross-validation results using trial and error.
Selected features: Age, APOE status, MMSE, ADAS-Cog13, RAVLT immediate and CDRSOB
Missing data: Imputed using mean value
Confounder correction: None
Method category: Statistical regression and machine learning
Prediction method: A linear model was used to predict future values of ADAS-Cog13 and Ventricles, as well as other cognitive tests: MMSE, RAVLT and CDRSOB. For the prediction of clinical diagnosis, forecasted values of the previous five measures were used as input to an SVM. APOE4 status and education were also used as inputs to the SVM. Adding extra features did not seem to increase prediction scores based on cross-validation. Two submissions were made:

  • Sunshine-Conservative: CN and AD subjects were forecasted to have the same diagnosis (i.e. no conversion) for all future timepoints, after observing that a small proportion of them convert after 1 year.

  • Sunshine-Std: Without the above modification.

Team: Threedays    (Members: Paul Moore1, Terry J. Lyons1, John Gallacher2, Institution: 1Mathematical Institute, University of Oxford, 2Department of Psychiatry, University of Oxford, UK)
Feature selection: Manual
Selected features: Age, months since baseline, gender, race, marital status, diagnosis, cognitive tests (MMSE, CDRSB, ADAS11, ADAS-Cog13, RAVLT immediate, learning, forgetting and percent forgetting, FAQ) and APOE status.
Missing data: Random forest method deals with missing data automatically, by finding optimal splits with existing data only.
Confounder correction: None
Method category: Machine learning
Prediction method: For the prediction of clinical diagnosis, two random forest models are trained, the first for transitions from a healthy diagnosis, and the second for transitions from an MCI diagnosis. For AD individuals, authors assume that the diagnosis will not change. The training data was generated by ordering each participant’s data by time, then associating the feature vector x with diagnosis y for each time horizon available for the participant. ADAS-Cog13 and Ventricles were not predicted. The PLOS paper describes a method similar to the original, but using different predictors and a single random forest.
Publication link: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0211558


Team: Tohka-Ciszek (Members: Jussi Tohka, Robert Ciszek Institution: A.I. Virtanen Institute for Molecular Sciences, University of Eastern Finland, Finland)
Feature selection: Manual
Selected features:

  • D2: diagnosis, gender, education, race, marital and APOE status, age, cognitive tests (CDRSB, ADAS11, ADAS-Cog13, MMSE, RAVLT learning, immediate and perc. forgetting, FAQ, MOCA, all Ecog), MRI volumes (ventricles, hippocampus, whole brain, entorhinal, fusiform, middle temporal, ICV)

  • D3: diagnosis, age, gender, education, ethnicity, race, marital status, ADAS-Cog13, MMSE, MRI volumes as above

Missing data:

  • SMNSR   : A sub-model is trained for each subset of features which occurs without missing values. A specific catch-all subset is used for patients for which well performing whole measurement set cannot be found. For this data set, values are imputed using the median of k-nearest neighbours or replaced with -1.

  • RandomForestLin   : Imputation using mean values from the timepoints of the same subject (D2) or diagnostic category (for D3).

Confounder correction: None
Method category: Machine learning
Prediction method:

  • Tohka-Ciszek-SMNSR   : For ADAS-Cog13 prediction, a Sparse Multimodal Neighborhood Search Regression was used. This method first uses a linear regression model to estimate ADAS-Cog13 from the selected features belonging to the current subject and neighbour subjects, estimated based on a K-nearest neighbour algorithm. The forecasts from this model were passed to a gradient-boosted tree model, providing the final prediction. Ventricles and clinical diagnosis were not predicted.

  • Tohka-Ciszek-RandomForestLin   : To predict ADAS-Cog13 and Ventricles, a weighted average of two models was used: 1) a unimodal linear model 2) a linear model taking the response variable from the final time point and predictor variables from a time point before that. For diagnosis prediction, a random forest was trained using ADAS-Cog13, Ventricle/ICV, age and APOE status.

Confidence Intervals: Predicted score +/- cross-validation MAE.
Repository link: https://github.com/jussitohka/tadpole(RandomForestLin),https://github.com/rciszek/SMNSR (SMNSR)

Team: VikingAI    (Members: Bruno Jedynak1, Kruti Pandya1, Murat Bilgel2, William Engels1, Joseph Cole1, Institutions: 1Portland State University, USA, 2Laboratory of Behavioral Neuroscience, National Institute on Aging, National Institutes of Health, Baltimore, MD, USA)
Feature selection: Manual
Selected features: Diagnosis, age, ADAS-Cog13, CDRSOB, MMSE, RAVLT immediate, tau,
ventricles/ICV, hippocampus/ICV. Features were normalized prior to model fitting.
Missing data: Method automatically deals with missing data through Bayesian inference
Confounder correction: None
Method category: Data-driven disease progression model
Prediction method: For the prediction of ADAS-Cog13 and Ventricles, a latent-time parametric model was used, estimating a linear subject-specific model over the age of subjects, resulting in a disease progression score (DPS). The trajectories of biomarkers were assumed to be either sigmoidal functions or a sum of logistic basis functions over the DPS space. Some feature-specific parameters are estimated from the features’ histograms, while the rest are optimized. Priors over the parameters to be optimized are set a-priori. Two submissions were made:

