Deep Learning Approaches to Detect Atrial Fibrillation Using Photoplethysmographic Signals: Algorithms Development Study

Abstract

Background: Wearable devices have evolved as screening tools for atrial fibrillation (AF). A photoplethysmographic (PPG) AF detection algorithm was developed and applied to a convenient smartphone-based device with good accuracy. However, patients with paroxysmal AF frequently exhibit premature atrial complexes (PACs), which result in poor unmanned AF detection, mainly because of rule-based or handcrafted machine learning techniques that are limited in terms of diagnostic accuracy and reliability.

Objective: This study aimed to develop deep learning (DL) classifiers using PPG data to detect AF from the sinus rhythm (SR) in the presence of PACs after successful cardioversion.

Methods: We examined 75 patients with AF who underwent successful elective direct-current cardioversion (DCC). Electrocardiogram and pulse oximetry data over a 15-min period were obtained before and after DCC and labeled as AF or SR. A 1-dimensional convolutional neural network (1D-CNN) and recurrent neural network (RNN) were chosen as the 2 DL architectures. The PAC indicator estimated the burden of PACs on the PPG dataset. We defined a metric called the confidence level (CL) of AF or SR diagnosis and compared the CLs of true and false diagnoses. We also compared the diagnostic performance of 1D-CNN and RNN with previously developed AF detectors (support vector machine with root-mean-square of successive difference of RR intervals and Shannon entropy, autocorrelation, and ensemble by combining 2 previous methods) using 10 5-fold cross-validation processes.

Results: Among the 14,298 training samples containing PPG data, 7157 samples were obtained during the post-DCC period. The PAC indicator estimated 29.79% (2132/7157) of post-DCC samples had PACs. The diagnostic accuracy of AF versus SR was 99.32% (70,925/71,410) versus 95.85% (68,602/71,570) in 1D-CNN and 98.27% (70,176/71,410) versus 96.04% (68,736/71,570) in RNN methods. The area under receiver operating characteristic curves of the 2 DL classifiers was 0.998 (95% CI 0.995-1.000) for 1D-CNN and 0.996 (95% CI 0.993-0.998) for RNN, which were significantly higher than other AF detectors (P<.001). If we assumed that the dataset could emulate a sufficient number of patients in training, both DL classifiers improved their diagnostic performances even further especially for the samples with a high burden of PACs. The average CLs for true versus false classification were 98.56% versus 78.75% for 1D-CNN and 98.37% versus 82.57% for RNN (P<.001 for all cases).

Conclusions: New DL classifiers could detect AF using PPG monitoring signals with high diagnostic accuracy even with frequent PACs and could outperform previously developed AF detectors. Although diagnostic performance decreased as the burden of PACs increased, performance improved when samples from more patients were trained. Moreover, the reliability of the diagnosis could be indicated by the CL. Wearable devices sensing PPG signals with DL classifiers should be validated as tools to screen for AF.

Keywords: atrial fibrillation; deep learning; diagnosis; photoplethysmography; pulse oximetry.

Figure 1

Flowchart illustrating the deep learning process. For each subject, 15-min PPG data during pre- and post- direct-current cardioversion periods were obtained. Each 15-min sample was preprocessed by removing bias, applying bypass filters, and normalization. Then, each sample was subdivided into 30-second samples with 20-second overlaps for data augmentation. The 30-second samples were trained and tested by 1-dimensional convolutional neural network (1D-CNN) and recurrent neural network (RNN) methods. Each sample was labeled as atrial fibrillation (AF) or sinus rhythm (SR). The number in parenthesis shows an example of the corresponding confidence level. In this example, the confidence level for diagnosing AF was 0.9 in 1D-CNN and RNN models, whereas 0.1 in SR for both models. 1D-CNN: 1-dimensional convolutional neural network; AF: atrial fibrillation; PPG: photoplethysmography; RNN: recurrent neural network; SR: sinus rhythm.

Figure 2

Typical examples of 15-second-long photoplethysmography and corresponding synchronized electrocardiogram samples for (A) stereotypic normal sinus rhythm and (B) atrial fibrillation with suggested confidence level using the 1-dimensional convolutional neural network algorithm. AF: atrial fibrillation; CL: confidence level; ECG: electrocardiogram; PPG: photoplethysmography; SR: sinus rhythm.

Figure 3

The receiver operating characteristic (ROC) curves of 2 deep learning classifiers (1-dimensional convolutional neural network, 1D-CNN and recurrent neural network, RNN) compared with other previous high-end atrial fibrillation (AF) detectors. (A) A Comparison of several ROC curves by different AF-detection algorithms. (B) The area under the curve and corresponding 95% CI by different algorithms. Both 1D-CNN and RNN methods showed significantly better diagnostic performance than previous detectors. 1D-CNN: 1-dimensional convolutional neural network; RMSSD: root-mean square of the successive differences of RR intervals; RNN: recurrent neural network; ShE: Shannon entropy; SVM: support vector machine.

Figure 4

Comparison of performances of deep learning classifiers and previous state-of-the-art atrial fibrillation detectors by premature atrial complexes (PACs) burden. The performance of classifying photoplethysmography samples during post- direct-current cardioversion period as sinus rhythm by each algorithm was measured by specificity. (A) Scenario A was obtained by the 5-fold cross-validation with random assignment of patients. In this case, each algorithm faced new patient’s data during testing. (B) Scenario B was obtained by the 5-fold cross-validation with random assignment of samples. This approach assumed that the training distribution could emulate the test distribution. Regardless of the method, both 1-dimensional convolutional neural network and recurrent neural network maintained higher specificity over burden of PACs. Both DL classifiers showed higher specificity in Scenario B than Scenario A. 1D-CNN: 1-dimensional convolutional neural network; PAC: premature atrial complex; PPG: photoplethysmography; RNN: recurrent neural network root-mean square of successive difference of RR intervals.

Figure 5

The characteristics of confidence level (CL) calculated by deep learning (DL) classifiers. Data were obtained by repeating the 5-fold cross-validation test over 10 times. (A) Comparison of true and false CLs of 1-dimensional convolutional neural network (1D-CNN) and recurrent neural network (RNN) methods by Box-and-Whiskers plot. True CLs indicate the cases where the diagnosis of a DL classifier was correct. Conversely, false CLs indicate cases where a DL classifier was incorrect. In the both 1D-CNN and RNN methods, the distributions of true or false CLs were significantly different (P<.001) for both 1D-CNN and RNN methods. If cut-off level of CL is to be 95% (dashed line), the diagnostic accuracy was 99.6% in 1D-CNN and 99.2% in RNN. Therefore, a diagnosis with a CL ≥95% can be regarded as certain. (B) The association between the probability of misdiagnosis and sample proportions and the respective CLs. Because 91% of the tested samples showed CL ≥95%, most diagnoses made by DL classifiers were valid. The probability of false diagnoses decreases from 50% to 0% as the CL increases from 50% to 100%. Comparing 1D-CNN and RNN, there was no significant difference in CLs (P=.98). *P<.001, calculated by Student t test. **P=.98, calculated by Student t test. 1D-CNN: 1-dimensional convolutional neural network; RNN: recurrent neural network.