Multiclass Arrhythmia Detection and Classification From Photoplethysmography Signals Using a Deep Convolutional Neural Network

Abstract

Background Studies have reported the use of photoplethysmography signals to detect atrial fibrillation; however, the use of photoplethysmography signals in classifying multiclass arrhythmias has rarely been reported. Our study investigated the feasibility of using photoplethysmography signals and a deep convolutional neural network to classify multiclass arrhythmia types. Methods and Results ECG and photoplethysmography signals were collected simultaneously from a group of patients who underwent radiofrequency ablation for arrhythmias. A deep convolutional neural network was developed to classify multiple rhythms based on 10-second photoplethysmography waveforms. Classification performance was evaluated by calculating the area under the microaverage receiver operating characteristic curve, overall accuracy, sensitivity, specificity, and positive and negative predictive values against annotations on the rhythm of arrhythmias provided by 2 cardiologists consulting the ECG results. A total of 228 patients were included; 118 217 pairs of 10-second photoplethysmography and ECG waveforms were used. When validated against an independent test data set (23 384 photoplethysmography waveforms from 45 patients), the DCNN achieved an overall accuracy of 85.0% for 6 rhythm types (sinus rhythm, premature ventricular contraction, premature atrial contraction, ventricular tachycardia, supraventricular tachycardia, and atrial fibrillation); the microaverage area under the microaverage receiver operating characteristic curve was 0.978; the average sensitivity, specificity, and positive and negative predictive values were 75.8%, 96.9%, 75.2%, and 97.0%, respectively. Conclusions This study demonstrated the feasibility of classifying multiclass arrhythmias from photoplethysmography signals using deep learning techniques. The approach is attractive for population-based screening and may hold promise for the long-term surveillance and management of arrhythmia. Registration URL: www.chictr.org.cn. Identifier: ChiCTR2000031170.

Keywords: arrhythmias; deep convolutional neural networks; photoplethysmography.

Figure 1. Overall structure of proposed method for detecting multiple arrhythmia types including signal acquisition, signal preprocessing, and algorithm development.

AF indicates atrial fibrillation; DCNN, deep convolutional neural network; PAC, premature atrial contraction; PPG, photoplethysmography; PVC, premature ventricular contraction; SR, sinus rhythm; SVT, supraventricular tachycardia; and VT, ventricular tachycardia.

Figure 2. Architecture, development, and evaluation of the DCNN for classification of multiple arrhythmia types.

A, The DCNN model has 13 one‐dimensional convolutional layers, 5 one‐dimensional max‐pooling layers, and 2 fully connected layers. B, Workflow illustrating the data sets used to train, tune, and evaluate the DCNN model. The symbol expressed as Conv_k represents a 1‐dimensional convolution layer with k number of filters; for example, Conv_32 denotes a 1‐dimensional convolution layer with 32 filters. DCNN indicates deep convolutional neural network; MaxPool, max‐pooling; and PPG, photoplethysmography.

Figure 3. Distribution of rhythm types across data sets.

AF indicates atrial fibrillation; PAC, premature atrial contraction; PVC, premature ventricular contraction; SR, sinus rhythm; SVT, supraventricular tachycardia; and VT, ventricular tachycardia.

Figure 4. Results of the confusion matrix and ROC curves.

A, Confusion matrix with numbers and relative percentages to evaluate the performance of the DCNN for 6‐rhythm discrimination. B, Microaverage ROC curves of the DCNN and ML‐based detectors for 6‐rhythm discrimination. In (A), rows represent the categories given by the reference standard, and columns represent the categories predicted by the DCNN. Percentages were calculated by normalizing the results horizontally. AF indicates atrial fibrillation; ANN, artificial neural network; AUC, area under the ROC curve; DCNN, deep convolutional neural network; KNN, k‐nearest neighbors; PAC, premature atrial contraction; PVC, premature ventricular contraction; RF, random forest; ROC, receiver operating characteristic; SR, sinus rhythm; SVM, support vector machine; SVT, supraventricular tachycardia; and VT, ventricular tachycardia.

Figure 5. t‐SNE visualizations of learned features from representative layers in the DCNN: (A) second, (B) fourth, (C) seventh, and (D) 13th convolutional layers

AF indicates atrial fibrillation; DCNN, deep convolutional neural network; PAC, premature atrial contraction; PVC, premature ventricular contraction; SR, sinus rhythm; SVT, supraventricular tachycardia; t‐SNE, t‐distributed stochastic neighbor embedding; and VT, ventricular tachycardia.

Figure 6. Examples of PPG waveforms with a Guided Grad‐CAM visualization showing crucial regions for the DCNN to predict a certain triage category: (A) sinus rhythm, (B) premature ventricular contraction, (C) premature atrial contraction, (D) ventricular tachycardia, (E) supraventricular tachycardia, and (F) atrial fibrillation. In each panel, the I‐lead ECG waveform corresponding to the PPG is also shown.

DCNN indicates deep convolutional neural network; Guided Grad‐CAM: guided gradient‐weighted class activation mapping; and PPG, photoplethysmography.