Machine learning‐based model for worsening heart failure risk in Chinese chronic heart failure patients

Abstract

AbstractAimsThis study aims to develop and validate an optimal model for predicting worsening heart failure (WHF). Multiple machine learning (ML) algorithms were compared, and the results were interpreted using SHapley Additive exPlanations (SHAP). A clinical risk calculation tool was subsequently developed based on these findings.Methods and resultsThis nested case–control study included 200 patients with chronic heart failure (CHF) from the China‐Japan Friendship Hospital (September 2019 to December 2022). Sixty‐five variables were collected, including basic information, physical and chemical examinations, and quality of life assessments. WHF occurrence within a 3‐month follow‐up was the outcome event. Variables were screened using LASSO regression, univariate analysis, and comparison of key variables in multiple ML models. Eighty per cent of the data was used for training and 20% for testing. The best models were identified by integrating nine ML algorithms and interpreted using SHAP, and to develop a final risk calculation tool. Among participants, 68 (34.0%) were female, with a mean age (standard deviation, SD) of 68.57 (12.80) years. During the follow‐up, 60 participants (30%) developed WHF. N‐terminal pro‐brain natriuretic peptide (NT‐proBNP), creatinine (Cr), uric acid (UA), haemoglobin (Hb), and emotional area score on the Minnesota Heart Failure Quality of Life Questionnaire were critical predictors of WHF occurrence. The random forest (RF) model was the best model to predict WHF with an area under the curve (AUC) (95% confidence interval, CI) of 0.842 (0.675–1.000), accuracy of 0.775, sensitivity of 0.900, specificity of 0.833, negative predictive value of 0.800, and positive predictive value of 0.600 for the test set. SHAP analysis highlighted NT‐proBNP, UA, and Cr as significant predictors. An online risk predictor based on the RF model was developed for personalized WHF risk assessment.ConclusionsThis study identifies NT‐proBNP, Cr, UA, Hb, and emotional area scores as crucial predictors of WHF in CHF patients. Among the nine ML algorithms assessed, the RF model showed the highest predictive accuracy. SHAP analysis further emphasized NT‐proBNP, UA, and Cr as the most significant predictors. An online risk prediction tool based on the RF model was subsequently developed to enhance early and personalized WHF risk assessment in clinical settings.