  • VikingAI-Sigmoid: sigmoidal function as biomarker trajectory

  • VikingAI-Logistic: the sum of 15 logistic basis functions as the biomarker trajectory

Confidence Intervals: Bayesian predictive intervals

In addition to the above entries, the organisers also included several benchmark algorithms as well as some extra predictions: (1) two ensemble predictions averaging all the predictions submitted by the participants, and (2) 62 randomised predictions which can tell how likely it is that top submissions obtained their scores due to chance, by building a null distribution of the performance metrics. The
Source code of some benchmarks (BenchmarkLastVisit, BenchmarkMixedEffectsAPOE, BenchmarkSVM) was offered to participants before the conference deadline, as a starting point for making predictions.

Benchmark: BenchmarkLastVisit    (Authors: Daniel Alexander, Razvan Marinescu, Institutions: University College London, UK, Massachusetts Institute of Technology, USA)
Feature selection: None
Selected features: ADAS-Cog13, ventricle volume, diagnosis
Missing data: Not required
Confounder correction: None
Method category: Regression
Prediction method: For ADAS-Cog13 and Ventricles, the last available measure is used, otherwise the average for the current diagnostic group is used. Confidence intervals are set to default widths of 2 for ADAD13 and 0.001 for Ventricles/ICV. For prediction of clinical diagnosis, the last available diagnosis is used with probability 100%, and 0% probability for the other diagnoses.
Confidence Intervals: hard-coded
Repository link: https://github.com/noxtoby/TADPOLE/blob/master/evaluation

Benchmark: BenchmarkMixedEffects    (Author: Razvan Marinescu, Daniel Alexander, Institution: University College London, UK, Massachusetts Institute of Technology, USA)
Feature selection: None
Selected features: ADAS-Cog13, ventricle volume, diagnosis (, APOE status)
Missing data: Automatic, since model is univariate.
Confounder correction: APOE status was used as covariate in the linear mixed effects model
Method category: Regression
Prediction method: Linear Mixed Effects Model with age at visit as the predictor variable. Model was fitted independently for ADAS-Cog13 and Ventricles. Predictions for clinical diagnosis were derived from the corresponding ADAS-Cog13 forecasts, using three Gaussian likelihood models for CN, MCI and AD. The likelihoods for diagnostic classes were finally converted to probabilities by normalisation. Default values were used for confidence intervals. Two predictions were made:

  • BenchmarkMixedEffectsAPOE: the slope of the population trajectory was stratified by APOE status

  • BenchmarkMixedEffects: as above but without APOE

Confidence Intervals: hard-coded
Repository link: https://github.com/noxtoby/TADPOLE/blob/master/evaluation

Benchmark: BenchmarkSVM    (Author: Esther Bron, Institution: Erasmus MC)
Feature selection: Manual
Selected features: Diagnosis, age, ADAS-Cog13, Ventricles, ICV, APOE
Missing data: Fill-in using average value of biomarker from past visits of the same subject, otherwise population average.
Confounder correction: None
Method category: Machine Learning
Prediction method: For the prediction of clinical diagnosis, a probabilistic SVM was used based on the selected features, while for the prediction of ADAS-Cog13 and Ventricles, a Support Vector Regressor (SVR) was used. All SVM/SVRs used linear kernels. Default values were used for confidence intervals.
Confidence Intervals: hard coded
Repository link: https://github.com/noxtoby/TADPOLE/blob/master/evaluation

Benchmark: RandomisedBest    (Author: Razvan Marinescu, Institutions: University College London, UK, Massachusetts Institute of Technology, USA)
Feature selection: Manual
Selected features: Diagnosis, age, ADAS-Cog13, Ventricles, ICV
Missing data: Fill-in using last available measurement
Confounder correction: None
Method category: Regression
Prediction method: The method aims to construct a null distribution of values, to check how likely a high score could be obtained by chance alone. Starting from the simplest prediction method, i.e. BenchmarkLastVisit which simply takes the last available measure, 62 randomised predictions were created (as many as the total number of predictions in D2) by adding random perturbations to the predictions. For Diagnosis prediction, the probability of controls and MCI subjects to convert within 1 year was computed from ADNI historical data, and then each control and MCI was randomly chosen to convert with those probabilities. For ADAS-Cog13 and Ventricles, random uniform noise was added to the predictions as follows:

  • ADAS: new_adas similar-to\sim last_available_adas + U(0, 7)

  • Ventricles: new_ventricles similar-to\sim last_available_ventricles + U(0, 0.01)

The new values new_adas and new_ventricles, as well as the new diagnosis, were assigned to all 60 months, thus assuming no change across the 60 month predictions. All 62 different predictions were evaluated, and the RandomisedBest entry shows the best scores obtained by all 62 submissions in each category separately.
Confidence Intervals: hard-coded
Repository link: https://github.com/noxtoby/TADPOLE/blob/master/evaluation

6 Acknowledgements

TADPOLE Challenge has been organised by the European Progression Of Neurological Disease (EuroPOND) consortium, in collaboration with the ADNI. We thank all the participants and advisors, in particular Clifford R. Jack Jr. from Mayo Clinic, Rochester, United States and Bruno M. Jedynak from Portland State University, Portland, United States for useful input and feedback.

