Fast CNN Based Electrocardiogram Signal Quality Assessment Using Fourier Magnitude Spectrum for Resource-Constrained ECG Diagnosis Devices

Automatic assessment of electrocardiogram (ECG) signal quality plays a vital role in reducing false alarms and improving trustworthiness of unobtrusive health monitoring devices under noisy ECG recordings, which are unavoidable in continuous health monitoring. In this paper, we present an ECG signal quality assessment (ECG-SQA) method based on the Fourier magnitude spectrum (FMS) as an input to the one-dimensional convolutional neural network (1D CNN) with optimal hyper-parameters and activation function, which significantly reduces CNN model size and computational load of resource-constrained devices. On the untrained ECG databases including single-lead and multi-lead ECG signals having different kinds of PQRST morphologies and various kinds of noise sources, the optimal 1D CNN based ECG-SQA method had a sensitivity of 99.30%, and specificity of 95.40% for the three convolution layers (CLs), three dense layers (DLs), and kernel size of $3\times 1$ . This study demonstrated that optimal parameter selection can reduce computational resources of 52% with the CNN model size of 852 kB and 67697 parameters as compared with other CNN models. Real-time implementation on Raspberry Pi computing shows that the processing time is 124.4±42.5 ms for checking the quality of 5 s ECG signal.