Answers are grounded in this paper and cite their sources.
Artificial intelligence-based clinical decision support systems can materially improve risk prediction and resource allocation in critical care, yet their sustained governance raises an underexamined problem for the health informatics community: model fairness and interpretability are typically certified once, at deployment, even though both properties evolve as models are periodically retrained within feedback-laden sociotechnical environments. This study reframes algorithmic fairness as a dynamic, lifecycle property rather than a static deployment checkpoint, and develops a multi-model framework that jointly addresses predictive performance, longitudinal fairness, and explainability. Using the MIMIC-IV cohort, the LSTM achieves an AUC of 0.884 (95% CI: 0.859-0.908) on a held-out test set, comparable to contemporary tabular baselines (LightGBM 0.892, Random Forest 0.889, XGBoost 0.886, Logistic Regression 0.867) and substantially above a SOFA-only clinical baseline (0.770). Because three development-model predictors (vasopressors, mechanical ventilation, and the full SOFA score) are not consistently recorded in eICU, cross-institutional generalization is assessed with a common-feature Random Forest, yielding AUC 0.743 (95% CI: 0.738-0.749). We conduct a longitudinal simulation on the real MIMIC-IV cohort, with simulated performative feedback, across 15 retraining cycles and 16 intersectional subgroups, simultaneously monitoring TPR, FPR, and FOR gaps. Across 30 independent runs, all three strategies converge to final TPR gaps with no significant difference (Friedman χ ²(2)=1.195, $p=0.550$); however, the Dynamic Adaptive Fairness controller exhibits bounded, damped trajectories where static reweighing shows larger transient spikes. Under an auxiliary exogenous covariate shift, adaptive control attenuates the post-shift TPR-gap spike (Δ =+0.091) relative to the no-intervention baseline (Δ =+0.139). A cross-institutional eICU fairness audit reveals measurable per-hospital variability (TPR gap 0.134, FPR gap 0.060, FOR gap 0.032); subgroup estimates are reported with event counts and Wilson intervals. SHAP feature-importance rankings remain highly stable across cycles (ρ =0.986), anchored by clinical severity markers. The framework supports pre-deployment fairness stress testing and longitudinal governance of clinical AI.
No takes yet. Share an insight, caveat, or question.
Zafar et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: