HAI Book 2025 - Flipbook - Page 317
Kang, Min Su
64
Graph convolutional networks to predict cognitive decline at the
individual level in Alzheimer9s disease
Min Su Kang1,2, Jacquline Heaton1,2, Alexander J. Nyman1,2, Mahdi Biparva1, Melissa
McSweeney1,2, Julie Ottoy1,2, Mario Masellis2,3, Zahra Shirzadi4,5, Keith Johnson4,5, Reisa
Sperling4,5, Aaron Schultz4,5, Chhatwal Jasmeer4,5, Walter Swardfager2,6, Nir Lipsman2,7,
Sandra E. Black2,3, Jennifer Rabin2,3,6, Maged Goubran1,2,6,8
1
Artificial Intelligence and Computational Neurosciences lab, Sunnybrook Research Institute, University of Toronto,
Toronto, ON, CA
2
Hurvitz Brain Sciences Program, Sunnybrook Research Institute, University of Toronto, Toronto, ON, CA
3
Department of Medicine (Division of Neurology), University of Toronto, Toronto, ON, CA
4
Department of Neurology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, US
5
Center for Alzheimer Research and Treatment, Department of Neurology, Brigham and Women’s Hospital, Harvard
Medical School, Boston, MA, US
6
Department of Pharmacology & Toxicology, University of Toronto, Toronto, ON, CA
7
Harquail Centre for Neuromodulation, Sunnybrook Health Sciences Centre, University of Toronto, Toronto, ON, CA
8
Department of Medical Biophysics, University of Toronto, Toronto, ON, CA
Background: Cognitive trajectories in Alzheimer9s disease (AD) are highly heterogeneous. There is a pressing need
to develop biomarker models that predict cognitive trajectories at the individual level. Our goal was to develop
advanced deep-learning models to predict cognitive trajectories at the individual level as measured by the
Preclinical Alzheimer Cognitive Composite (PACC). We employed graph convolutional networks (GCNs) using
multimodal longitudinal biomarkers.
Methods: 597 ADNI participants (245 CN, 252 MCI, 100 AD dementia) and 200 HABS participants (196 CN, 4 MCI)
underwent baseline 3T-MRI, A´-PET and cognitive assessments. Baseline predictors included demographics (age,
sex, education), genetic (APOEe4), cognitive (MMSE, Trails A, Boston naming test and PACC), and imaging
variables (A´-PET, cortical volume and thickness). The outcome was the longitudinal PACC scores. We trained
GCN and GraphSage (a GCN that aggregates features across neighbouring nodes) to predict longitudinal PACC
trajectories using 80% of the ADNI data and validated the models in the remaining 20% of ADNI data and the
entire HABS dataset. We compared GCNs against a traditional gradient-boosting regressor (GBR).
Results: GraphSage performed the best overall in predicting longitudinal PACC scores with the highest R2
(GraphSage: 0.89±0.33 vs. GRB: 0.65±0.63) and lowest mean square error (MSE) (Fig.1). GraphSage significantly
outperformed GBR on both ADNI and HABS (p
