Search papers, labs, and topics across Lattice.
This paper introduces a drift-adaptive two-stream architecture for ICU intervention prediction that separates physiological and treatment representations, allowing for targeted updates to treatment parameters while preserving stable physiological data. By utilizing a Temporal RAG module, the system grounds predictions in patient-specific, era-matched PubMed evidence, enhancing the model's interpretability and relevance. Experiments demonstrated that selective adaptation improved discrimination and calibration for vasopressor and septic shock cases, outperforming a fully retrained baseline in identifying critical cases without compromising retrieval consistency.
Selective adaptation in clinical AI can enhance prediction accuracy for critical interventions while preserving the integrity of patient physiology.
Background: Clinical decision support systems degrade silently as treatment protocols evolve, yet standard adaptation methods treat models as monolithic blocks, unable to distinguish stable patient physiology from shifting institutional practice. Methods: We propose an adaptive clinical intelligence architecture for ICU intervention prediction that structurally decouples physiological from treatment representations, confining parameter updates to the treatment stream upon a dual distributional and accuracy trigger. Automated audit logs record which treatment features drove each adaptation event and how their importance shifted. At inference, an attribution-driven Temporal RAG module grounds each prediction in patient-specific, era-matched PubMed evidence anchored to the patient's dominant physiological features. Experiments used 84,792 MIMIC-IV stays (2008-2022) under strict chronological split. Results: Drift localised entirely to the treatment stream, validating the structural prior. Selective adaptation improved vasopressor and septic shock discrimination and calibration over the static source model. A fully retrained baseline yielded marginally higher aggregate discrimination but missed 26 septic shock cases the framework correctly identified, with none in the reverse direction; retrieval consistency with the pre-adaptation source model was preserved by the framework but degraded substantially in the retrained baseline. Conclusions: Structurally constraining adaptation to drifting components while preserving stable physiological representations enables clinical AI to evolve with practice without distorting learned patient biology. This architecture offers a template for governable, interpretable deployment of adaptive models in high-stakes clinical environments.
