Predicting hospital readmission from electronic health records in heart failure patients : a machine learning approach

Abstract

Introduction: Unplanned hospital readmissions pose a burden on both patients and healthcare organisations, leading to adverse health outcomes and increased healthcare costs. Heart failure is associated with high unplanned readmission rates. The prediction of hospital readmission can therefore help healthcare providers identify patients with the highest risk of readmission and ensure closer follow-up monitoring and care coordination to prevent readmission. This study utilised a Machine Learning (ML) approach to develop and validate a readmission risk prediction model predicting thirty-day hospital readmission, focusing on patients with heart failure. Methods: Two Electronic Health Record datasets were utilised in this study: the open-access Medical Information Mart for Intensive Care IV (MIMIC-IV) and a Public Hospitals dataset, sourced from the Directorate of Health Information and Research (DHIR) in Malta. Exploratory Data Analysis of these datasets indicated that the thirty-day hospital readmission rate of heart failure patients was 21% (MIMIC-IV) and 27% (local Public Hospitals). Predictive risk variables were extracted from the datasets and four Supervised Learning models were developed in RapidMiner Studio: Decision Tree (DT), Random Forest (RF), Gradient Boosted Trees (GBT) and EXtreme Gradient Boosting (XGBoost). The models were trained and validated on data from MIMIC-IV, using ten-fold cross-validation. Results: The highest performing model for predicting thirty-day hospital readmission was the XGBoost model, which achieved performance metrics: F1 Score 77.8 and Area Under the Receiver Operator Curve (AUROC) 0.87, when trained on the MIMIC-IV Heart Failure cohort. The RF and GBT algorithms also performed well (RF: F1-Score 67.6, AUROC 0.73; GBT: F1 Score 68.9, AUROC 0.70) while the DT algorithm performed poorly possibly due to overfitting (F1 Score 37.5, AUROC 0.50). Lower performance was also observed when the model was tested on the local Public Hospitals dataset (AUROC 0.51). This was mainly attributed to an imbalanced class distribution and a lack of predictive features in the test dataset. Conclusion: Results compare well with other studies applying similar models utilising MIMIC data. This demonstrated that the proposed Machine Learning approach can be used to predict thirty-day hospital readmission through Electronic Health Record data. Since relatively high readmission rates were observed in the first week post-discharge, further studies investigating the prediction of hospital readmission in the early post-discharge period, may provide additional insights on the observed high readmission rates in heart failure patients. Disclosures: Funded in part by the Malta Digital Innovation Authority (MDIA), as part of the Pathfinder MDIA Digital Scholarship scheme.