The organisers are extremely grateful to The Alzheimer’s Association, The Alzheimer’s Society and Alzheimer’s Research UK for sponsoring the challenge by providing the £30,000 prize fund and providing invaluable advice into its construction and organisation. Similarly, we thank the ADNI leadership and members of our advisory board and other members of the EuroPOND consortium for their valuable advice and support.

RVM was supported by the EPSRC Centre For Doctoral Training in Medical Imaging with grant EP/L016478/1 and by the Neuroimaging Analysis Center with grant NIH NIBIB NAC P41EB015902. NPO, FB, SK, and DCA are supported by EuroPOND, which is an EU Horizon 2020 project. ALY was supported by an EPSRC Doctoral Prize fellowship and by EPSRC grant EP/J020990/01. PG was supported by NIH grant NIBIB NAC P41EB015902 and by grant NINDS R01NS086905. DCA was supported by EPSRC grants J020990, M006093 and M020533. FB was supported by the NIHR UCLH Biomedical Research Centre and the AMYPAD project, which has received support from the EU-EFPIA Innovative Medicines Initiatives 2 Joint Undertaking (AMYPAD project, grant 115952). EEB was supported by the Dutch Heart Foundation (PPP Allowance, 2018B011) and Medical Delta Diagnostics 3.0: Dementia and Stroke. The UCL-affiliated researchers received support from the NIHR UCLH Biomedical Research Centre. This project has received funding from the European Union Horizon 2020 research and innovation programme under grant agreement No 666992. Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012).

The work of VikingAI was facilitated in part by the Portland Institute for Computational Science and its resources acquired using NSF Grant DMS 1624776. MB was supported by the Intramural Research Program of the National Institute on Aging, NIH. BTTY, NS and MN were supported by the Singapore National Research Foundation (NRF) Fellowship (Class of 2017). Team SBIA was supported by grant R01 AG054409. Team DIKU has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skodowska-Curie grant agreement No 721820.

7 Author Contributions

DCA, RVM, NPO, ALY, EEB and SK designed the challenge. RVM performed the evaluation of the algorithms and the analysis of the results, and drafted the manuscript. All authors provided feedback on the manuscript. RVM, NPO, ALY, EB and DCA created the TADPOLE website. RVM and NPO constructed the TADPOLE D1-D4 datasets. RVM, DCA and EB wrote benchmark scripts which were offered to participants before the deadline. DCA and NPO organised and ran the TADPOLE webinars. NF and FB provided valuable suggestions on the challenge design. AT and MW offered access to the ADNI database. All other co-authors participated in the challenge. EB lead the TADPOLE-SHARE effort to make the algorithms openly available for further reuse, in a standardised format.

8 Competing Interests Statement

SJK received fees for participation in a Roche Diagnostics advisory board and does paid consulting for DIADEM outside of the submitted work.

