AToP: An Explainability Framework for Transformer-Based EHR Predictions Using Attribution-Guided Temporal Pattern Mining

AToP converts Transformer per-token Integrated Gradients on MIMIC-IV v3.1 into temporal pattern narratives, revealing AUROC 0.77 and PR-AUC 0.34 readmission signals.

• Why it matters: Token-level explanations miss within-visit co-occurrences and cross-visit trajectories, while rare but highly attributed clinical markers are hard to summarize for EHR risk predictions.
• What they did: AToP mines frequent episodes from salient-token subsequences, aggregates attribution across “carriers,” collapses near-duplicates by patient overlap, and validates with masking, reversed-order analysis, token faithfulness tests, and KernelSHAP comparison.
• The result: Carrier-conditional attribution highlights rare psychiatric markers (including psychosis, self-harm overdose, SI, and severe MDD), while pattern-level attribution identifies ordered trajectories like {MDD, SI} [->] SI and cancer anemia [->] ascites, plus lower-risk obstetric and chest pain patterns.

AI summary of the abstract · may contain errors · not medical advice · read the paper.