Introduction: Seismocardiography (SCG) is the measure of the local vibrations on the chest surface produced by the beating heart. Multiple fiducial points in the SCG have been correlated to events in the cardiac cycle such as valve openings, valve closures, and blood flow. Therefore, these fiducial points are of great interest in the research of non-invasive measures of cardiac mechanics. Automatic identification of these fiducial points could make them relevant for a better clinical application. The aim was to identify multiple fiducial points related to cardiac events automatically on a beat-to-beat basis.
Methods: SCG was measured from 198 subjects. 28 of these subjects had cardiovascular disease (CVD) including hypertrophic cardiomyopathy, dilated cardiomyopathy, aortic valve disease, ischemic heart disease, heart failure and left bundle branch block. A semi-automatic annotation process was used to annotate the fiducial points in 42,452 individual SCG beats within three weeks. The data was split into 80% training and 20% testing. A UNet model was developed and trained to predict the multiple fiducial points on a beat-to-beat basis. To evaluate the model, an acceptable error margin of fiducial point prediction was set to 10 ms. Thus, the number of true positives and false positives was obtained, which was used to obtain macro-averaged positive predictive value and sensitivity for each fiducial point for subjects with and without CVD.
Results: The median positive predictive value was between 0.917 and 1.00 for no CVD subjects, while it was between 0.812 and 0.977 for CVD subjects. The median sensitivity for no CVD subjects was between 0.843 and 0.918, while it was between 0.645 and 0.847 for CVD subjects.
Conclusion: A novel UNet-based algorithm for detection of multiple fiducial points in subjects with CVD and no CVD has been developed providing the potential for an efficient and extensive analysis of the SCG.