Title:  Automatic Multiple Fiducial Point Delineation for the Non-contact Seismocadiogram Signals Using Deep Learning Technology

Committee: 

Dr. Ying Zhang, Advisor

Dr. Durgin, Chair

Dr. Inan

Abstract: The objective of the proposed research is to develop a standalone automatic multiple fiducial-point delineation for the non-contact seismocardiogram (SCG) signals. SCG is the precordial vibration that contains temporal information about cardiac micro-events, and this vibration can be recorded in a non-contact fashion using a Doppler radar device. Non-contact SCG measurement alleviates patients’ annoyance and can advance the development of the wireless healthcare system. However, delineation work for non-contact SCG signals is more difficult since they are more vulnerable to interference, and any assistant contact signals are avoided to achieve the fully non-contact measurement. To address this challenge, we formulated the multiple fiducial point delineation as a sequence-to-sequence task and leveraged multiple deep learning technologies to build up a standalone delineation framework for non-contact SCG signals. First, a SCG-CRF network consisting of a feature extraction block, a time series analysis block, and a joint tagging block was constructed to learn the conversion of the SCG signals and their corresponding labels. The SCG-CRF network was validated using both the contact SCG signal from the combined measurement of electrocardiography, breathing, and seismocardiogram (CEBS) database and the radar acceleration waveforms. As a part of the proposed work, a generative data-augmentation network will be developed to enhance the generalization of the SCG-CRF network. In addition, a segment filter will also be built to identify recognizable segments of non-contact SCG signals, and a waveform transformation model will be investigated to convert the non-contact SCG signals to contact-like SCG signals for the subsequent delineation. The accuracy and generalizability of the overall delineation framework will be evaluated using non-contact SCG signals captured in various states.