RaiLED-AD: Rationale-Guided Knowledge Transfer for Alzheimer's Disease Prediction from Electronic Health Records

In Alzheimer's Disease and Related Dementia (ADRD) prediction, Electronic Health Records (EHRs) provide rich but fragmented information. Without a coherent clinical narrative, models tend to rely on a few dominant signals (e.g., age-related patterns) rather than capturing the underlying clinical mechanisms. Such reliance becomes problematic in younger-onset cases, where these signals are less informative. To address this challenge, we propose RaiLED-AD, a dual-encoder teacher-student framework where the student learns from serialized EHR data and the teacher leverages LLM-generated narratives that capture temporal and relational patterns. A hybrid objective with soft-label supervision and hierarchical contrastive alignment transfers these reasoning signals to the student, which operates independently at inference. On a real-world EHR cohort, RaiLED-AD consistently improves ADRD prediction over baselines and achieves substantial gains in the challenging younger-onset subgroup (index age <65). These results highlight the potential of integrating LLM-derived reasoning signals with structured EHR models for early-stage ADRD risk prediction.