[Development and validation of an interpretable machine learning model for stage prediction of sepsis-associated...

To develop a stage prediction model for sepsis-associated acute kidney injury (SA-AKI) based on interpretable machine learning, thereby providing support for individualized clinical intervention.

Adult patients with sepsis were selected from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database and the eICU-Collaborative Research Database (eICU-CRD). Intensive care unit (ICU) admission was defined as time zero. Demographic data were extracted, along with clinical variables collected within 24 hours after ICU admission, including vital signs, laboratory parameters, organ function scores, and therapeutic interventions. The occurrence of acute kidney injury (AKI) and its highest stage within 7 days after ICU admission were used as the study outcomes. Data from the MIMIC-IV database were divided into training and internal validation sets in a 7 : 3 ratio, stratified by AKI stage, while data from eICU-CRD were used as the external validation set. LASSO regression and multivariable ordinal Logistic regression were used to select features associated with SA-AKI stage. Six machine learning models were developed, including eXtreme Gradient Boosting, Light Gradient Boosting Machine (LightGBM), random forest, K-nearest neighbors, support vector machine, and logistic regression. The optimal model was selected based on the area under the receiver operator characteristic curve (AUC), accuracy, F1-score, precision, specificity, and sensitivity. Calibration curve and decision curve analysis (DCA) were used to evaluate model performance, and SHapley Additive exPlanations (SHAP) were used to interpret the optimal model.

A total of 7 415 and 2 727 patients were included from the MIMIC-IV and eICU-CRD databases, respectively. Thirteen features were selected using LASSO regression and multivariable ordinal Logistic regression, including 24-hour urine output [odds ratio (OR)=0.183, 95% confidence interval (95%CI) was 0.168-0.200, P<0.001], body mass index (OR=2.016, 95%CI was 1.887-2.155, P<0.001), invasive mechanical ventilation (OR=1.395, 95%CI was 1.315-1.478, P<0.001), congestive heart failure (OR=1.405, 95%CI was 1.321-1.494, P<0.001), male (OR=1.088, 95%CI was 1.028-1.152, P=0.004), mean serum potassium level (OR=1.109, 95%CI was 1.044-1.178, P<0.001), mean activated partial thromboplastin time (OR=1.212, 95%CI was 1.141-1.288, P<0.001), maximum lactate level (OR=1.373, 95%CI was 1.282-1.470, P<0.001), mean body temperature (OR=0.875, 95%CI was 0.823-0.930, P<0.001), minimum pulse oxygen saturation (OR=0.811, 95%CI was 0.763-0.862, P<0.001), diabetes mellitus (OR=1.131, 95%CI was 1.067-1.199, P<0.001), maximum total bilirubin level (OR=1.345, 95%CI was 1.246-1.451, P<0.001), and vasoactive drug use (OR=1.160, 95%CI was 1.092-1.233, P<0.001). Compared with other models, LightGBM was identified as the best-performing model, and its predicted probabilities were well calibrated. The model achieved AUCs of 0.872 and 0.851 for predicting SA-AKI stage in the internal and external validation sets. In the internal validation set, the AUCs for predicting non-AKI and AKI stages 1, 2 and 3 were 0.909, 0.837, 0.838, and 0.904, respectively, while the corresponding AUCs in the external validation set were 0.837, 0.818, 0.815, and 0.937. The DCA results showed that the highest net benefit was achieved within a threshold range of 0.1-0.8 for patients with non-AKI and AKI stage 2 and stage 3, and within a threshold range of 0.1-0.4 for patients with AKI stage 1. Differences in feature importance across SA-AKI stages were revealed by SHAP analysis. The 24-hour urine output was ranked first in most stages, whereas invasive mechanical ventilation was identified as the most important feature among non-AKI patients in the external validation set.

The LightGBM-based model for predicting SA-AKI stages was successfully developed, and key features affecting each stage were identified using SHAP, providing an interpretable tool for the clinical identification of different SA-AKI stages.