Development and external validation of interpretable machine learning models for 7-, 14-, and 28-day mortality in sepsis-associated liver injury

Mortality in sepsis-associated liver injury (SALI) remains high, and early risk stratification is difficult. Existing prognostic models commonly focus on a single endpoint, usually 28-day mortality, and rarely undergo independent external validation.

We developed machine-learning models using clinical features recorded during the first 24 h to predict all-cause in-hospital mortality within 7, 14, and 28 days. In MIMIC-IV (v3.1), 1,550 adults meeting the prespecified SALI criteria were allocated to a training cohort (n = 1,119) and a held-out internal validation cohort (n = 431). An independent Chinese cohort from Xiaogan Central Hospital (n = 229; 28-day in-hospital mortality, 45.9%) was used for external validation. One hundred candidate pipelines were evaluated using stratified 10-fold cross-validation. SHAP was used to describe feature attribution across prediction horizons.

AUROCs in the held-out internal validation cohort were 0.840, 0.841, and 0.850 for 7-, 14-, and 28-day mortality, respectively; corresponding external AUROCs were 0.720, 0.717, and 0.763. The external 28-day model outperformed modified SOFA (0.763 vs. 0.673; P = 0.011) and MELD-Na (0.763 vs. 0.636; P = 0.002). A parsimonious logistic-regression sensitivity model achieved an external 28-day AUROC of 0.762, not significantly different from the full model (P = 0.970). The 14-day model was miscalibrated externally; recalibration improved the calibration slope from 0.563 to 0.970 but did not materially improve the Brier score. SHAP rankings varied across horizons and were interpreted descriptively rather than causally.

The models showed moderate transportability to one independent retrospective cohort. Multicentre prospective validation is needed before clinical implementation.