SiamAF: Learning Shared Information from ECG and PPG Signals for Robust Atrial Fibrillation Detection

TL;DR

This work proposes a new approach to predict AF from either PPG or ECG and outperforms baseline methods on three external test sets, and achieves comparable performance to traditional learning regimes while requiring much fewer training labels, providing a potential approach to reduce future reliance on manual labeling.

Abstract

Atrial fibrillation (AF) is the most common type of cardiac arrhythmia. It is associated with an increased risk of stroke, heart failure, and other cardiovascular complications, but can be clinically silent. Passive AF monitoring with wearables may help reduce adverse clinical outcomes related to AF. Detecting AF in noisy wearable data poses a significant challenge, leading to the emergence of various deep learning techniques. Previous deep learning models learn from a single modality, either electrocardiogram (ECG) or photoplethysmography (PPG) signals. However, deep learning models often struggle to learn generalizable features and rely on features that are more susceptible to corruption from noise, leading to sub-optimal performances in certain scenarios, especially with low-quality signals. Given the increasing availability of ECG and PPG signal pairs from wearables and bedside monitors, we propose a new approach, SiamAF, leveraging a novel Siamese network architecture and joint learning loss function to learn shared information from both ECG and PPG signals. At inference time, the proposed model is able to predict AF from either PPG or ECG and outperforms baseline methods on three external test sets. It learns medically relevant features as a result of our novel architecture design. The proposed model also achieves comparable performance to traditional learning regimes while requiring much fewer training labels, providing a potential approach to reduce future reliance on manual labeling.

Figures (6)

• Figure 1: Here we show the architecture of our proposed framework for learning from PPG and ECG and predicting (i.e., testing) on only either ECG or PPG. At training time, the model takes both ECG and PPG signals as inputs. In each training iteration, the two signal modalities take turns flowing through the network following each of the colored paths; i.e., the PPG signal flows through the red path and the ECG signal flows through the purple path, then the ECG signal flows through the red path and the PPG signal flows through the purple path. In the configuration shown in the figure, the PPG and ECG signal inputs will go through the encoder $f$ and the projector $g$. The learned features of the PPG signal will pass through the predictor $q$ to map to the latent space of ECG features. We optimize an agreement loss between the predicted latent feature vector of the PPG signal and the projected latent feature vector of the ECG signal; we also optimize a supervised cross-entropy loss for the output of the classifier $h$ (which takes PPG latent features as inputs). For a detailed description of the training process, please see Section \ref{['sec:method']} and Algorithm \ref{['alg:algo_1']}. After the training is complete, we save only the encoder and the classifier for future predictions.
• Figure 2: This figure shows the performance comparisons of all models on the test sets. Note the Stanford dataset does not contain ECG signals. In both plots, the horizontal axis is the AUPRC score, and the vertical axis is the AUROC score. The closer to the top right corner, the better. Here confidence intervals are generally smaller than the dot sizes, so we provide a zoom-in of one of the dots to demonstrate the 95% CI. For exact CI and statistical significance test results, please refer to Sec.\ref{['sec:perfvals']} and Sec.\ref{['sec:stattests']}.
• Figure 3: The four plots visualize latent features after each pooling layer of the ResNet-34 encoder used in our proposed framework. Figure \ref{['fig:simsiamlatent1']}, \ref{['fig:simsiamlatent2']}, \ref{['fig:simsiamlatent3']}, \ref{['fig:simsiamlatent4']}, \ref{['fig:simsiamlatent5']}, each represent one of the stages. In Figure \ref{['fig:simsiamlatent5']}, we also visualize the separation between AF and non-AF classes in the same latent space.
• Figure 4: This figure shows visualizations of latent features at the final average pooling layer of the encoders in the baseline models.
• Figure 5: This figure shows an example of the peak detection results on a pair of time-synchronized (simultaneous) 30s ECG and PPG signals and the corresponding encoded peak sequences. Here the red dots indicate the detected peaks.