Machine Learning Models for Predicting Medium-Term Heart Failure Prognosis: Discrimination and Calibration Analysis

BackgroundThe number of patients with heart failure (HF) is increasing with an aging population, shifting care from hospitals to clinics. Predicting medium-term prognosis after discharge can improve clinical care and reduce readmissions; however, no established model has been evaluated with both discrimination and calibration. ObjectivesThis study aimed to develop and assess the feasibility of machine learning (ML) models in predicting the medium-term prognosis of patients with HF. MethodsThis study included 4,904 patients with HF admitted to four affiliated hospitals at Nippon Medical School (2018-2023). Four ML models--logistic regression, random forests, extreme gradient boosting, and light gradient boosting--were developed to predict the endpoints of death or emergency hospitalization within 180 days of discharge. The patients were randomly divided into training and validation sets (8:2), and the ML models were trained on the training dataset and evaluated using the validation dataset. ResultsAll models demonstrated acceptable performance as assessed by the area under the precision-recall curve. The models showed favorable agreement between the predicted and observed outcomes in the calibration evaluations with the calibration slope and Brier score. Successful risk stratification of medium-term outcomes was achieved for individual patients with HF. The SHapley Additive exPlanations algorithm identified nursing care needs as a significant predictor alongside established laboratory values for HF prognosis. ConclusionsML models effectively predict the 180-day prognosis of patients with HF, and the influence of nursing care needs underscores the importance of multidisciplinary collaboration in HF care. Clinical Trial RegistrationURL: https://www.umin.ac.jp/ctr; unique identifier: UMIN000054854