Predictive Analytics
Competition 2019

This year, we invite you to develop a model predicting brain age from healthy individuals based on structural Magnetic Resonance Imaging (sMRI) data.


Ready, Steady, Go!

It was an excellent journey. Thank you guys

The PAC 2019 is over and ended in amazing results, valuable insights and great enthusiastic discussion. Thanks to all the teams involved. See the results

We are now very happy to announce that we are offering all teams the opportunity to contribute to the PAC special issue in Frontiers in Psychiatry (Impact Factor 3.161). All the teams will receive further information per mail.

Frontiers in Psychiatry

What To Do

Background

The brain changes as we age, and these changes are associated with cognitive decline, neurodegenerative disease and dementia. Although brain ageing is universal, rates of brain ageing differ markedly; some people suffer cognitive decline in later middle-adulthood, while others remains cognitively normal into their tenth decade. The process of brain ageing includes morphological and functional changes to the brain, which can be assessed using neuroimaging. This raises the possibility that the variability in brain ageing can be measured, and research has focused on developing such a neuroimaging biomarker of brain ageing; the so-called ‘brain-age’ paradigm.

The idea with brain-age is that if statistical models can be developed to accurately predict chronological age in healthy people (using neuroimaging data), then the apparent age of a new individual’s brain can be calculated. Where someone’s brain-age is older than their real age, this is thought to reflect poorer brain health, relative to their age. Older-appearing brains have been associated with psychiatric and neurological diseases, with greater risk of developing dementia and a shorter lifespan. Younger-appearing brains have been found in people who exercise more, have greater years of education, meditate or play musical instruments.

The hope is that brain-age can provide a sensitive, if unspecific, global measure of brain health, that could be used in many contexts. These include clinical trials of neuroprotective therapies, screening groups of people at-risk of poorer cognitive ageing, and providing mechanistic insights into the downstream consequences of different diseases.

Critical to the success of brain-age models, is the accuracy of the healthy training model. Hence, the goal of this year’s PAC is to build the most accurate model, using the training data supplied. Specifically, we would like to minimize brain predicted age difference (brain-PAD) which is calculated as brain-predicted age minus chronological age.

Data

We provide you with 1) raw nifti data and 2) analogous to Cole et al. with fully pre-processed data as described in the paper. Of course, you are free to use either one of the datasets and/or process them further in any way you like.

We will evaluate model performance by comparing the uploaded predictions to actual chronological age in the test dataset for each individual. This year, there are two distinct objectives:

  1. The team submitting the model with the smallest Mean Absolute Error for the test dataset will win this year’s PAC Best Model Award.
  2. The team submitting the model with the smallest Mean Absolute Error for the test dataset while keeping the Spearman correlation between brain predicted age difference and chronological age below r = .10 will win this year’s PAC Bias Reduction Award.

PAC 2019 Results

Results Objective 1

Lowest Mean Absolute Error

Team MAE ρ
1 BrainAgeDifference
Donders Institute, Radboud University
2.9043 -0.3914
2 BrainAGE
University Hospital Jena
3.0857 -0.3423
3 ARAMIS
Brain and Spine Institute Paris
3.3284 -0.2103
4 Quantum Pika
National Yang Ming University
3.3315 -0.3939
5 sablab
Cornell University
3.3716 -0.2469
6 DRAGN
University of Pennsylvania
3.5464 -0.3189
7 Procrastination
Beijing Tiantan Hospital
3.6026 -0.3828
8 Milan_buaa
Beihang University
3.6358 -0.3294
9 inteneural
Inteneural
3.6787 -0.3861
10 Mind the Gap
King's College London
3.7597 -0.4271

Results Objective 2

Lowest Mean Absolute Error while keeping the absolute spearman correlation of the brain age gap and the true age < 0.1

Team MAE ρ
1 BrainAgeDifference
Donders Institute, Radboud University
2.9503 -0.0311
2 BrainAGE
University Hospital Jena
3.7989 -0.0867
3 DRAGN
University of Pennsylvania
3.9242 0.0205
4 Quantum Pika
National Yang Ming University
3.9439 -0.0147
5 PACMEN
Stellenbosch University Cape Town
4.7326 -0.0147
6 ARAMIS
Brain and Spine Institute Paris
4.8320 -0.0211

PAC 2019 Teams

We are 274 participants in 79 teams

PAC ADMINS

Tim Hahn

Tim Hahn

Head of Machine Learning in Psychiatry Group
University of Muenster

Udo Dannlowski

Udo Dannlowski

Head of Institute of Translational Psychiatry
University of Muenster

James Cole

James Cole

Centre for Neuroimaging Science
King's College London


Christian Gaser

Christian Gaser

Structural Brain Mapping
University of Jena

Nils Winter

Nils Winter

PHOTON Team
University of Muenster

Ramona Leenings

Ramona Leenings

PHOTON Team
University of Muenster

Dominik Grotegerd

Dominik Grotegerd

PHOTON Team
University of Muenster

Editors Psychiatry

Editorial Team

Frontiers in Psychiatry
Frontiers Media S.A.