Applying Integrating Event-Based Models and Deep Learning to Predict Baseline Diagnosis and Tau Progression in Alzheimer's Disease

Motivation: Our previous Deep Learning Network succeeded in predicting baseline cognition and diagnosis for AD subjects using MRI. We propose to combine it with the difficult task of predicting individualized tau spread task and on-set disease stages. Goal(s): Our goal was to determine if a custom UNet could predict the Event-Based Model (SuStaIn - EBM) results accurately. Approach: We designed a multitask UNet with MRI and Tau-PET inputs, customized loss functions, trained and tested it on ADNI 3 MCI and AD subjects. Results: The algorithm was able to predict the EBM seeding patterns and stages with R-squared scores over 0.7. Impact: Our method enhances understanding of inter-subject heterogeneity in AD, bridges the information misalignment in Tau-PET and MRI, and supports precision treatment by predicting individual tau progression and seeding patterns with only baseline MRI and demographic inputs.