1D Convolutional Neural Network Impact on Heart Rate Metrics for ECG and BCG Signals

The presence of motion artifacts (MA) in cardiac signals negatively impacts the reliability of higher-level information such as the Heart Rate (HR), and therefore the correct diagnosis of pathologies. This paper proposes an MA detection method, based on One-Dimensional Convolutional Neural Networks (1D CNN), to label noisy zones of signals as unreliable, and subsequently avoid them for metric calculations.

To validate the concept, we first design a CNN to detect MAs in electrocardiogram (ECG) recordings from MIT–BIH Arrhythmia and Noise Stress Test Databases. This network extracts features from 1 s data segments, and then classifies them as clean or noisy. Also, we then train a tuned version of the model with semi-synthetic ballistocardiogram (BCG) signals.

The classification in ECG achieves an accuracy of 95.9% and the BCG classification obtains an accuracy of 91.1%. Both classifiers are incorporated into beat detection systems, which produce an increase in the sensitivity of the detection algorithms from 75 to 98.5% in the ECG case, and from 72.1 to 94.5% in the case of BCG, for signals contaminated at 0 dB of SNR.

We propose that this method will improve accuracy of any processing algorithm on BCG signals by identifying useful segments where a high accuracy can be achieved.