L2C-FNN: Longitudinal to Cross-sectional Feedforward Neural Network for generalizable AD-dementia progression prediction

Motivation: Current longitudinal AD-dementia progression prediction studies lack cross-cohort evaluation, raising concerns about the clinical applicability of prediction models. Goal(s): Our goal was to develop a generalizable ML algorithm, L2C-FNN, and assess its generalizability across entirely distinct test cohorts. Approach: L2C-FNN and baseline models were trained solely on ADNI and subsequently evaluated on AIBL, MACC, and OASIS. Multimodal biomarkers were leveraged for forecasting future clinical diagnosis, cognition, and ventricle volume. Results: Our algorithm compares favorably against strong baseline models across all test datasets, confirming its superior generalizability. Impact: The demonstrated potential for improved generalizability in L2C-FNN signifies progress toward enhancing AI prediction models for clinical application. This underscores the continued need for cross-cohort evaluation in future AD-dementia progression modeling studies.