References

  • [1] Dev Mehta, Robert Jackson, Gaurav Paul, Jiong Shi, and Marwan Sabbagh. Why do trials for Alzheimer’s disease drugs keep failing? A discontinued drug perspective eps for 2010-2015. Expert opinion on investigational drugs, 26(6):735–739, 2017.
  • [2] Rachael I Scahill, Jonathan M Schott, John M Stevens, Martin N Rossor, and Nick C Fox. Mapping the evolution of regional atrophy in Alzheimer’s disease: unbiased analysis of fluid-registered serial MRI. Proceedings of the National Academy of Sciences, 99(7):4703–4707, 2002.
  • [3] Mert R Sabuncu, Rahul S Desikan, Jorge Sepulcre, Boon Thye T Yeo, Hesheng Liu, Nicholas J Schmansky, Martin Reuter, Michael W Weiner, Randy L Buckner, Reisa A Sperling, et al. The dynamics of cortical and hippocampal atrophy in Alzheimer disease. Archives of neurology, 68(8):1040–1048, 2011.
  • [4] Rachelle S Doody, Valory Pavlik, Paul Massman, Susan Rountree, Eveleen Darby, and Wenyaw Chan. Erratum to: Predicting progression of Alzheimer’s disease. Alzheimer’s research & therapy, 2(3):14, 2010.
  • [5] Stefan Klöppel, Cynthia M Stonnington, Carlton Chu, Bogdan Draganski, Rachael I Scahill, Jonathan D Rohrer, Nick C Fox, Clifford R Jack Jr, John Ashburner, and Richard SJ Frackowiak. Automatic classification of MR scans in Alzheimer’s disease. Brain, 131(3):681–689, 2008.
  • [6] Daoqiang Zhang, Yaping Wang, Luping Zhou, Hong Yuan, Dinggang Shen, Alzheimer’s Disease Neuroimaging Initiative, et al. Multimodal classification of Alzheimer’s disease and mild cognitive impairment. Neuroimage, 55(3):856–867, 2011.
  • [7] Jonathan Young, Marc Modat, Manuel J Cardoso, Alex Mendelson, Dave Cash, Sebastien Ourselin, Alzheimer’s Disease Neuroimaging Initiative, et al. Accurate multimodal probabilistic prediction of conversion to Alzheimer’s disease in patients with mild cognitive impairment. NeuroImage: Clinical, 2:735–745, 2013.
  • [8] Jussi Mattila, Juha Koikkalainen, Arho Virkki, Anja Simonsen, Mark van Gils, Gunhild Waldemar, Hilkka Soininen, Jyrki Lötjönen, Alzheimer’s Disease Neuroimaging Initiative, et al. A disease state fingerprint for evaluation of Alzheimer’s disease. Journal of Alzheimer’s Disease, 27(1):163–176, 2011.
  • [9] Hubert M Fonteijn, Marc Modat, Matthew J Clarkson, Josephine Barnes, Manja Lehmann, Nicola Z Hobbs, Rachael I Scahill, Sarah J Tabrizi, Sebastien Ourselin, Nick C Fox, et al. An event-based model for disease progression and its application in familial Alzheimer’s disease and Huntington’s disease. NeuroImage, 60(3):1880–1889, 2012.
  • [10] Alexandra L Young, Neil P Oxtoby, Pankaj Daga, David M Cash, Nick C Fox, Sebastien Ourselin, Jonathan M Schott, and Daniel C Alexander. A data-driven model of biomarker changes in sporadic Alzheimer’s disease. Brain, 137(9):2564–2577, 2014.
  • [11] Bruno M Jedynak, Andrew Lang, Bo Liu, Elyse Katz, Yanwei Zhang, Bradley T Wyman, David Raunig, C Pierre Jedynak, Brian Caffo, Jerry L Prince, et al. A computational neurodegenerative disease progression score: method and results with the Alzheimer’s disease Neuroimaging Initiative cohort. Neuroimage, 63(3):1478–1486, 2012.
  • [12] Michael C Donohue, Hélène Jacqmin-Gadda, Mélanie Le Goff, Ronald G Thomas, Rema Raman, Anthony C Gamst, Laurel A Beckett, Clifford R Jack Jr, Michael W Weiner, Jean-François Dartigues, et al. Estimating long-term multivariate progression from short-term data. Alzheimer’s & Dementia, 10(5):S400–S410, 2014.
  • [13] Marco Lorenzi, Maurizio Filippone, Giovanni B Frisoni, Daniel C Alexander, Sébastien Ourselin, Alzheimer’s Disease Neuroimaging Initiative, et al. Probabilistic disease progression modeling to characterize diagnostic uncertainty: application to staging and prediction in Alzheimer’s disease. NeuroImage, 2017.
  • [14] Neil P Oxtoby, Alexandra L Young, David M Cash, Tammie LS Benzinger, Anne M Fagan, John C Morris, Randall J Bateman, Nick C Fox, Jonathan M Schott, and Daniel C Alexander. Data-driven models of dominantly-inherited Alzheimer’s disease progression. Brain, 141(5):1529–1544, 2018.
  • [15] Jean-Baptiste Schiratti, Stéphanie Allassonnière, Olivier Colliot, and Stanley Durrleman. A bayesian mixed-effects model to learn trajectories of changes from repeated manifold-valued observations. Journal of Machine Learning Research, 18(133):1–33, 2017.
  • [16] Murat Bilgel, Jerry L Prince, Dean F Wong, Susan M Resnick, and Bruno M Jedynak. A multivariate nonlinear mixed effects model for longitudinal image analysis: Application to amyloid imaging. Neuroimage, 134:658–670, 2016.
