The performance of preliminary version of irregular heart rhythm notification algorithm of a consumer smartwatch for screening of atrial fibrillation

Abstract Background Smartwatch photoplethysmography (PPG) is a useful non-invasive tool for atrial fibrillation (AF) screening. However, the process of signal acquisition is vulnerable to variable factors such as motion, respiration, and unstable sensor contact which may lead to false positive results. Thus, an optimal signal acquisition and processing algorithm is needed to balance the inherent trade-off between sensitivity and specificity for smartwatch PPG to be used for AF screening in daily life. Purpose This study aimed to evaluate the efficacy of the preliminary version of the Irregular Heart Rhythm Notification (IHRN) feature of a consumer smartwatch for AF screening. Methods We obtained time-synchronized PPG and ECG signals using smartwatches and wearable ECG patches in 49 patients with paroxysmal and persistent AF in a free-living setting for up to 7 days. The IHRN algorithm selectively schedules PPG tachogram data acquisition at specific time interval, classifies heart rhythm for each interval, and integrates the classification results to determine whether to alert the user of their irregular heart rhythm. We tested the diagnostic performance of the IHRN algorithm for identifying AF episodes lasting 30 minutes or more. Results We obtained a total of 364,260 minutes of ECG recording and time-synchronized 27,524 minutes of PPG signal in 49 AF patients. Out of the PPG signals collected, 19,341 minutes (70.3%) were considered valid for the analysis. On the ECG analysis, 754 AF episodes (≥30 seconds) were identified. Longer-lasting AF episodes had a higher likelihood of being captured by PPG and consequently classified as AF by the algorithm (Figure 1). AF episodes lasting over 30 minutes were identified in 28 out of 49 patients. Regarding the subjective-level analysis, the sensitivity and specificity of IHRN for AF ≥ 30 minutes were 82.1% and 100.0%, respectively. In tachogram-level analysis, the positive predictive value was 98.7%. Conclusions The preliminary version of the IHRN algorithm of the smartwatch demonstrated excellent performance in identifying AF in a free-living setting. Further large-scale research is needed to determine whether this algorithm is useful for AF screening in a real-world setting.Figure 1