External validation of a machine learning-based classification algorithm for ambulatory heart rhythm diagnostics in...

The aim of this study was to perform an external validation of an automatic machine learning (ML) algorithm for heart rhythm diagnostics using smartphone photoplethysmography (PPG) recorded by patients with atrial fibrillation (AF) and atrial flutter (AFL) pericardioversion in an unsupervised ambulatory setting.

Patients undergoing cardioversion for AF or AFL performed 1-min heart rhythm recordings pericardioversion at least twice daily for 4-6 weeks, using an iPhone 7 smartphone running a PPG application (CORAI Heart Monitor) simultaneously with a single-lead electrocardiogram (ECG) recording (KardiaMobile). The algorithm uses support vector machines to classify heart rhythm from smartphone-PPG. The algorithm was trained on PPG recordings made by patients in a separate cardioversion cohort. Photoplethysmography recordings in the external validation cohort were analysed by the algorithm. Diagnostic performance was calculated by comparing the heart rhythm classification output to the diagnosis from the simultaneous ECG recordings (gold standard). In total, 460 patients performed 34 097 simultaneous PPG and ECG recordings, divided into 180 patients with 16 092 recordings in the training cohort and 280 patients with 18 005 recordings in the external validation cohort. Algorithmic classification of the PPG recordings in the external validation cohort diagnosed AF with sensitivity, specificity, and accuracy of 99.7%, 99.7% and 99.7%, respectively, and AF/AFL with sensitivity, specificity, and accuracy of 99.3%, 99.1% and 99.2%, respectively.

A machine learning-based algorithm demonstrated excellent performance in diagnosing atrial fibrillation and atrial flutter from smartphone-PPG recordings in an unsupervised ambulatory setting, minimizing the need for manual review and ECG verification, in elderly cardioversion populations.

Clinicaltrials.gov, NCT04300270.