  • [17] Răzvan V Marinescu, Arman Eshaghi, Marco Lorenzi, Alexandra L Young, Neil P Oxtoby, Sara Garbarino, Sebastian J Crutch, Daniel C Alexander, Alzheimer’s Disease Neuroimaging Initiative, et al. Dive: A spatiotemporal progression model of brain pathology in neurodegenerative disorders. NeuroImage, 192:166–177, 2019.
  • [18] Igor Koval, Alexandre Bône, Maxime Louis, Simona Bottani, Arnaud Marcoux, Jorge Samper-Gonzalez, Ninon Burgos, Benjamin Charlier, Anne Bertrand, Stéphane Epelbaum, et al. Simulating Alzheimer’s disease progression with person-alised digital brain models. Inria preprint, 2018.
  • [19] Ashish Raj, Amy Kuceyeski, and Michael Weiner. A network diffusion model of disease progression in dementia. Neuron, 73(6):1204–1215, 2012.
  • [20] Yasser Iturria-Medina, Roberto C Sotero, Paule J Toussaint, José María Mateos-Pérez, Alan C Evans, Michael W Weiner, Paul Aisen, Ronald Petersen, Clifford R Jack, William Jagust, et al. Early role of vascular dysregulation on late-onset Alzheimer’s disease based on multifactorial data-driven analysis. Nature communications, 7:11934, 2016.
  • [21] Juan Zhou, Efstathios D Gennatas, Joel H Kramer, Bruce L Miller, and William W Seeley. Predicting regional neurodegeneration from the healthy brain functional connectome. Neuron, 73(6):1216–1227, 2012.
  • [22] Lena Maier-Hein, Matthias Eisenmann, Annika Reinke, Sinan Onogur, Marko Stankovic, Patrick Scholz, Tal Arbel, Hrvoje Bogunovic, Andrew P Bradley, Aaron Carass, et al. Why rankings of biomedical image analysis competitions should be interpreted with care. Nature communications, 9(1):5217, 2018.
  • [23] Keelin Murphy, Bram Van Ginneken, Joseph M Reinhardt, Sven Kabus, Kai Ding, Xiang Deng, Kunlin Cao, Kaifang Du, Gary E Christensen, Vincent Garcia, et al. Evaluation of registration methods on thoracic CT: the EMPIRE10 challenge. IEEE Transactions on Medical Imaging, 30(11):1901–1920, 2011.
  • [24] Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (BRATS). IEEE Transactions on Medical Imaging, 34(10):1993–2024, 2014.
  • [25] Klaus H Maier-Hein, Peter F Neher, Jean-Christophe Houde, Marc-Alexandre Côté, Eleftherios Garyfallidis, Jidan Zhong, Maxime Chamberland, Fang-Cheng Yeh, Ying-Chia Lin, Qing Ji, et al. The challenge of mapping the human connectome based on diffusion tractography. Nature communications, 8(1):1349, 2017.
  • [26] Esther E Bron, Marion Smits, Wiesje M Van Der Flier, Hugo Vrenken, Frederik Barkhof, Philip Scheltens, Janne M Papma, Rebecca ME Steketee, Carolina Méndez Orellana, Rozanna Meijboom, et al. Standardized evaluation of algorithms for computer-aided diagnosis of dementia based on structural MRI: the CADDementia challenge. NeuroImage, 111:562–579, 2015.
  • [27] Isabella Castiglioni, Christian Salvatore, Javier Ramírez, and Juan Manuel Górriz. Machine-learning neuroimaging challenge for automated diagnosis of mild cognitive impairment: Lessons learnt. Journal of neuroscience methods, 302:10, 2018.
  • [28] Michael W Weiner, Dallas P Veitch, Paul S Aisen, Laurel A Beckett, Nigel J Cairns, Robert C Green, Danielle Harvey, Clifford R Jack Jr, William Jagust, John C Morris, et al. Recent publications from the Alzheimer’s Disease Neuroimaging Initiative: Reviewing progress toward improved AD clinical trials. Alzheimer’s & Dementia, 13(4):e1–e85, 2017.
  • [29] Genevera I Allen, Nicola Amoroso, Catalina Anghel, Venkat Balagurusamy, Christopher J Bare, Derek Beaton, Roberto Bellotti, David A Bennett, Kevin L Boehme, Paul C Boutros, et al. Crowdsourced estimation of cognitive decline and resilience in Alzheimer’s disease. Alzheimer’s & Dementia, 12(6):645–653, 2016.
  • [30] Razvan V Marinescu, Neil P Oxtoby, Alexandra L Young, Esther E Bron, Arthur W Toga, Michael W Weiner, Frederik Barkhof, Nick C Fox, Stefan Klein, Daniel C Alexander, et al. Tadpole Challenge: prediction of longitudinal evolution in Alzheimer’s disease. arXiv preprint arXiv:1805.03909, 2018.
  • [31] Elaheh Moradi, Antonietta Pepe, Christian Gaser, Heikki Huttunen, Jussi Tohka, Alzheimer’s Disease Neuroimaging Initiative, et al. Machine learning framework for early MRI-based Alzheimer’s conversion prediction in MCI subjects. Neuroimage, 104:398–412, 2015.
  • [32] Xiaojing Long, Lifang Chen, Chunxiang Jiang, Lijuan Zhang, Alzheimer’s Disease Neuroimaging Initiative, et al. Prediction and classification of Alzheimer disease based on quantification of MRI deformation. PloS one, 12(3):e0173372, 2017.
  • [33] Joseph H Grochowalski, Ying Liu, and Karen L Siedlecki. Examining the reliability of ADAS-Cog change scores. Aging, Neuropsychology, and Cognition, 23(5):513–529, 2016.
