Simon Meske

Advisors

Robert Richer (M. Sc.), Thomas Altstidl (M. Sc.), Daniel Krauß (M. Sc.), Nils Albrecht (M.Sc.) (Technical University Hamburg, Germany), Prof. Dr.-Ing. habil. Alexander Kölpin (Technical University Hamburg, Germany), Prof. Dr. Björn Eskofier

Duration

02/2024 – 08/2024

Abstract

The prevention of illnesses is an essential factor in prolonging life expectancy and ensuring an overall healthier life next to the treatment of such. In recent years, this area of health care has gotten increasingly more attention. One prominent example is the German Health Ministry’s plan to create a new institute responsible for finding and exploring new ways to prevent illnesses [1]. The leading causes of death globally are cardiovascular diseases (CVD), with a death rate of over 51% in the year 2019 [2]. Thus, monitoring vital signs, such as heart rate, is a helpful way to detect CVD. The current gold standard for measuring heart rate in hospitals is the electrocardiogram (ECG). This technique requires direct contact with the patient and trained personnel to apply the electrodes correctly, limiting its feasibility for ubiquitous and continuous measurements. Because of these drawbacks, researchers focused on methods that are based on contactless measurement principles, such as radar. Here, the heart rate can be estimated by measuring and evaluating the changes in the reflection of electromagnetic waves caused by body movements or cardiac vibrations [3]. This method improves usability in a clinical scenario since no direct contact with the patient is required and, consequently, opens possibilities for scenarios outside this clinical setting since there is no need to apply any electrodes by trained professionals. The heart rate can be monitored at home or in the car, resulting in considerable coverage throughout the day and increasing the possibility of detecting CVD early on [4].

However, state-of-the-art radar-based heart rate detection suffers from some considerable limitations. Large random body movements (RLBM), which occur naturally when the examined individual moves, are a significant issue in detecting heart rates correctly. These RLBMs overshadow the small signal amplitudes that occur due to heart sounds. Previous studies in this field circumvent this issue by utilizing very rigid settings for the participants to prevent movement. The lower amount of movement also reduces the heart rate variability (HRV), making the datasets used more homogenous and the heart rates easier to predict. This becomes especially visible when models trained on data solely trained on lying participants are compared to models trained on sitting participants [3, 4, 5]. Nevertheless, to make radar-based heart rate measurements a viable alternative to ECG, they must be more robust against RLBMs and deliver good results for a more extensive range of heart rates.

Therefore, this thesis intends to overcome this limitation by explicitly focusing on data containing artifacts from RLBM. The data for training the machine learning models is from recordings in different settings. These settings include data from individuals in bed with low movement and heart rate variability, sitting participants during mood induction with slightly more movement and a more extensive heart rate range, and data during acute stress induction with the highest amount of RLBMs and heart rate ranges. Combining these different datasets has the potential to improve the robustness of radar-based heart sound detection.  In comparison to previous work [6], this thesis will combine a deep learning-based approach with a wavelet transformation-based preprocessing pipeline since it has been shown that frequency- or wavelet-based preprocessing approaches are more promising for radar-based heart sound estimation [7, 8, 9].

After training the proposed pipeline, its performance will be compared to that of an existing pipeline, which uses the filtered signals as input to an LSTM model to evaluate whether the model is more robust to movement.

 

Full Thesis

 

References

[1] Ministry of Health Germany. (2023, October 04). https://www.bundesgesundheitsministerium.de/presse/pressemitteilungen/praeventions-institut-im-aufbau-pm-04-10-23
[2] Gaidai, O., Cao, Y., & Loginov, S. (2023). Global cardiovascular diseases death rate prediction.Current Problems in Cardiology, 101622.
[3] Will, C., Shi, K., Schellenberger, S., Steigleder, T., Michler, F., Fuchs, J., … & Koelpin, A. (2018). Radar-based heart sound detection.Scientific reports, 8(1), 11551.
[4] Lee, K. J., Park, C., & Lee, B. (2016, August). Tracking driver’s heart rate by continuous-wave Doppler radar. In2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) (pp. 5417-5420). IEEE.
[5] Iwata, Y., Thanh, H. T., Sun, G., & Ishibashi, K. (2021). High accuracy heartbeat detection from CW-Doppler radar using singular value decomposition and matched filter.Sensors, 21(11), 3588.
[6] Albrecht, N. C., Langer., D., Krauss D., Richer, R., Abel, L., Eskofier B. M., Rohleder, N., Koelpin, A. (2024) EmRad: Ubiquitous Vital Sign Sensing using Compact Continous-Wave Radars. (submitted for publication).
[7] Li, M., & Lin, J. (2017). Wavelet-transform-based data-length-variation technique for fast heart rate detection using 5.8-GHz CW Doppler radar.IEEE Transactions on Microwave Theory and Techniques, 66(1), 568-576.
[8] Yamamoto, K., & Ohtsuki, T. (2020). Non-contact heartbeat detection by heartbeat signal reconstruction based on spectrogram analysis with convolutional LSTM.IEEE Access, 8, 123603-123613.
[9] Yamamoto, K., Hiromatsu, R., & Ohtsuki, T. (2020). ECG signal reconstruction via Doppler sensor by hybrid deep learning model with CNN and LSTM.IEEE Access, 8, 130551-130560
[10] Küderle, A., Richer, R., Sîmpetru, R. C., & Eskofier B. M. (2023). tpcp: Tiny Pipelines for Complex Problems-A set of framework independent helpers for algorithms development and evaluation. Journal of Open Source Software, 8(82), 4953.