AToP: An Explainability Framework for Transformer-Based EHR Predictions Using Attribution-Guided Temporal Pattern Mining
Year: 2026
Type: preprint
Abstract: Objective: To develop and evaluate AToP (Attribution to Temporal Patterns), a post-hoc explainability framework that converts per-token Integrated Gradients (IG) from a Transformer into structured temporal pattern narratives. Materials and Methods: AToP was applied to a single-stream Transformer trained on MIMIC-IV v3.1 (218,196 patients; 12.6% unplanned 30-day readmission). It mines frequent episodes from salient-token subsequences in the training cohort, aggregates attribution among patients whose records contain each token or pattern, and collapses near-duplicates by patient overlap. Validation used masking, reversed-order analysis, token-level faithfulness tests, and KernelSHAP comparison. Results: The Transformer achieved AUROC 0.77 and PR-AUC 0.34. Population-weighted attribution emphasized high-prevalence codes with small per-patient impact. Carrier-conditional attribution, the mean among patients with the token, surfaced rare high-attribution psychiatric markers including psychosis, self-harm overdose, suicidal ideation (SI), and severe major depressive disorder (MDD). Pattern-level attribution identified within-visit co-occurrences and cross-visit trajectories not captured by token rankings. Risk-increasing patterns included psychiatric sequences ({MDD, SI} → SI) and oncologic progressions (cancer anemia → ascites). Obstetric and chest pain evaluation patterns were associated with lower predicted readmission. Discussion: The framework provides two complementary views of model predictions: carrier-conditional aggregation highlights rare, highly attributed clinical features, while temporal pattern analysis identifies within-visit combinations and cross-visit trajectories missed by token-level rankings. Masking and visit-order reversal provided evidence of correspondence with model behavior, but pattern attributions do not represent unique effects or establish clinical validity. Conclusion: AToP turns a Transformer's attributions into visit-anchored temporal pattern narratives, showing which events, in which order, contribute to its readmission predictions.
Source: medRxiv
Authors: Anas Belouali, Hadi Kharrazi
Institution: Johns Hopkins University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225859660