  • [34] Clifford R Jack Jr, David S Knopman, William J Jagust, Leslie M Shaw, Paul S Aisen, Michael W Weiner, Ronald C Petersen, and John Q Trojanowski. Hypothetical model of dynamic biomarkers of the Alzheimer’s pathological cascade. The Lancet Neurology, 9(1):119–128, 2010.
  • [35] Thomas G Beach, Sarah E Monsell, Leslie E Phillips, and Walter Kukull. Accuracy of the clinical diagnosis of Alzheimer disease at National Institute on Aging Alzheimer Disease Centers, 2005–2010. Journal of neuropathology and experimental neurology, 71(4):266–273, 2012.
  • [36] Xiao Han, Jorge Jovicich, David Salat, Andre van der Kouwe, Brian Quinn, Silvester Czanner, Evelina Busa, Jenni Pacheco, Marilyn Albert, Ronald Killiany, et al. Reliability of MRI-derived measurements of human cerebral cortical thickness: the effects of field strength, scanner upgrade and manufacturer. Neuroimage, 32(1):180–194, 2006.
  • [37] S Kiebel and AP Holmes. The general linear model, volume 8. chapter, 2007.
  • [38] Daniel J Stekhoven and Peter Bühlmann. Missforest—non-parametric missing value imputation for mixed-type data. Bioinformatics, 28(1):112–118, 2011.
  • [39] Rahul Mazumder, Trevor Hastie, and Robert Tibshirani. Spectral regularization algorithms for learning large incomplete matrices. Journal of machine learning research, 11(Aug):2287–2322, 2010.
  • [40] Michael J Pencina, Ralph B D’Agostino Sr, Ralph B D’Agostino Jr, and Ramachandran S Vasan. Evaluating the added predictive ability of a new marker: from area under the ROC curve to reclassification and beyond. Statistics in medicine, 27(2):157–172, 2008.
  • [41] Rebecca R Andridge and Roderick JA Little. A review of hot deck imputation for survey non-response. International statistical review, 78(1):40–64, 2010.
  • [42] Guangyu Chen, Hao Shu, Gang Chen, B Douglas Ward, Piero G Antuono, Zhijun Zhang, and Shi-Jiang Li. Staging Alzheimer’s disease risk by sequencing brain function and structure, cerebrospinal fluid, and cognition biomarkers. Journal of Alzheimer’s Disease, 54(3):983–993, 2016.
  • [43] Chen S, Qi K, and Qiu D. Is MRI all we need? Prediction of conversion between normal cognitive function, mild cognitive impairment and Alzheimer’s disease. ISMRM, 2018.
  • [44] Mostafa Mehdipour Ghazi, Mads Nielsen, Akshay Pai, Marc Modat, M Jorge Cardoso, Sébastien Ourselin, and Lauge Sørensen. Robust parametric modeling of Alzheimer’s disease progression. arXiv preprint arXiv:1908.05338, 2019.
  • [45] Vikram Venkatraghavan, Esther E Bron, Wiro J Niessen, Stefan Klein, and Alzheimer’s Disease Neuroimaging Initiative. Disease progression timeline estimation for Alzheimer’s disease using discriminative event based modeling. NeuroImage, 186:518–532, 2019.
  • [46] Alexandra L Young, Razvan V Marinescu, Neil P Oxtoby, Martina Bocchetta, Keir Yong, Nicholas C Firth, David M Cash, David L Thomas, Katrina M Dick, Jorge Cardoso, et al. Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases with Subtype and Stage Inference. Nature communications, 9(1):4273, 2018.
  • [47] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794. ACM, 2016.
  • [48] JD Kalbfleisch and Jerald Franklin Lawless. The analysis of panel data under a Markov assumption. Journal of the American Statistical Association, 80(392):863–871, 1985.
  • [49] Lars Lau Raket. Disease progression modeling in alzheimer’s disease: insights from the shape of cognitive decline. medRxiv, 2019.
  • [50] Martin Reuter, Nicholas J Schmansky, H Diana Rosas, and Bruce Fischl. Within-subject template estimation for unbiased longitudinal image analysis. Neuroimage, 61(4):1402–1418, 2012.
  • [51] Karl J Friston, Andrew P Holmes, Keith J Worsley, J-P Poline, Chris D Frith, and Richard SJ Frackowiak. Statistical parametric maps in functional imaging: a general linear approach. Human brain mapping, 2(4):189–210, 1994.
  • [52] Gautam Prasad, Talia M Nir, Arthur W Toga, and Paul M Thompson. Tractography density and network measures in Alzheimer’s disease. In 2013 IEEE 10th International Symposium on Biomedical Imaging, pages 692–695. IEEE, 2013.
  • [53] Kenichi Oishi, Andreia Faria, Hangyi Jiang, Xin Li, Kazi Akhter, Jiangyang Zhang, John T Hsu, Michael I Miller, Peter CM van Zijl, Marilyn Albert, et al. Atlas-based whole brain white matter analysis using large deformation diffeomorphic metric mapping: application to normal elderly and Alzheimer’s disease participants. Neuroimage, 46(2):486–499, 2009.
  • [54] Bradley Efron and Robert J Tibshirani. An introduction to the bootstrap. CRC press, 1994.
  • [55] Elizabeth R DeLong, David M DeLong, and Daniel L Clarke-Pearson. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics, pages 837–845, 1988.

