Evaluating Patient Representation Strategies for Early Prediction of Alzheimer’s Disease and Related Dementias from Longitudinal EHRs

Abstract

The neurodegenerative processes underlying Alzheimer’s disease and related dementias (ADRD) begin years before clinical diagnosis, motivating scalable approaches for earlier identification. Electronic health records (EHRs) provide longitudinal clinical information for early ADRD prediction but are sparse, temporal, and hierarchical, and alternative EHR representations encode these characteristics differently. We systematically evaluated how EHR representation and temporal distance from diagnosis influence ADRD prediction and whether representation-specific performance generalizes across health systems. Using All of Us EHRs, we compared interpretable count-based representations with pretrained clinical foundation-model representations at clinically meaningful lead times across multiple ADRD case definitions. Count-based models achieved the strongest overall discrimination and calibration. Performance generally declined with increasing prediction horizons, while differences between representations narrowed; at 36 months, pretrained representations achieved comparable AUROC (0.719 versus 0.738) with higher sensitivity and F1 at a fixed operating threshold. Temporal feature attribution showed greater model reliance on healthcare-utilization indicators at longer horizons and increasing importance of cognitive phenotypes, including amnesia and mild cognitive impairment, nearer diagnosis. Zero-shot evaluation on independent UChicago Medicine EHRs showed substantial performance degradation across representations, demonstrating challenges in cross-institutional transportability. These findings characterize representation-dependent tradeoffs in early ADRD prediction across temporal horizons and healthcare systems.