Abstract: Clinical natural language processing (NLP) models for hospital discharge prediction are vulnerable to explicit discharge-related lexical cues in clinical notes, even when note timestamps are temporally filtered. These cues may act as proxies for future outcomes, leading models to rely on shortcut-like signals rather than clinically meaningful evidence. We propose AuditBERT-Discharge, a lightweight framework that reframes interpretability as a training-time auditing mechanism. Token-level SHAP attributions from an initial BioClinicalBERT model are used to identify highly attributed discharge-related tokens, which are subsequently masked in the training corpus before model retraining. We evaluate the framework on a retrospective cohort of elective spine surgery encounters from a single academic medical center. Qualitative attribution analysis and preliminary evaluation suggest that SHAP-guided auditing reduces reliance on explicit discharge-proxy language and yields more conservative, better-calibrated predictions, although improvements in discrimination are not uniform. These findings demonstrate the potential of interpretability-guided auditing as a practical strategy for identifying and constraining shortcut-like lexical dependencies in temporally sensitive clinical NLP tasks.
Read the original article:
Interpretability-Guided Auditing for Clinical Language Models for Discharge Prediction
