Noise-Aware Atrial Fibrillation Detection for Resource-Constrained Wearable Devices

Atrial fibrillation (AF) is characterized by RR intervals of unequal lengths, fibrillatory waves, and absent P-wave. The AF raises the risk of ischemic stroke. So, early diagnosis is essential for detection of AF. Due to the intermittent nature of AF, early diagnosis is achieved through continuous electrocardiogram (ECG) monitoring. This research work presents a lightweight, single-stage, and noise-aware AF detection method, developed using 1D-convolutional neural network (CNN), which is implemented on computing platform with limited resources in terms of memory space and battery capacity. With 5 datasets, the 5-layer CNN with optimal hyperparameters (kernel size: 4×1, number of kernels: 8, 16, 32, 64, and 128, loss function: sparse categorical cross entropy, and optimizer: adaptive moment estimation) demonstrated an accuracy, sensitivity, and specificity of 99.89%, 99.95%, and 99.81%, respectively with model size 3.15 MB and latency of 0.30 ms for ECG segment of 5 s duration. The CNN model is deployed on Raspberry Pi 4B computing platform and detects AF with an accuracy and sensitivity of 99.67% and 99.62%, respectively. Our results show feasibility for implementation of the method on wearable health monitoring devices for reduction in false alarm rates and increase in performance.