Temporal Validation of Explainable Machine Learning for 28-Day Mortality Prediction in Adults with Peripheral Artery Disease in Intensive Care

Background Whether first-24-hour hemodynamic summaries improve mortality prediction beyond established severity measures in critically ill patients with peripheral artery disease (PAD) is uncertain. We evaluated temporal model performance and adjusted hemodynamic associations with 28-day mortality. Methods This secondary analysis included first intensive care unit stays of adults with ICD-coded PAD who reached a 24-hour landmark. Models developed in 2008–2016 were evaluated without refitting in 2017–2022. Extreme Gradient Boosting (XGBoost), clinical comparators, and hemodynamic and severity-score ablations used 2000-bootstrap confidence intervals. Landmark Cox models used 20 multiply imputed datasets. Results Among 2459 patients, 435 (17.7%) died by day 28; the temporal test cohort included 535 patients and 106 deaths. Full XGBoost achieved an area under the receiver-operating-characteristic curve (AUROC) of 0.830 (95% CI 0.786–0.871), area under the precision-recall curve of 0.602 (0.524–0.682), Brier score of 0.117 (0.104–0.130), and expected calibration error of 0.029. The model without hemodynamics had a numerically higher AUROC than the full model (0.837 vs. 0.830), but the paired difference was imprecise and compatible with small changes in either direction (full minus no hemodynamics, − 0.007; 95% CI − 0.024 to 0.010). Higher shock index, lower pulse pressure, and greater mean arterial pressure variability were associated with mortality. The composite was not independently associated (HR, 1.07; 95% CI 0.82–1.39). Conclusions The model retained good discrimination in an internal temporal test cohort, but early hemodynamic summaries did not demonstrate measurable incremental predictive value. Individual measures were risk markers, not validated treatment targets. External validation is required.