Ballistocardiography Signal Quality Assessment Using Machine Learning

Ballistocardiography (BCG) is a non-invasive and versatile technique for detecting cardiac activity. Its flexibility makes it suitable for use in everyday and clinical settings, and it has also demonstrated promise for application in magnetic resonance imaging (MRI) gating. The paper focuses on the automatic identification of the BCG signal quality, which aims to increase the heartbeat detection accuracy, crucial for precise MR triggering. We analyzed 21 signal features based on their statistical properties and evaluated 14 machine learning models to classify BCG segments as high- or low-quality. Results showed the best performance of 4 ensemble classifiers, with the the Extra Trees Classifier achieving the highest overall accuracy of 86.65%, sensitivity of 80.54%, and precision of 73.62%. We also examined feature importance scores, finding that time-domain features generally outperformed those in the frequency domain. These insights can help guide future research and improve the robustness of BCG-based applications.Clinical relevance-This BCG signal-processing method was developed specifically for multi-channel sensory systems to automate channel selection rather than relying on visual assessment. By incorporating this technique into cardiac MRI examinations, we could improve heartbeat detection accuracy, reduce staff workload, and streamline acquisition of MRI data, which is often time-consuming and challenging in terms of image quality lacking motion or other artifacts due to insufficient gating.