Performance of an artificial intelligence algorithm to detect atrial fibrillation on a 24-hour continuous photoplethysmography recording using a smartwatch: ACURATE study

Abstract Background In the awakening era of mobile health, wearable devices capable of detecting atrial fibrillation (AF) are on the rise. Smartwatches and wristbands are equipped with photoplethysmography (PPG) technology that enables (semi)continuous rhythm monitoring. These devices have been pioneered already in a few screening trials. However, such devices are being spread among consumers at a pace that is not paralleled by the evidence supporting their clinical performance. This imbalance reflects the urgent need for validation studies. Purpose To determine the diagnostic performance of an artificial intelligence algorithm to detect AF using photoplethysmography acquired by a smartwatch. Methods One hundred patients (≥18 years) without a pacemaker-dependent heart rhythm who were referred to a university hospital or a large tertiary hospital for elective 24-hour ECG Holter monitoring were asked to wear a continuous PPG monitoring smartwatch (i.e. Samsung GWA2 or Empatica E4) simultaneously with the Holter. All activities of daily life were allowed. The ECG trace and PPG waveform were synchronised and fragmented in one-minute fragements. The one-minute ECG fragments were labelled as AF, non-AF, or insufficient quality based on the routine clinical interpretation of the 24-hour Holter (i.e. software + physician overreading). The one-minute PPG fragments were analysed by an artificial intelligence (AI) algorithm (i.e. FibriCheck) and were given the same labels. Diagnostic metrics of the PPG AI algorithm were calculated with respect to the ECG interpretation, for all fragments with sufficient quality for both PPG and ECG. Results Four patients had to be excluded due to technical error (3 Holter errors, 1 smartwatch error). The mean age in the remaining study population (n=96) was 59±16 years, 51 (53%) were men and 15 (15.6%) were known with permanent AF. In this population, simultaneous ECG and PPG monitoring was recorded for 115,245 one-minute fragments. Fragments of insufficient quality for ECG (n=1,454; 1.3%), PPG (n=25,704; 22.3%) or both (n=15,362; 13.3%) were excluded. PPG fragments were more frequently of insufficient quality (p<0.001). AF was present in 10,255 (14.1%) of the resulting 72,725 high-quality one-minute fragments. The sensitivity of PPG to detect AF was 93.4% (CI 92.9% - 93.8%). The specificity of PPG to exclude AF was 98.4% (CI 98.3% - 98.5%). As a result, the overall accuracy of the PPG algorithm on one-minute fragment level was 97.7% (CI 97.6%- 97.8%). Conclusion Continuous out-of-hospital PPG monitoring using a smartwatch in combination with an AI algorithm can accurately discriminate between AF and non-AF rhythms in a heterogenous patient population. PPG quality is more often affected than ECG quality during daily life activities. Funding Acknowledgement Type of funding sources: Foundation. Main funding source(s): Research Foundation-Flanders, Strategic Basic Research Fund