Interpretable machine learning prediction of in-hospital mortality in people with HIV: A cohort study in southeastern China

Background Hospitalized individuals with human immunodeficiency virus (HIV) remain at high risk of in-hospital mortality despite advances in combination antiretroviral therapy. Accurate and interpretable prediction tools are needed to support timely risk stratification and clinical decision-making. Methods We conducted a retrospective cohort study of 2015 adults with HIV admitted to the AIDS ward of a tertiary hospital in southeastern China between 2010 and 2021. Clinical and laboratory data at admission were used to develop machine learning models predicting in-hospital mortality. Eleven algorithms were compared using area under the receiver operating characteristic curve (AUC), precision-recall curve, and decision curve analysis. Model interpretability was achieved using Shapley Additive exPlanations (SHAP). To enhance usability, we reduced predictors through SHAP-guided selection and deployed the final model as an interactive web application. Results Among 2015 patients, 293 (14.54%) died during hospitalization. The Light Gradient Boosting Machine model demonstrated the best performance (AUC = 0.9008). A simplified model incorporating five admission variables, cancer antigen 125, direct bilirubin, white blood cell count, C-reactive protein, and age, achieved comparable accuracy (AUC = 0.9008). SHAP analysis provided transparent explanations at both the population and individual levels. The final model was implemented as a user-friendly online tool for individualized risk estimation. Conclusions We developed and validated an interpretable machine learning model that accurately predicts in-hospital mortality among patients with HIV using five readily available clinical features. This approach enhances transparency, supports clinical decision-making, and may improve early identification of high-risk patients.