Explainable machine learning for predicting 28-day mortality in ICU patients with stroke complicated by chronic kidney disease

Patients with stroke complicated by chronic kidney disease (CKD) face high mortality in the intensive care unit (ICU) due to complex “brain-kidney crosstalk.” We aimed to develop and externally validate an explainable machine learning (ML) model based on routine bedside indicators to predict 28-day all-cause mortality, in strict adherence to the TRIPOD-AI reporting guidelines. Clinical data were extracted from the MIMIC-IV and MIMIC-III (external validation) databases. To prevent data leakage, the internal cohort was partitioned into training (70%) and testing (30%) sets before any preprocessing. Class imbalance was addressed using the Synthetic Minority Over-sampling Technique (SMOTE) exclusively within the training set. Core predictors were identified using a dual dimensionality-reduction strategy that combined LASSO regression and the Boruta algorithm. Thirteen ML algorithms were evaluated, and the optimal model’s discriminative performance was formally compared to traditional severity scores. The Shapley Additive exPlanations (SHAP) framework was utilized for clinical interpretability. A total of 1,550 eligible patients were included. The feature selection pipeline identified 14 primary physiological and biochemical baseline indicators. The Random Forest (RF) model demonstrated stable generalizability, yielding an area under the curve (AUC) of 0.654 in the internal testing set and 0.672 in the external validation cohort. SMOTE application substantially improved the sensitivity for identifying non-survivors (66.7% and 55.6% in the respective cohorts), albeit with an expected trade-off in specificity. DeLong’s test confirmed that the RF model performed comparably to complex clinical scores (e.g., SOFA, SAPS II) with no statistically significant difference (all P > 0.05). SHAP analysis identified creatinine, white blood cell count, anion gap, heart rate, and respiratory rate as the top five drivers of mortality, elucidating critical non-linear pathological thresholds. By relying solely on 14 easily accessible routine clinical indicators, the proposed RF model offers a streamlined and transparent alternative to complex traditional scoring systems. It offers data-driven, non-linear insights for early high-risk screening and individualized clinical decision-making in critically ill stroke patients.