Machine-learning prediction of 30-day infection-related hospitalization in advanced CKD

Abstract Background Patients with advanced chronic kidney disease (CKD), defined as having an estimated glomerular filtration rate (eGFR) < 45 ml/min/1.73 m², are vulnerable to infections leading to hospitalization. Predicting such events using routine clinical data may facilitate early intervention. This study aimed to develop a machine-learning models to predict infection-related hospitalization within 30 days of a nephrology visit. Methods We conducted a single center retrospective cohort study of patients in the pre-ESRD pay-for-performance program at Taichung Veterans General Hospital (2010–2022). The primary outcome was infection-related hospitalization within 30 days of the index visit; qualifying admissions required intravenous antibiotic use. We excluded visits from patients < 18 years and those with < 2 nephrology visits in the prior year. Predictor variables included longitudinal laboratory parameters (means and variances within the prior 3 months), comorbidities, and recent medication use. Five machine learning models (XGBoost, random forest, logistic regression, support vector machine, and k-nearest neighbor) were trained and evaluated, with feature importance assessed using Shapley Additive explanations (SHAP). Results The analysis included 11,502 index visits from 11,018 patients, with 2,195 infection-related hospitalizations. Among tested models, the XGBoost algorithm achieved the highest predictive performance (AUROC 85.5%, sensitivity 71.3%, specificity 84.8%). The top ten predictors, as ranked by SHAP importance, were recent use of diuretics, serum albumin, Charlson comorbidity index (CCI) excluding renal disease, mean eGFR, corticosteroid use, albumin variance, age, eGFR variance, mean potassium, and mean uric acid. SHAP dependence plots were used to illustrate nonlinear relationships and approximate thresholds associated with increased risk of infection. Conclusions Machine learning models using routinely available outpatient data can effectively predict short-term infection-related hospitalization in advanced CKD patients. This approach has the potential to inform individualized surveillance and early intervention strategies in nephrology care. Clinical trial number Not applicable.