Detecting Atrial Fibrillation in Real Time Based on PPG via Two CNNs for Quality Assessment and Detection

Abstract

Real-time detection of atrial fibrillation (AFib) is made possible by the quality assessment via a 1-D convolutional neural network (1D-CNN) in a processor of a photoplethysmography (PPG) sensor patch and a 2-D convolutional neural network (2D-CNN) for AFib detection in cloud. The 1D-CNN is able to filter out the unqualified PPG that is contaminated by motion artifacts and/or ambient light interference. The remaining qualified PPG is then inputted to another built 2D-CNN for detecting AFib. This 1D-CNN consists of four layers of convolutions and max pooling, one long short-term memory (LSTM), and an output dense layer. The 2D-CNN is pretrained based on the electrocardiography (ECG) data from multiparameter intelligent monitoring in intensive care (MIMIC) III database, for which the RR-intervals (RRIs) of ECG data are first extracted in Poincaré images and then regarded as input features to the model for training. This 2D-CNN has also four layers of convolutions and max pooling and four output dense layers. The pretrained model is next fine-tuned based on peak-to-peak intervals (PPIs) of PPG measured from wearable devices as input features for detecting AFib effectively. The quality-assessment 1D-CNN model is implemented in the wearable device to transmit only qualified data to the 2D-CNN model in cloud for AFib detection, achieving power efficiency. Both models are trained by the Adam optimizer. To validate the models, the PPIs of PPG were collected to evaluate the performance of the established models in real time. Experimental results show that the fine-tuned 2D-CNN for AFib detection achieves the accuracy, sensitivity, and specificity were 98.08%, 96.82%, and 98.86%, respectively, the most favorable as opposed to other reported works based on PPG. The models are able to not only assist clinicians in AFib detection but also provide a mechanism to detect AFib via wearable devices in real time.