Unobtrusive and Continuous Measurements of Dynamic Beat-to-beat Blood Pressure Using Multimodal Physiological Signals

Abstract：Wearable cuffless blood pressure (BP) technology is emerging as a critical tool for monitoring hypertension, the leading risk factor of most cardiovascular diseases. However, current cuffless BP methods are not precise enough for clinical use, because they mainly use single or dual modalities/features as input for prediction. To address this challenge, inspired by our previously developed mechanism-driven multimodal physiological model (MPM), we propose McBP-Net, a deep learning model in hybrid CNN- LSTM architecture using multimodal signals along with handcrafted features as input for continuous dynamic beat- to-beat BP estimation. Multimodal signals includes photo- plethysmographic, impedance-plethysmographic (IPG), electro-cardiographic and temperature signals. Validated on 23 individuals undergoing a cold pressor test to induce large BP variations, our approach achieves mean absolute errors of 4.19 mmHg for systolic and 2.98 mmHg for diastolic BP. By fusing four multimodal signals, the performance improves on average by 16.20% over three modalities, 37.37% over dual modalities, and 49.52% over single- modality approaches, with significant improvements attributed to the integration of temperature and IPG signals. Additionally, McBP-Net provides a balanced solution between accuracy and computational efficiency, outperforming other state-of-the-art data-driven models and our previous MPM. Notably, it maintains robust performance with only a 0.21 mmHg degradation in SBP estimation even when trained on rest-stage data. Validated across different BP levels, our approach shows strong generalizability and consistent accuracy for estimated BP in both normotensive and hypertensive ranges. This approach demonstrates promising potential in medical-grade wearable cuffless dynamic BP measurements.

Keywords：Dynamic blood pressure, multimodality, cuffless, convolutional neural network, long short-term memory