SMARtCARE Uses Bounded-Autonomy AI to Surface Hidden ICU Risk Patterns

Summary

SMARtCARE is a privacy-aware, agentic clinical decision-support architecture designed to address a long-context problem: relevant patient history from earlier admissions may fall outside the system's active reasoning context. In ICU monitoring, an early drift in vital signs can therefore look nonspecific even when it resembles a previous deterioration pattern. The system has four states: Stable, Meta-cognitive, Assisted, and Regulated (Revoked). It does not automatically retrieve prior records; instead, it compares current drift with a lossy six-channel fingerprint of the patient's prior trajectory. A match when the prior record is absent triggers a Meta-cognitive escalation for clinician review, and complete retrieval requires clinician action in the Assisted state. A patient-identity guard is intended to maintain correct attribution across data loading, logs, and audits. Evaluation included a synthetic Monte Carlo study of state transitions and estimator stability, which was not a clinical-performance test, plus runs on MIMIC-III and MIMIC-IV Clinical Database Demos. The pipeline found one prior-pattern recurrence among 14 two-admission MIMIC-III patients and no matches among nine such MIMIC-IV patients, highlighting the limitation of a fixed canonical pattern library. All logged decisions in both runs were traceable and correctly attributed. The authors present SMARtCARE as a mechanism for surfacing middle-context risk, not as evidence of clinical effectiveness.