A CNN - ELM-Based Method for Ballistocardiogram Classification in a Clinical Environment

Abstract

The ballistocardiogram (BCG) represents rich signals that have been adopted in various clinical applications. Still, vital signs detection via BCG is a troublesome task because the BCG waveform morphology depends on the measurement device. Additionally, BCG can be different between and within-subjects, hence in this paper, we applied a deep learning-based approach, namely convolutional neural network (CNN) and extreme learning machine (ELM), to discriminate between BCG and non-BCG signals. BCG signals were acquired with an IoT -based microbend fiber optic sensor mat from ten patients diagnosed with obstructive sleep apnea and underwent drug-induced sleep endoscopy. Three methods, including undersampling, oversampling, and generative adversarial networks (GANs), were used to balance the number of BCG vs. non-BCG signals. Furthermore, the system performance was assessed using 10-fold cross-validation. Overall, the best results were achieved using the CNN-ELM with GANs as a data balancing method. The average accuracy, precision, recall F -score were 94%, 90%, 98%, and 94%, respectively.