ECG Quality Detection and Noise Classification for Wearable Cardiac Health Monitoring Devices

Electrocardiogram (ECG) signals are continuously acquired from the body surface and analyzed or transmitted with the help of wearable devices for continuous cardiac health monitoring under resting, ambulatory, and exercise conditions. Automatic checking of ECG signal quality has become most essential to reduce false alarms and improve trustworthiness of automatic ECG diagnosis. Furthermore, identifying the types of ECG noise sources can also improve noise removal effectiveness with selection of noise-specific denoising approach with reduced computational load. In this paper, we present convolutional neural network (CNN) based ECG quality detection and classification method by exploring optimal hyperparameters to achieve lightweight CNN model with acceptable performance in classifying noises into electrode movement artifacts, muscle artifacts, and random noises. The proposed CNN-based method achieves an accuracy of 94.87% in detecting quality of ECG signals and achieves a sensitivity of above $98 \%$ in identifying three types of noises with two convolutional layers and three dense layers with best activation function of exponential linear unit. The proposed method has a model size of 1409 kB and computational time of 71.62 ms for processing 5 s ECG signal. The proposed CNN-based ECG quality checking and noise type identification has great potential in automated cardiovascular disease diagnosis with reduced false alarms by discarding severely corrupted signals and enabling noise-specific ECG denoising in achieving better noise reduction capabilities by selecting noise specific signal processing techniques.