Diagnostic assessment of a deep learning system for detecting atrial fibrillation in pulse waveforms

Correspondence to Dr Ming-Zher Poh, Cardiio, Inc., Cambridge, MA 02139, USA; [email protected]

Introduction

Atrial fibrillation (AF) is associated with a third of all strokes,¹ but is asymptomatic in over one-third of patients² and often goes undiagnosed. Although treatment of patients with AF with oral anticoagulants is effective in reducing stroke risk by 60%–70%,³ nearly 25% of patients with stroke only discover the presence of AF after the potentially preventable stroke event.⁴ As the use of smartphone apps, wearable fitness trackers and smartwatches capable of acquiring pulse waveforms via photoplethysmography (PPG) becomes increasingly common, these tools may present a new avenue for early detection of undiagnosed AF and timely anticoagulant treatment to prevent stroke.

Prior work on PPG-based AF detection algorithms relied predominantly on explicit rules and handcrafted features derived from a sequence of interbeat intervals of the PPG waveform aimed at capturing pulse irregularity, the hallmark of AF. Published methods include coefficient of variation (CoV),⁵ coefficient of sample entropy (CoSEn),⁶ normalised root mean square of successive differences (nRMSSD) + Shannon entropy (ShE),⁷ nRMSSD + Poincaré plot geometry (SD1/SD2),⁸ Poincaré plot patterns⁹ and autocorrelation analysis using a support vector machine (SVM).¹⁰ Thus far, achieving both very high sensitivity and specificity remains challenging because of other arrhythmias such as ectopic beats and the presence of motion or noise artefacts in the PPG signal that can mimic AF.

In this work, we trained a deep convolutional neural network (DCNN) to distinguish between noise, sinus rhythm, ectopic rhythms and AF using a large set of PPG signals. In contrast to handcrafted features, the DCNN automatically learns the most predictive features directly from the raw PPG waveform based on the training examples.

Methods

Data sets and reference standards

To develop the DCNN, we constructed a data set (PPG-RHYTHM) from several publicly accessible PPG repositories, including the MIMIC-III critical care database,¹¹ the Vortal data set from healthy volunteers¹² and the IEEE-TBME PPG Respiratory Rate Benchmark data set.¹³ All PPG recordings were resampled to 30 Hz and divided into segments of 512 samples (approximately 17 s long). A total of 186 317 PPG segments with concurrent ECG from 3373 unique persons were analysed and assigned to one of...