Prediction Model for Delirium in Patients with Sepsis-Associated Liver Injury: An Interpretable Machine Learning Approach

Background Patients with sepsis-associated liver injury (SALI) are at marked risk of delirium, a severe complication strongly linked with poor neurological outcomes. Early identification remains challenging, as existing predictive tools lack specificity for this distinct population. An interpretable machine learning (ML) model was designed and validated to enable prediction of delirium among SALI patients. Methods De-identified data from MIMIC-IV were retrospectively assessed in this cohort study. The dataset was randomly partitioned, with 70% assigned to model training and 30% reserved for evaluation. Independent predictors were identified through univariate analysis followed by stepwise multivariable logistic regression. Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), Gradient Boosting Machine (GBM), and Extreme Gradient Boosting (XGBoost) were constructed and tested regarding discrimination, calibration, and clinical usefulness. Model interpretation was performed using the Shapley Additive Explanations (SHAP) framework. Results Among 1461 patients with SALI, 917 (62.8%) developed delirium. Seven independent risk factors were identified: diabetes with chronic complications, reduced SpO 2 , decreased hemoglobin, lower Glasgow Coma Scale (GCS) score, and treatment with continuous renal replacement therapy (CRRT), vasopressin, or mechanical ventilation. The GBM model demonstrated optimal performance, achieving an area under the receiver operating characteristic curve (AUROC) of 0.831 in the training set and 0.811 (95% CI: 0.766-0.855) in the testing set. SHAP analysis revealed that mechanical ventilation, GCS score, CRRT requirement, and hemoglobin levels were the most influential predictors, indicating that delirium risk is primarily driven by markers of multi-organ dysfunction. Conclusions An interpretable gradient boosting model was established and assessed for its ability to predict delirium among individuals with SALI. The model's transparency, achieved through SHAP analysis, establishes clear associations between delirium and quantifiable markers of multi-organ dysfunction. This tool enables early, individualized risk stratification, facilitating targeted preventive interventions to improve outcomes in this vulnerable population.