Generating Synthetic Mechanocardiograms for Machine Learning Based Peak Detection

Abstract

Acquiring labeled data for machine learning algorithms in healthcare is expensive due to the laborious expert annotation and privacy concerns. This challenge is further complicated in the case of Mechanocardiogram (MCG) data, which are characterized by high inter- and intrapersonal complexity, compounded further by sensor variability. In this paper, we introduce an innovative method for generating synthetic mechanocardiogram (MCG) signals to address the scarcity of labeled data necessary for training machine learning models in healthcare. Our approach involves generating RR-intervals, adding wavelets, and incorporating noise to create realistic synthetic MCG signals. These synthetic signals were used to train a convolutional neural network (CNN) for peak detection in real MCG data. Our key contributions include developing a detailed methodology for realistic synthetic MCG signal generation, reducing the mean absolute error (MAE) in peak detection by 4.88 beats per minute (BPM) using synthetic data, enhancing the training of machine learning models, creating a new peak detection method, and addressing data scarcity in biomedical signal processing. These contributions emphasize the methodological innovations and the significance of our results, underscoring the potential impact of synthetic data in improving healthcare diagnostics.