Smartwatch-based machine learning for the detection of cardiac dysfunction

Background Asymptomatic heart failure, a target for early intervention per guideline staging, is frequently missed without systematic screening. Objective We aimed to develop a machine learning model based on smartwatch monitoring technology to screen individuals for cardiac dysfunction risk. Methods In this prospective cohort study, we developed a real-time cardiac dysfunction risk model based on smartwatch monitoring technology. A total of 270 patients underwent synchronized acquisition of photoplethysmography (PPG) and electrocardiography (ECG) signals with smartwatch, with contemporaneous echocardiography as the golden standard for model labels between December 6, 2024 and November 11，2025 in a Chinese Hospital. The cohort was randomly divided into a training set (n = 198) and an internal validation set (n = 49). Cardiac dysfunction was defined as ejection fraction (EF) < 50%. Leveraging multimodal feature engineering in conjunction with SHAP feature importance analysis, we developed an XGBoost classifier for enhanced risk screening.The model was rigorously validated by its discriminative power (the area under the receiver operating characteristic curve, AUC-ROC and AUC-PR), clinical utility (sensitivity, specificity, PPV, NPV), calibration (calibration curve and Brier score), and generalizability (in an external cohort, n=93). Results The mean age was 55 ± 14 years (62.6% male) in the training cohort, 57 ± 14 years (53.1% male) in the validation cohort, and 47 ± 21 years (49.5% male) in the testing cohort; the corresponding mean ejection fractions (EFs) were 53.93 ± 13.12%, 55.51 ± 12.31%, and 53.88 ± 14.16%, respectively. The numbers (proportions) of participants with EF<50% were 64 (32.3%), 16 (32.7%), and 39 (41.9%) in the training, validation, and testing cohorts. The patients with asymptomatic or mildly symptomatic HF (NYHA I–II) accounted for 26.25%. Among 27 ECG features and 26 PPG features, QRS duration, QT interval and PPG velocity features were predominantly concentrated in the positive SHAP region, indicating a positive association with increased cardiac risk (Figure 1). The AUCs of cardiac dysfunction model were 0.99 in training cohort, 0.93 in internal validation cohort, and 0.95 in external cohort (Figure 2). The accuracy, precision, sensitivity, specificity, and F1 score reached 0.96, 0.94, 0.92, 0.98, and 0.93 in training cohort, respectively, while 0.87, 0.93, 0.74, 0.96, and 0.83 in external validation cohort, respectively. Conclusion The cardiac dysfunction risk model demonstrated excellent discriminative ability (AUC 0.93–0.99) and good calibration across internal and external cohorts, highlighting its potential to facilitate guideline-recommended early detection of asymptomatic heart failure, although further prospective validation in broader populations is warranted.12