Physiological Stability Index (PSI): A Retrospective Computational Framework for Assessing Dynamic Physiological Instability

Background: Acute clinical deterioration in intensive care may be preceded by gradual changes in the dynamics of physiological signals. Current monitoring systems, however, are often based on static thresholds for individual vital signs and may not fully capture loss of systemic resilience before overt decompensation. Objective: This study presents an open retrospective evaluation of the Physiological Stability Index (PSI) as an interpretable computational framework for assessing dynamic physiological instability from clinical and physiological time series. PSI is not proposed as a diagnostic device or a substitute for clinical judgment, but as a testable model of systemic fragility. Methods: The framework combines components representing the intensity of physiological perturbations, autocorrelation persistence, relaxation dynamics, and changes in coupling relationships between subsystems. The open retrospective computational evaluation used publicly available ICU data from the PhysioNet/CinC Challenge 2009, eICU Demo 2.0.1, Challenge 2012, and Challenge 2019 datasets. In the main analyses, 51,875 records met computational eligibility criteria; the complete Challenge 2019 recomputation included 38,774 analyzable records and 459,406 PSI windows. Results: Global PSI showed measurable but limited discriminative performance against publicly available endpoints: eICU mortality AUROC 0.666, Challenge 2012 mortality AUROC 0.567, and Challenge 2019 presence of at least one positive SepsisLabel AUROC 0.633. The local audit reanalysis of Challenge 2012 Set A included 3,908 episodes and 163,667 valid windows; the window-level proxy endpoint yielded AUROC 0.593 (95% CI 0.558–0.599) and AUPRC 0.443 (95% CI 0.398–0.450). Important negative findings were that the isolated DC component outperformed global PSI in the local proxy task, while a simple component mean in Challenge 2019 slightly outperformed PSI-v2. Adding PSI did not improve episode-level SAPS-I, SOFA, or SAPS-I + SOFA mortality models, and alternative mortality aggregations of PSI were weak. Interpretation: The results support the computational feasibility of PSI and the presence of a non-random but limited dynamic signal in open ICU data. They do not constitute clinical validation, evidence of causality, a calibrated alarm system, or reliable pre-event prediction in individual patients. Further development requires a locked protocol, patient-level external testing, raw-vital baseline models, prospective shadow-mode evaluation, calibration, and assessment of false alarms per patient-day.