Estimating stroke volume using consumer-grade smartwatches: toward predictive, real-world cardiac assessment

Background

The clinical assessment of stroke volume (SV) remained confined to hospital settings due to its reliance on non-portable techniques, creating a critical gap in continuous hemodynamic monitoring.

Objective

The present study aimed to develop and validate an SV estimation model based on consumer-grade smartwatch technology, enabling non-invasive and accessible cardiac assessment in real-world environments.

Methods

Total 114 participants were prospectively enrolled between December 6, 2024 and November 11，2025, from a Chinese hospital. Participants wore smartwatches for ECG and PPG acquisition, while physicians concurrently conducted cardiac ultrasound to assess parameters including stroke volume (SV). A multi-model framework was developed using rigorously signal-quality-checked ECG and PPG data from smartwatches, along with demographic information. Separate steady-state and dynamic tracking models were constructed via XGBoost, and their outputs were subsequently fused to generate the final stroke volume estimations.

To assess the model’s predictive power and clinical applicability, we evaluated its accuracy, trend-tracking ability, clinical acceptability, and capacity to distinguish adequate cardiac function. Among 114 participants, 86 subjects were randomly selected as training cohort, and 28 subjects as internal validation cohort. The model was further validated in external cohort of 76 subjects.

Results

The mean age was 49 ± 15 years (50.0% male) in the training cohort, 48 ± 18 years (39.3% male) in the validation cohort, and 36 ± 15 years (51.3% male) in the testing cohort; the corresponding mean SVs were 62.58 ± 14.16 ml, 63.20 ± 13.37ml, and 63.33 ± 12.816 ml, respectively.

The model demonstrated high absolute accuracy, with a Mean Absolute Error (MAE) of 5.63 ml/m² and a Root Mean Square Error (RMSE) of 6.75 ml/m². For trend prediction, it achieved a moderate-to-strong correlation with the gold standard (adjusted R² = 0.63). Clinical agreement analysis revealed a small mean bias (1.28 ml/m²) and excellent acceptability, with a percentage error of 9.97% within the 95% limits of agreement(Figure 1). Furthermore, the model showed robust discriminative capacity for identifying adequate cardiac function (SV > 60 ml), with an Area Under the Curve (AUC) ranging from 0.822 to 0.885 across training and validation cohorts （Table 1）.

Conclusion

We present the first multimodal model leveraged consumer smartwatch ECG and PPG signals to enable non-invasive estimation of SV, demonstrating strong predictive performance against the echocardiographic gold standard. By overcoming the critical challenges of motion artifact and the sensitivity-specificity trade-off inherent in traditional single-modality wearables, it paved the way for early detection of decompensation in heart failure and personalized management of cardiovascular disease.