Biological Amnesia in ICU Time-Series Prediction: A Drift-Adaptive Two-Stream Architecture with Temporal Retrieval

Research Question and Approach

Clinical decision support systems (CDSS) in intensive care units (ICU) often suffer from performance degradation as institutional treatment protocols evolve—a phenomenon known as concept drift. Standard approaches, such as full model retraining, often lead to 'biological amnesia,' where the model unintentionally overwrites stable physiological knowledge while trying to learn new treatment patterns. The authors propose a two-stream architecture that structurally separates physiological data (processed by an LSTM) from treatment context (processed by an MLP). This design allows for selective adaptation: when a dual-signal trigger (based on distributional shift and accuracy) detects drift, only the treatment and fusion parameters are updated, while the physiological stream remains frozen.

Key Findings

Experiments using 84,792 ICU stays from the MIMIC-IV database (2008–2022) demonstrated that drift was almost entirely localized to the treatment stream, validating the architectural prior. The selective adaptation framework significantly outperformed static models and monolithic retrained baselines in predicting vasopressor needs and septic shock. Notably, while a fully retrained baseline achieved similar aggregate discrimination, it missed 26 septic shock cases that the proposed framework correctly identified, with zero cases missed in the reverse direction. Furthermore, the framework maintained retrieval consistency with the original source model, whereas full retraining caused significant degradation in evidence retrieval.

This research provides a template for building 'governable' clinical AI by generating transparent audit logs that explain exactly how a model's reasoning shifts during an update. By constraining adaptation to drifting components, the system can evolve alongside changing medical guidelines without losing its foundational understanding of human physiology. The inclusion of an attribution-driven Temporal RAG module further ensures that clinical predictions are grounded in era-matched, patient-specific medical literature, enhancing the interpretability and safety of AI deployment in high-stakes environments.