Supplementary material

Appendix A Creating the D1-D4 datasets

The data used from ADNI consists of: (1) CSF markers of amyloid-beta and tau deposition; (2) various imaging modalities such as magnetic resonance imaging (MRI), positron emission tomography (PET) using several tracers: Fluorodeoxyglucose (FDG, hypometabolism), AV45 (amyloid), AV1451 (tau) as well as diffusion tensor imaging (DTI); (3) cognitive assessments acquired in the presence of a clinical expert; (4) genetic information such as alipoprotein E4 (APOE4) status extracted from DNA samples; and (5) general demographic information. Extracted features from this data were merged together into a final spreadsheet and made available on the LONI ADNI website.

The imaging data has been pre-processed with standard ADNI pipelines. For MRI scans, this included correction for gradient non-linearity, B1 non-uniformity correction and peak sharpening. [ADNI MRI pre-processing]. Meaningful regional features such as volume and cortical thickness were extracted using the Freesurfer cross-sectional and longitudinal pipelines [50]. Each PET image (FDG, AV45, AV1451) had their frames co-registered, averaged across the six five-minute frames, standardised with respect to the orientation and voxel size, and smoothed to produce a uniform resolution of 8mm full-width/half-max (FWHM) (see http://adni.loni.usc.edu/methods/pet-analysis/pre-processing/). Standardised uptake value ratio (SUVR) measures for relevant regions-of-interest were extracted after registering the PET images to corresponding MR images using the SPM5 software [51]. Further details have been provided in the ADNI procedures manual. DTI scans were corrected for head motion and eddy-current distortion, skull-stripped, EPI-corrected, and finally aligned to the T1 scans using the pipeline from [52]. Diffusion tensor summary measures were estimated based on the Eve white-matter atlas [53].

In addition to the standard datasets, we also created three leaderboard datasets LB1, LB2 and LB2 which mimick the D1, D2 and D4 datasets. These datasets were used by participants to preliminarily evaluate their algorithms before the competition deadline, and to compare their results on the leaderboard system (https://tadpole.grand-challenge.org/Leaderboard/).

Appendix B Statistical testing

B.1 Differences in MAUC scores

For analysing whether the MAUC scores obtained by top algorithms are significantly different, we performed a bootstrapped hypothesis test [54], since the significance test for comparing two AUC scores [55] does not extend to multiple classes. For two TADPOLE entries A𝐴A and B𝐵B, where A𝐴A scored better than B𝐵B on the full D4 test set, we want to confirm either the null hypothesis H0:mauc(A)=mauc(B):subscript𝐻0𝑚𝑎𝑢𝑐𝐴𝑚𝑎𝑢𝑐𝐵H_{0}:mauc(A)=mauc(B) or the alternative hypothesis H1:mauc(A)>mauc(B):subscript𝐻1𝑚𝑎𝑢𝑐𝐴𝑚𝑎𝑢𝑐𝐵H_{1}:mauc(A)>mauc(B). We then proceed as follows:

  • Sample N=50𝑁50N=50 random bootstraps Disubscript𝐷𝑖D_{i} of the D4 test set with replacement.

  • Compute the mauc(A,Di)𝑚𝑎𝑢𝑐𝐴subscript𝐷𝑖mauc(A,D_{i}) and mauc(B,Di)𝑚𝑎𝑢𝑐𝐵subscript𝐷𝑖mauc(B,D_{i}) based on the bootstrapped dataset. Repeat for all N𝑁N bootstraps.

  • Compute the p-value as iI[mauc(A,Di)<mauc(B,Di)]/Nsubscript𝑖𝐼delimited-[]𝑚𝑎𝑢𝑐𝐴subscript𝐷𝑖𝑚𝑎𝑢𝑐𝐵subscript𝐷𝑖𝑁\sum_{i}I[mauc(A,D_{i})<mauc(B,D_{i})]/N, which is the proportion of bootstrapped datasets where A𝐴A performed worse than B𝐵B.

  • Accept/reject H0subscript𝐻0H_{0} based on a 5% significance level.

B.2 Differences in MAE scores

For comparing differences in MAE scores, we applied the non-parametric Wilcoxon signed-rank test on paired samples of absolute errors across all visits of the D4 subjects. We chose the non-parametric Wilcoxon test because the input samples are not normally distributed, as they represent absolute errors and are always positive. We report results based on a 5% significance level.

B.3 Differences between D2 and D3 forecasts

For comparing differences between the scores obtained by two algorithms on D2 vs D3 forecasts, we use an approach similar to comparing MAUC scores (section 8.4.2).

