ID 2360: Improving the robustness of radar-based heart sound detection
Monitoring vital signs is crucial for assessing the physiological status of individuals. The heart rate is a vital sign typically measured continuously using wearables or electrocardiogram (ECG) recording. Contactless vital sign monitoring is desirable in various healthcare settings to avoid the inconvenience and discomfort of traditional contact-based sensing [1]. Continuous wave Doppler radar (CWDR) is a promising approach for non-contact measurement of heartbeats as it measures the relative displacement of the human body caused by the contraction of the heart muscle [2, 3], and radiofrequency waves can penetrate clothing and bedding.
As Will et al. and Shi et al. [3, 4] show, heartbeat information can be derived from the received radar echo by detecting two characteristic heart sounds occurring during a single heartbeat. The mechanical cause of these sounds is, with respect to the underlying heart rate, high-frequent with spectral components in the recorded radar echo in the band of 16-80 Hz [5]. However, existing processing pipelines for detecting heart sounds must be robust against random large body movements (RLBMs) which are an integral part of human behavior. A recent study by Herzer et al. [6] has shown that controlled large-body movements significantly impact the accuracy of these approaches due to the noise they induce in the 16-80 Hz frequency band.
To address this limitation, this thesis aims to build up on the method developed by Shi et al. [4] which has proven itself in stationary settings and improved its performance on noisy data. The goal is to derive a heart sound detection algorithm that is applicable in real-life scenarios by accounting for RLBMs more precisely. In addition, since currently published datasets containing radar and reference signals, such as [7], do not contain RLBMs, this thesis aims to collect and curate a more realistic dataset for testing novel radar-based heart sound detection pipelines. Combining multiple datasets from sleep and stress research will be used in this thesis to enhance the generalizability & robustness of existing models. The state-of-the-art processing pipelines [3, 4] will be improved stepwise using different approaches, such as retraining the pipeline on the new dataset containing RLBMs, choosing different pre-processing approaches, modifying the network architecture, combining the information of multiple radar nodes, etc.
Supervisors
References
[1] Wenjin Wang, Xuju Wang, “Contactless Vital Signs Monitoring”, 2021, pp. 1-24, https://doi.org/10.1016/C2019-0-04343-9.
[2] Hu, Z. Zhao, Y. Wang, H. Zhang and F. Lin, “Noncontact Accurate Measurement of Cardiopulmonary Activity Using a Compact Quadrature Doppler Radar Sensor,” in IEEE Transactions on Biomedical Engineering, vol. 61, no. 3, pp. 725-735, March 2014, doi: 10.1109/TBME.2013.2288319.
[3] Will, C., Shi, K., Schellenberger, S. et al. Radar-Based Heart Sound Detection. Sci Rep 8, 11551 (2018). https://doi.org/10.1038/s41598-018-29984-5.
[4] Shi et al., “Segmentation of Radar-Recorded Heart Sound Signals Using Bidirectional LSTM Networks,” 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Berlin, Germany, 2019, pp. 6677-6680, doi: 10.1109/EMBC.2019.8857863.
[5] Holldack, K. & Wolf, D. Atlas und kurzgefaßtes Lehrbuch der Phonokardiographie und verwandter Untersuchungsmethoden. (Thieme, 1966).
[6] Herzer et al., “Influence of Sensor Position and Body Movements on Radar-Based Heart Rate Monitoring,” 2022 IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI), Ioannina, Greece, 2022, pp. 1-4, doi: 10.1109/BHI56158.2022.9926775.
[7] Shi, K., Schellenberger, S., Will, C. et al. A dataset of radar-recorded heart sounds and vital signs including synchronised reference sensor signals. Sci Data 7, 50 (2020). https://doi.org/10.1038/s41597-020-0390-1.