Standalone Systolic Profile Detection of Non-Contact SCG Signal With LSTM Network

This research aims to propose a standalone systolic profile (SP) detection method that could locate systolic profiles without referring to concurrent R-peaks of electrocardiogram (ECG) signal. This technique is specifically used to facilitate the study of seismocardiogram (SCG) obtained by noncontact sensing methods, such as radar acceleration waveform (RAW). In our research, a single-layer bi-directional long short-term memory (LSTM) network with 300 hidden units is used to distinguish SP, which is the part of SCG that depicts the phase of systole, including fiducial points such as isovolumic contraction (IM) and aortic valve opening (AO) points. The network is trained on a semi-automatically constructed dataset that consists of 2115 positive segments (PSs) and 2115 negative segments (NSs) with the length of 256 milliseconds. The trained network is applied to RAW signal in the form of a sliding window, judging whether each segment is SP. The performance of the proposed approach is evaluated using the prediction accuracy of SP and average inter-beat intervals (IBIs) on 6 subjects. The results show that this approach is promising in replacing the ECG R-peaks as the reference signal for clean or partially noisy RAWs. In addition, the proposed approach demonstrates better performance than the existing heart rate envelope-based methods.