Predicting 30-day readmissions from skilled nursing facilities using interoperable FHIR data and explainable machine learning

Skilled Nursing Facilities (SNF) hospital readmissions continue to be a significant issue in terms of healthcare quality, patient safety and cost management in the Centres for Medicare and Medicaid Services (CMS) Hospital Readmissions Reduction Program (HRRP). A large number of SNFs do not have sophisticated analytical software to integrate clinical and social data to determine high-risk residents of early readmission. By training and testing a machine learning model that is interpretable and based on interoperable Fast Healthcare Interoperability Resources (FHIR) data, this study will fulfill this gap and predict 30-day hospital readmissions among SNF residents. The analysis was based on de-identified, FHIR-mapped data of 14,250 SNF residents, namely medications, vital sign, functional status, prior utilisation and social risk indicators. The gradient-boosted machine (GBM) model was constructed and compared to a basis of logistic regression. The performance of the models was assessed in terms of the AUROC, AUPRC, calibration analysis, and the decision curve analysis. The explainability was done by SHapley Additive exPlanations (SHAP) which allowed transparent understanding of the individual risk factors. SHAP analysis gave easily understandable, clinically significant explanations, which justified actionable care planning. The unmanned pilot ensured stable performance over a period of time with slight drift. On the whole, this paper proves that interoperable FHIR data combined with explainable machine learning can help to make SNFs predict readmission risks ethically, transparently, and effectively. The strategy complies with policy, privacy and quality improvement objectives, and provides value to work conveniently to clinicians, administrators and policymakers aiming to minimize preventable hospital readmissions.