AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Using Smartwatch Photoplethysmography to Identify CHA ₂ DS ₂ -VASc Risk: Subanalysis from the HEARTBEAT Study
By Massad C, Tsakiris E, Liu Y et al.
Background: Wearable photoplethysmography (PPG) from wearable devices offers continuous, noninvasive cardiovascular data. While PPG-derived heart rate and rhythm detection are well-validated, whether passive PPG signal morphology encodes broader thromboembolic and cardiovascular risk burden remains largely unexplored. Methods: Participants enrolled in the HEARTBEAT Study with available first-month passive smartwatch PPG recordings were included. PPG signals were processed using previously trained deep-learning encoder, and patient-level embedding means and standard deviations were used as model inputs. A shallow random-forest classifier was trained to distinguish high-thromboembolic risk (CHA ₂ DS ₂ -VASc ≥4) from low-risk (CHA ₂ DS ₂ -VASc ≤1) patients. Performance was evaluated on a held-out validation cohort (n=72) and further assessed via internal 5-fold cross-validation (n=401). Area under the precision-recall curve (AUPRC) was emphasized as the primary performance metric, with area under the receiver operating characteristic curve (AUROC) reported as a complementary measure. Results: On the held-out validation cohort, the random-forest classifier achieved AUPRC 0.819 and AUROC 0.813 for distinguishing participants with high vs low CHA ₂ DS ₂ -VASc scores. Internal 5-fold CV confirmed stability, with tuned random-forest models achieving mean-fold AUPRC up to 0.825 and mean-fold AUROC of approximately 0.77. Overall, AUPRC remained consistently strong across the held-out and cross-validation analysis, and AUROC further supported model discrimination and risk stratification. Conclusion: These findings suggest that wearable PPG morphology encodes cardiovascular risk information beyond just heart rate or arrhythmia detection.