A Conv -Transformer network for heart rate estimation using ballistocardiographic signals. (February 2023)

• Record Type: Journal Article
• Authors: Zhang, Miao; Qiu, Lishen; Chen, Yuhang; Yang, Shuchen; Zhang, Zhiming; Wang, Lirong
• Is Part Of: Biomedical signal processing and control. Volume 80(2023)Part 1
• Publication Date: 2023-02
• Subjects: Ballistocardiogram (BCG) -- Heart Rate (HR) -- Transformer -- HRCTP-net; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104302
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A deep learning method of a Conv-Transformer network with Pyramid input (HR CTP -net) is proposed to directly evaluate the HR from BCG signals. Transformer is used for the first time to estimate the heart rate of BCG. For the first time, the MIT-BIH noise stress test database is used in BCG signal analysis and proves its effectiveness. Abstract: Objective: Continuous heart rate (HR) monitoring has great implications for the prevention of chronic diseases, and we use non-contact ballistocardiography(BCG) technology to estimate HR. Methods: In this paper, overnight BCG data are acquired from 78 patients using a 10-channel piezoelectric sensor matrix with a sampling rate of 50 Hz. 200 sets of non-overlapping 10 s quiet period data are selected for follow-up work, with a total of 15, 600 segments. A Conv-Transformer network with Pyramid input (HR CTP -net) is used to estimate HR values for these segments, where local features are obtained by CNN and global features are calculated by the transformer. This is an end-to-end network without additional post-processing. During the experiment, electrocardiogram (ECG) noise which from the MIT-BIH noise stress test database is also introduced for data augmentation to further improve the network generalization ability. Results: Taking the synchronously collected ECG as the ground truth, the results in the 6-fold cross-validation show that the proposed method achieves the best results on mean absolute error (MAE), standard deviation of absolute error (SDAE) and pearson correlation coefficient (PCC) with 0.93 bpm, 1.31 bpm and 0.97, respectively. Conclusion: To the best of our knowledge, this paper is the first time to introduce transformer and ECG noise into BCG signal analysis and demonstrate their effectiveness. Significance: Our proposed HR CTP -net has potential and promise in healthcare applications.