Explainable AI for evaluating hospitalization outcomes based on electronic health records for chronic disease patients

Background Machine learning models can leverage electronic health record data to support predictive analytics and guide proactive interventions for chronic patients.Methods This study analyzed over 40,000 hospital admissions for chronic obstructive pulmonary disease (COPD), liver disease, diabetes, and acute myocardial infarction (AMI) using electronic health records (EHR) data collected in the emergency department. Prediction models were developed to estimate three key outcomes: 30-day readmission, medium or long length of stay, and single-day admission. To enhance interpretability, the study employed several explainable AI techniques, including SHapley Additive exPlanations (SHAP), and introduced two new visualizations, the SHAP Exploration Plot and the Unified SHAP Exploration Plot, which provide refined insights into feature contributions.Results Among the target outcomes, 30-day readmission consistently yielded the highest predictive performance across all diagnoses, followed by single-day admission and medium or long length of stay (average AUCs of 0.71, 0.67, and 0.64, respectively). SHAP was used to identify both global and patient-level predictors, including age, number of admissions, Charlson Comorbidity Index, gender, EHR screen use, and insurance status.Conclusions These findings can inform prevention-oriented medical decision-making tailored to individual patients and subgroups, potentially improving outcomes in the management of chronic diseases.