Explainable machine learning for predicting infections that require hospitalization in patients with systemic lupus erythematosus

OBJECTIVES: To develop and validate an explainable artificial intelligence (XAI)-based machine learning (ML) model for predicting infections requiring hospitalization in patients with systemic lupus erythematosus (SLE). METHODS: Outpatient data at Taichung Veterans General Hospital (2012-2023) were analyzed retrospectively. Infections were identified using ICD-9 and ICD-10 codes. Clinical variables, including demographic information, laboratory results, and medication history were used to train ML algorithms. Lasso regression was used for feature selection and narrowing the model to key predictors. Model performance was evaluated using AUROC, sensitivity, specificity, precision, and negative predictive value (NPV). RESULTS: Data from 18 331 patients were included in the study. The number of key predictors was narrowed to 15 from 70 clinical variables. XGBoost achieved the highest predictive performance, with an AUROC of 88% in the testing cohort and 84% in the temporal validation cohort. Shapley Additive Explanations values were used to interpret model outputs, identifying corticosteroid use, haemoglobin level, and white blood cell count as the top predictors. The model demonstrated high specificity (85%) and excellent NPV (98%). CONCLUSION: We developed an interpretable ML model as a supporting tool for predicting serious infections in SLE patients. The integration of SHAP-based explanations enhances clinical transparency and supports early identification of high-risk individuals, although the single-center, retrospective design warrants external validation in future studies.