Predicting incident risk from longitudinal electronic health records (EHRs) is difficult because relevant signals are multimodal, weak when isolated, and scattered across irregular patient histories. We introduce structured evidence routing, a router‑predictor‑reviewer workflow that separates full‑record access from disease‑specific assessment.
The router first organizes the complete pre‑index EHR into a compact summary and extracts targeted evidence slices. The predictor then builds an evidence‑linked risk assessment using these slices, and the reviewer critically evaluates the assessment.
To compare with supervised EHRSHOT baselines, we pair the routed evidence summaries with a supervised classifier readout. Across five 1‑year incident diagnosis tasks, our approach attains AUROC comparable to established supervised EHRSHOT baselines and remains competitive on AUPRC, while exposing a patient‑specific evidence trail. Internal pre‑readout ablations further indicate that routing, laboratory evidence, task guidance, and review each contribute positively to performance.
Blogger's Review: This work cleverly decouples "full‑record" from "disease‑focused" analysis through explicit evidence routing, preserving information richness while enhancing interpretability, and offers a traceable decision path for clinical risk prediction.