Cardiac arrhythmias are a critical health problem associated with a variety of heart-related complications, such as stroke or heart failure. Due to their transient and asymptomatic nature in the early stages, arrhythmias can go undetected with current screening methods relying on electrocardiography (ECG). With recent advances in photoplethysmography (PPG), wearables using this optical sensing technique have shown a large potential for detecting cardiac abnormalities in large populations.
Herein, we propose to assess the performance of an abnormal rhythm classifier based on PPG signals. The algorithm first extracts the sequence of interbeat intervals (IBIs) from a 30-second PPG window and classifies it as either normal or abnormal. With the aim of its embedding into a wearable device, the classifier architecture was kept simple and consists of two gated recurrent unit (GRU) layers. It has been trained on IBIs extracted from ECG signals from 283 subjects. The algorithm was evaluated on 54 patients undergoing a diagnostic or therapeutic electrophysiological procedure at the Lausanne University Hospital, Switzerland. This dataset includes the recording of a PPG signal from a wrist-bracelet and a simultaneous 12-lead ECG. The latter is used as gold-standard for cardiac rhythm annotation.
Although some cardiac rhythms, such as atrial fibrillation, atrial tachycardia, and atrioventricular re-entry tachycardia, are well detected as abnormal, the classification is more difficult for slow arrhythmias, such as atrioventricular blocks. Additionally, bigeminies are often misclassified as normal. This mainly comes from the non-detection of premature contractions by the pulse detector resulting in long and regular IBIs. Overall, the algorithm achieves a sensitivity of 0.83 and specificity of 0.71 for abnormal rhythms detection. These results indicate that this algorithm can be used to detect abnormal rhythms in a sensitive cohort and show its potential application for cardiac arrhythmias screening in large populations.