AFibri-Net: A Lightweight Convolution Neural Network Based Atrial Fibrillation Detector

By considering limited resource-constraints of medical devices and advanced deep learning networks, in this paper, we explore a lightweight convolutional neural network (CNN) based AFibri event detector by finding suitable hyperparameters and activation function with best trade-off between the detection accuracy and model size (or computational time). This study presents extensive evaluation results of CNN-AFibri event detection methods that are obtained for different combination of model parameters: number of convolutional layers (CLs of 3, 4, and 5), number of filters (8, 16, 32, 64 and 128), activation functions (including the rectified linear unit (ReLU), leakyReLU (LReLU), exponential linear unit (ELU)), and kernel sizes ( 3 \times 1~ , 4 \times 1 ). In addition to different CNN-AFibri models, we validate their performances under different ECG segment duration of 5, 10 and 30 seconds. On the standard databases and unseen databases, the CNN-AFibri model with the CLs of 5, ELU function and kernel size of 4 \times 1 had a highest accuracy of 99.97% (specificity of 99.98% and sensitivity of 99.95%) for 5 second ECG segments as compared to the performances of 54 CNN-AFibri models reported in this paper and other existing deep learning based methods on the same validation databases. Real-time implementation of the best CNN based method with a model size of 3.14 Megabyte is demonstrated using the Raspberry Pi computing platform with Broadcom BCM2711, 1.5 GHz Cortex-A72 quad-core CPU with 8 GB RAM. Results demonstrated that the average processing times are less than 3 ms and 11 ms for processing 5 s and 30 s ECG segments, respectively with an accuracy reduction of less than 1% as compared to the same model tested on the personal computer with Intel(R) Xeon(R) W-2133 3.60 GHz Processor with 6 core and 128 GB RAM.