SCG-CRF Network: A Standalone Sequence Labeling Framework for Seismocardiogram Signals Using Deep Learning Approach

The current seismocardiogram (SCG) signal annotations for multiple fiducial points (FPs) cannot achieve high precision or rely on assistant signals, which complicates experiments and hinders wireless measurement progress. In this article, we aim to improve the precision of predicting multiple SCG FPs using deep-learning techniques without the need for assistive signals. A standalone two-step framework is proposed to improve the detection accuracy by first identifying the fiducial regions (FRs), i.e., the vicinities around FPs, of the SCG and then detecting the FPs within the corresponding FRs. The SCG conditional-random-field (SCG-CRF) network is developed to capture spatial and temporal features of the SCG and produce sequential labels to indicate each FR, and the extremums inside are selected as the candidates. The leave-one-subject-out cross-validation (LOSO-CV) test is conducted using SCG signals from 27 healthy subjects to evaluate the prediction performance of the proposed work. The predicted candidates within ±1-ms error of the ground truth are counted as the correct predictions, and the average precisions, recalls, and$F1$-scores over all FPs of all subjects achieve 97.75%, 96%, and 96.87%, respectively. The proposed regional labeling approach can alleviate the imbalanced classification issue and increase the prediction accuracy compared with the methods that search FPs directly. The proposed framework can identify multiple FPs precisely without any assistant signals, which makes it a powerful annotation tool that can be extended to the labeling work of other similar time-series physiological signals.