Prediction of heart failure readmission risk based on survival machine learning: a prospective cohort study integrating novel wearable cardiac acoustic multimodal parameters and clinical indicators

Background Heart failure (HF) readmission remains a major clinical and economic burden. Early identification of high-risk patients is crucial for optimizing management and allocating resources. This study aimed to develop and validate a survival machine learning model integrating clinical, laboratory, echocardiographic, and wearable acoustic cardiographic indicators using an innovative wearable synchronous cardiac acoustic and electrocardiographic monitoring device to predict 180-day readmission risk in HF patients. Methods This prospective cohort study enrolled 498 patients hospitalized for HF between June 2024 and February 2025. Data were extracted from wearable monitoring devices, electronic medical records, questionnaires, and laboratory/imaging examinations. The primary outcome was HF-related readmission within 180 days after discharge. After handling missing data with multiple imputation, feature selection was performed using LASSO Cox regression and the Boruta algorithm, with variance inflation factor (VIF) analysis confirming the absence of collinearity. Seven survival machine learning algorithms were trained on a 7:3 split training set and evaluated on an independent test set. Model performance was assessed using the C-index, Brier score, time-dependent AUC, calibration curves, and decision curve analysis (DCA). Interpretability was evaluated using Shapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). An interactive web-based calculator was developed for clinical deployment. Results The 180-day HF readmission rate was 9.6% (48/498). The random survival forest (RSF) model achieved the best performance, with a C-index of 0.915 in the training set (Brier score 0.085) and 0.818 in the test set (Brier score 0.072). Key predictors included New York Heart Association (NYHA) functional class, total calcium, left atrial diameter, QRS duration, left ventricular ejection fraction (LVEF), and triglycerides. SHAP analysis revealed that novel digital biomarkers, including LVET% and EMAT%, demonstrated superior predictive performance compared to invasive or costly indicators such as BNP and LVEF. The integration of novel acoustic cardiographic indices, particularly EMAT%, provided significant incremental predictive value (NRI = 0.7318). A publicly accessible web-based calculator was developed. Conclusions The RSF model integrating routinely available clinical data with wearable multimodal acoustic cardiographic parameters effectively predicts 180-day HF readmission risk. The model demonstrates good discrimination, calibration, and clinical utility, serving as a decision-support tool for early risk stratification. LVET% and EMAT%, as non-invasive and easily monitored indicators, show promising utility for community-based population monitoring.