Appendix C Supplementary Results

Refer to caption
Figure 2: Distribution of performance metrics for clinical diagnosis (MAUC and BCA), ADAS-Cog13 (MAE, WES and CPA) and ventricle volume (MAE, WES and CPA) on the longitudinal D2 prediction set. For each entry, we plot the distribution of performance metrics derived using 50 bootstrap data sets drawn from the D4 test set. The submissions (rows) are in the same order as in Table 2. Entries are missing where teams did not make predictions for a particular target variable.
Refer to caption
Figure 3: Box plots of performance metrics for clinical diagnosis (MAUC and BCA), ADAS-Cog13 (MAE, WES and CPA) and ventricle volume (MAE, WES and CPA) on the cross-sectional D3 prediction set. The submissions (rows) are in the same order as in Table 3. Some entries are missing because teams did not make predictions for those target variables.
OverallDiagnosisADAS-Cog13Ventricles (% ICV)
SubmissionRankRankMAUCBCARankMAEWESCPARankMAEWESCPA
Billabong-UniAV45   110.7190.6241-28.718.550.333-43.493.400.50
Billabong-Uni   220.7170.6211-28.718.550.333-43.493.400.50
Billabong-MultiAV45   330.6610.5623-412.9512.710.421-23.163.080.47
Billabong-Multi   440.6580.5523-412.9512.710.421-23.163.080.47
Simple-SPMC-Plymouth2   -50.5000.504--------
Simple-SPMC-Plymouth1   -60.5000.499--------
Table 6: Results on custom prediction sets from two teams: Billabong and SPMC-Plymouth. SPMC-Plymouth predicted fewer subjects due to an incomplete submission, while Billabong used a prediction set similar to D3, but filled in missing data for cognitive tests and MRI with the last available measurement. SPMC-Plymouth only submitted predictions for clinical diagnosis, and obtained an MAUC score of 0.5. Results from Billabong show higher MAUC and BCA in diagnosis prediction compared to D3, but lower performance for ADAS-Cog13 and Ventricle volume prediction. Bold entries show best scores in this category.
Refer to caption
Figure 4: For D2 submissions, we show scatter plots of pairs of performance metrics for (top row) clinical diagnosis, (middle row) ADAS-Cog13 and (bottom row) Ventricles. Each dot is a participant submission, coloured according to the type of prediction algorithm used. Correlation coefficients and p-values are given above each subplot. A few outlier submissions with ADAS MAE >> 20, ADAS WES >> 40 or Ventricle WES >> 3 were excluded from the analysis.
Refer to caption
Figure 5: For D3 submissions, we show scatter plots of pairs of performance metrics for (top row) clinical diagnosis, (middle row) ADAS-Cog13 and (bottom row) Ventricles. Each dot is a participant submission, coloured according to the type of prediction algorithm used. Correlation coefficients and p-values are given above each subplot. A few outlier submissions with ADAS MAE >> 20, ADAS WES >> 40 or Ventricle WES >> 3 were excluded from the analysis.
Refer to caption
Figure 6: Bias in prediction of clinical diagnosis for MCI subjects only. X-axis shows individual subjects with designated MCI status at the clinical visit in D4, while the Y-axis shows TADPOLE algorithms. Red represents subjects which were predicted as CN with true diagnosis of MCI, while blue represents subjects predicted as AD with true diagnosis of MCI. Some algorithms show systematic biases either towards CN or AD.
Refer to caption
Figure 7: Bias in prediction of ADAS-Cog13. X-axis shows individual subjects with ADAS-Cog measurements in D4, while Y-axis shows TADPOLE algorithms. Red represents under-estimates while blue represents over-estimates. Most algorithms under-estimate ADAS-Cog measurements.
Refer to caption
Figure 8: Bias in prediction of ventricle volume. X-axis shows individual subjects with Ventricle volume measurements in D4, while Y-axis shows TADPOLE algorithms. Red represents under-estimates while blue represents over-estimates. Some algorithms systematically under-estimate or over-estimate ventricle volume.
Refer to caption
Figure 9: Distribution of ranks in clinical diagnosis MAUC for TADPOLE submissions using the longitudinal prediction set (D2), obtained from N=50𝑁50N=50 bootstraps of the test set (D4). More precisely, we computed the MAUC ranks given a specific bootstrap of the test set, and then for each TADPOLE submission (Y-axis) we plotted the number of times it achieved a specific rank. Figures 1014 use the same methodology.
Refer to caption
Figure 10: Distribution of ranks in ADAS-Cog13 MAE for TADPOLE submissions using the longitudinal prediction set (D2)
Refer to caption
Figure 11: Distribution of ranks in Ventricle Volume MAE for TADPOLE submissions using the longitudinal prediction set (D2).
Refer to caption
Figure 12: Distribution of ranks in clinical diagnosis MAUC for TADPOLE submissions using the longitudinal prediction set (D3).
Refer to caption
Figure 13: Distribution of ranks in ADAS-Cog13 MAE for TADPOLE submissions using the longitudinal prediction set (D3).
Refer to caption
Figure 14: Distribution of ranks in Ventricle Volume MAE for TADPOLE submissions using the longitudinal prediction set (D3).