Predictive Modeling of Hospital Readmissions with Sparse Bayesian Extreme Learning Machine

Abstract

Patients at high risk of readmissions are a strain to healthcare system and contribute to bed shortage. A tool that accurately identifies patients at high risk of hospital readmission is essential. In this study, we aimed to build a predictive model using machine learning to assess the risk of 30-day readmissions. We retrospectively collected a cohort of admitted patients from Singapore General Hospital. We extracted both clinical and administrative data from hospital electronic health records. The primary outcome was 30-day hospital readmissions. We used sparse Bayesian extreme learning machine (SBELM) to build a predictive model where patient demographics, indicators of socioeconomic status, prior healthcare utilization, markers of acute and chronic illness burden were included as predictors. We compared our model with the established LACE index in terms of predictive performance using receiver operating characteristic (ROC) analysis. Our SBELM model achieved better performance (AUC = 0.762 [95% confidence interval: 0.747–0.778], 70.0% sensitivity, 68.8% specificity, 29.8% positive predictive value (PPV) and 92.4% negative predictive value (NPV) at cutoff of 0.137) compared with the LACE index (AUC = 0.716 [95% confidence interval: 0.699–0.732], 65.9% sensitivity, 66.8% specificity, 27.3% PPV and 91.2% NPV at cutoff score of 9). Our model shows promising discriminatory ability and may be useful in further developing multidisciplinary interventions to prevent repeated hospital admissions.