Application of a machine learning model integrating T cell subsets and clinical markers in 28-Day mortality prediction of sepsis patients.

To develop a machine learning-based prediction model that integrating T cell subset data with clinical features to predict the 28-day mortality risk in sepsis patients. A retrospective cohort study was conducted using the MIMIC-IV database. Collected data included demographics, T cell subsets, laboratory results, SOFA score, GCS, and 28-day mortality. Feature selection was performed using LASSO regression combined with 10-fold cross-validation. We compared the performance of six machine learning models, namely RF, SVM, XGB, GLM, GBM, and LASSO logistic regression. Model performance was evaluated using the AUROC, calibration curves, and DCA, while the SHAP method was employed to interpret the optimal model. A total of 781 sepsis patients were included in our study, with a 28-day mortality rate of 18.5%. LASSO regression identified 12 key features: Age, CD + 4 Count, CD8 + T cell count, lactate, platelet, albumin, Na+, Bun, heart rate, anion gap, eosinophil count, monocyte count. Among the six compared machine learning models, GLM achieved the best comprehensive performance in the testing set, with an AUROC of 0.720, good calibration (Brier score: 0.134). SHAP analysis clarified the contribution degree and direction of each predictive factor to the model output. A GLM model integrating T cell subsets (notably CD8 + T cell count) and clinical features was successfully constructed. The model showed moderate discriminative ability and potential clinical application value in predicting the 28-day mortality risk of sepsis. SHAP-based interpretability clarifies individual risk factors, serving as an auxiliary prognostic tool for clinicians.