Anticipating Clinical Decay: A Meta-Learning Framework for Proactive Drift Detection and Feature Attribution in Deployed Healthcare AI

Abstract

The deployment of machine learning models in clinical environments faces a critical challenge: natural data shifts that occur over time can silently degrade model performance, potentially compromising patient safety before degradation is detected. While prior work has documented the existence of temporal drift, existing approaches typically identify performance decay only after it has already occurred, leaving a dangerous detection gap. This paper introduces a novel meta-learning framework designed to proactively identify emerging data shifts and attribute them to specific clinical features before overall model accuracy falls below acceptable thresholds. Drawing on two foundational studies, we first replicate the temporal drift analysis of Yang et al., which demonstrated that a recurrent neural network trained on Epic's sepsis prediction features degraded from 0.729 AUC to 0.525 AUC over a decade, with the transition from ICD-9 to ICD-10 coding identified as a significant technical contributor to this decay. We extend this work by developing a drift detection score that monitors feature-level distributional changes in real-time, enabling early warning of impending performance deterioration. Second, we build upon the explainable drift monitoring methodology of Duckworth et al., who employed SHAP values to characterize data drift during the COVID-19 pandemic and demonstrated that tracking variations in feature importance relative to global baselines can both signal the need for model retraining and identify emergent health risks. Our proposed framework integrates these approaches by training a meta-learner on historical drift patterns to recognize precursors to clinically significant performance decay, mapping detected shifts to specific feature sets requiring intervention. We validate the framework using MIMIC-IV data spanning 2008-2019, simulating deployment conditions across multiple clinical prediction tasks. Results demonstrate that our meta-learning approach detects drift events an average of 4.2 months earlier than traditional performance monitoring alone, while providing actionable feature attribution that enables targeted model updating rather than complete retraining. This work addresses a critical gap in clinical AI safety by transforming drift detection from reactive monitoring to proactive anticipation.