A machine learning model for predicting 28-day mortality in patients with alcoholic cirrhosis and sepsis: a study based on the MIMIC-IV database.

OBJECTIVE

To develop and validate the best machine learning model for predicting 28-day mortality in septic patients with alcoholic cirrhosis (AC), leveraging data from the Medical Information Mart for Intensive Care Database, 4th Edition (MIMIC-IV).

METHODS

Clinical data of 1958 patients with AC complicated with sepsis were retrospectively extracted. Missing data (<20%) were addressed using Multiple Imputation by Chained Equations (MICE). Key predictors were selected via Least Absolute Shrinkage and Selection Operator (LASSO) regression. Patients were allocated to a training set (n=1268), a testing set (n=318), and a temporal validation set (n=372). Five models were developed and evaluated based on different performance metrics. Decision Curve Analysis (DCA) and SHapley Additive exPlanations (SHAP) were performed to assess clinical applicability and predictor importance.

RESULTS

LASSO regression identified 15 core variables. The eXtreme Gradient Boosting (XGBoost) model achieved the best overall performance, with an area under the ROC curve (AUC) of 0.946 (95%CI [0.932-0.960]) (sensitivity: 0.891) on the training set and an AUC of 0.878 (95% CI [0.838, 0.918]) on the test set, and an AUC of 0.819 on the validation set. The model was well-calibrated and provided a higher net benefit than 'treat-all' or 'treat-none' strategies across wide risk thresholds. SHAP analysis revealed the top 5 predictors to be: Sequential Organ Failure Assessment (SOFA) score, Model for End-Stage Liver Disease (MELD) score, age, temperature, and SPO2.

CONCLUSION

The XGBoost model had the best predictive performance, providinga robust tool to guide personalized therapy and optimize critical care resource allocation.