Non-contact Blood Pressure Estimation Using a 300 GHz Continuous Wave Radar and Machine Learning Models

This work shows a novel system to measure the blood pressure (BP) values of subjects without body contact. For this purpose, a continuous wave (CW) radar consisting of a vector network analyzer (VNA), horn antennas, and frequency converters is operated at 300 GHz. By using discrete wavelet transformation and suitable signal processing, characteristics of heart sounds and certain features in the time and frequency domain are extracted from the radar signal. During that process, the heart rate of the subjects was also measured with a mean relative error (MRE) of 4.57 %. A data set of eight subjects is built up and combined with an existing database, thus creating enough instances to use machine learning (ML) models for blood pressure estimation. The models are trained, optimized and cross-validated with different subsets of the features. The ones with the best performance, support vector machine (SVM) and bagging, are also tested with the data of individual subjects, unknown to the model, which was trained with the remaining instances. Using the features in the frequency domain the best results were obtained with an MRE of 8.3 % for the diastolic BP (DBP) and 8.04 % for the systolic BP (SBP). These results suggest that this technique is of potential use for blood pressure monitoring without body contact and offer exciting possibilities for future work.