Deep Spectrogram Feature Based ECG Signal Quality Assessment for False Alarm Reduction

Abstract

Electrocardiogram (ECG) based continuous monitoring of vital signs and disease detection devices are susceptible to artifacts while recording the signals. Noisy ECG signal analysis for vital sign monitoring generates false alarms. There is a requirement for quality-aware health monitoring devices to reduce false alarms. The ECG signal quality assessment method contributes to efficient power utilization in resource-constrained wearable devices. This paper proposes an automatic, lightweight, and energy-efficient ECG signal quality assessment method based on spectrogram and convolutional neural network (CNN). For the optimal method, 24 CNNs are trained and tested. The CNNs are with 2, 3, and 4 numbers of convolutional layers

six different kernel sizes

and dense layers of 3 and 4. The optimal CNN for the quality assessment ECG signal has three convolutional layers, three dense layers, and a kernel size of 6×6. For a database unseen to the optimal CNN, the method has a sensitivity of 95.01% and a specificity of 95%. The optimal CNN has a testing time or latency of 18.49 ms for quality assessment of a 5 s ECG signal. The model size of the spectrogram and CNN-based optimal ECG signal quality assessment method is 723 kB. © 2023 IEEE.