An explainable machine learning model for predicting in-hospital infection in patients with systemic lupus erythematosus
TL;DR
A high-performance, explainable ML model using nine routinely available clinical variables to accurately predict in-hospital infection risk in SLE patients is developed and validated and has the potential to guide personalized clinical stratification and early intervention, ultimately improving patient outcomes.
Abstract
Abstract Infection is a leading cause of mortality in patients with systemic lupus erythematosus (SLE), yet effective tools for early identification of high-risk patients are lacking. This study aimed to develop an explainable machine learning (ML) model to predict in-hospital infection risk among SLE patients. We analyzed adult patients (≥18 years) with SLE (n = 7,833) from three departments using a population-based electronic medical record database (2000–2024). Among them, 3,157 (40.3%) patients developed an infection after 72 h of hospitalization. An initial comprehensive variable pool of 108 candidate predictors was included, encompassing demographics, comprehensive laboratory parameters, clinical features, disease activity, and treatment exposures. Ten machine learning models were applied. Model performance was evaluated using six metrics. Model interpretability was achieved using SHapley Additive exPlanations (SHAP). Nine predictors were selected: daily prednisone equivalent dose, albumin, hydroxychloroquine use, C-reactive protein, D-dimer, glucose, cystatin C, hemoglobin, and alpha1-globulin. Among all models tested, the Gradient Boosting model demonstrated the best overall performance on the independent validation set, with an area under the curve (AUC) of 0.858, with its robustness confirmed by 5-fold and 10-fold cross-validation (mean AUCs of 0.855 ± 0.002 and 0.854 ± 0.008, respectively). SHAP analysis revealed that daily prednisone equivalent dose, albumin, and hydroxychloroquine use were the most influential factors. We developed and validated a high-performance, explainable ML model using nine routinely available clinical variables to accurately predict in-hospital infection risk in SLE patients. This tool provides transparent, individualized risk assessment and has the potential to guide personalized clinical stratification and early intervention, ultimately improving patient outcomes.