A Novel Multimodal Physiological-Informed Model (MPIM) for the Unobtrusive Estimation of Dynamic Beat-to-beat Arterial Blood Pressure

Abstract

According to the latest report from the World Health Organization (WHO), cardiovascular diseases (CVDs) have been the leading cause of death and disability globally for the past 20 years, accounting for approximately 32% of all deaths worldwide. The number of CVD deaths increases steadily, though technologies and medicines are developed so fast, over the past two decades and it is expected to rise to over 23 million by 2030. Among the risk factors for CVDs, hypertension is strongly evidenced as a leading cause, with a high prevalence affecting approximately 1.4 billion adults worldwide. To effectively combat these deadly uncontrolled diseases, the international community has proposed a paradigm shift from the current reactive medicine to the future proactive health, emphasizing early prediction, early diagnosis, and early intervention of diseases. Driven by the vision of proactive health enabled by mobile health (mHealth), wearable technology has emerged as a significant area of research and development globally.

The overall aim of this thesis is to develop methods suitable for wearable devices that can monitor the cardiovascular risk factors—specifically arterial blood pressure (BP)—frequently, accurately, and unobtrusively in their daily life for early detection, early prediction, early diagnosis and early intervention of the diseases. However, current approaches face significant challenges in achieving the accuracy required for medical-grade applications, limiting their reliability and widespread adoption in clinical and everyday settings. Many existing models relying on single or dual modalities often provide insufficient information, failing to capture the complex physiological regulation of BP changes, and many are validated only under resting conditions, restricting their real-world applicability. While data-driven models demonstrate improved performance, they often lack interpretability and are heavily dependent on large good-quality datasets for training, which constrains their ability to generalize effectively and increases computational demands.

Thus, to address these challenges, this thesis focuses on investigating and developing multimodal signal-based models that extended beyond the widely used single/dual-modality approaches for wearables continuous BP monitoring, aiming to improve the accuracy even in dynamic situation. The key objectives include: 1) conducting cold pressor test (CPT) on human subjects to induce large BP variations and collect five physiological signals including continuous BP (cBP), electrocardiogram (ECG), photoplethysmogram (PPG), impedance plethysmogram (IPG) and skin temperature (ST) signals to establish the multimodal dataset with dynamic changes; 2) analyzing the effects of CPT on multiple physiological signals and related parameters; 3) developing a new mechanism-driven multimodal physiological model (MPM) that incorporates bioimpedance and temperature features beyond traditional pulse transit time (PTT) to enhance noninvasive and continuous dynamic beat-to-beat BP estimation under personalized calibration; 4) designing McBP-Net, a multimodal deep learning (DL) model, systematically comparing different input combinations, and varying amounts of ground truth data to evaluate and improve the dynamic BP estimation performance under hybrid calibration; 5) proposing a multimodal physiological-informed model (MPIM) that combines physiological mechanisms leveraging physiological-informed loss functions and sparse regression to further enhance accuracy, particularly under population calibration.

To achieve the mentioned specific objectives, we carried out the research activities as follows. Firstly, conducting CPT to induce large BP variations in 23 human subjects. During this procedure, multimodal physiological signals, including cBP, ECG, PPG, IPG, and ST were recorded simultaneously and the multimodal dataset with dynamic changes were established.

Secondly, we systematically investigated the effects of CPT on multiple physiological signals and related parameters. Analysis of CPT effects revealed significant increases in heart rate (HR) and BP (p < 0.001), reduced HR variability (HRV), and increased BP variability (BPV). Additionally, IPG-derived HR showed better alignment with ECG compared to PPG-derived HR. Therefore, wearable devices incorporating IPG sensors can offer a more reliable and gesture-free/hands-free alternative for HR estimation. Furthermore, significant changes in physiological parameters during CPT highlighted the critical importance of considering temperature as a factor in BP estimation models.

Thirdly, a new mechanism-based physio-photo-electro MPM was developed as a function of PTT, heartbeat cycle, bioimpedance features, and ST extracted from multimodal signals. The MPM models arterial dimension changes through IPG features, combined with PTT for the determination of pulse pressure (PP). By incorporating ST in BP estimation via total peripheral resistance (TPR) changes and utilizing the heartbeat cycle to reflect the temperature-mediated sympathetic nervous system response associated with BP regulation, the proposed MPM can provide a multi-angle assessment of BP. Validated under personalized calibration, MPM achieved mean absolute errors (MAEs) of 5.78 and 4.15 mmHg for SBP and DBP, outperforming two widely cited PTT-based models with at least a 32% improvement (p < 0.001). These findings highlight the effectiveness of leveraging multimodal signals for dynamic BP estimation while further accuracy improvement is still needed.

Fourthly, inspired by the MPM, we proposed McBP-Net, built with hybrid CNN-LSTM DL architecture to capture both local features and temporal dependencies of multimodal signals to continuously estimate dynamic beat-to-beat BP. Validated under hybrid calibration, it achieved the MAEs of 4.19 and 2.98 mmHg for SBP and DBP, respectively, falling within the accuracy range required by the Grade A of IEEE standard. The integration of four multimodal signals improved performance by 16.20%, 37.37%, and 49.52% over three-, dual-, and single-modality approaches, respectively, with significant contributions from IPG and ST signals. Notably, ST showed a strong nonlinear relationship with BP, with a high mutual information of 0.91 for SBP. Furthermore, McBP-Net achieved a balance between accuracy and computational efficiency, offering inference speed of 36.7% faster and reducing computational demands by 78% compared to tested transformer-based models, while demonstrating only minimal degradation (0.21 mmHg) in dynamic SBP estimation when trained on rest-stage data. Despite these advancements, McBP-Net remained insufficient under population calibration with MAEs of 13.68 mmHg for SBP and 8.10 mmHg for DBP, necessitating further improvements.

Finally, to further enhance accuracy especially under population calibration, we developed the multimodal physiological-informed model (MPIM) that integrates physiological mechanisms into neural network training. This was achieved by constructing two key physiological-informed loss components and sparse regression alongside conventional neural network loss, ensuring the network was optimized for both data fitting and physiological consistency, thereby enhancing interpretability and computational efficiency. Under population calibration, MPIM achieved MAEs of 7.13 and 4.52 mmHg for SBP and DBP, with DBP accuracy falling within the thresholds of the AAMI standard and IEEE/BHS Grade A criteria. Compared to a conventional DNN with the same DL architecture but no physiological constraints, MPIM demonstrated an average accuracy improvement of 40.88% over DNN (11.59 and 7.97 mmHg for SBP and DBP) and 46.04% over McBP-Net (13.68 and 8.10 mmHg for SBP and DBP) from Chapter 5. Under personalized calibration, MPIM achieved MAEs of 4.95 mmHg (SBP) and 4.08 mmHg (DBP), improving accuracy by 20.25% over DNN and 8.03% over MPM from Chapter 4. In hybrid calibration, MPIM continued to outperform DNN, but its MAE values (4.30 mmHg for SBP and 3.15 mmHg for DBP) were slightly higher than McBP-Net. Under both personalized and hybrid calibrations, SBP and DBP accuracy aligned with the AAMI standard and IEEE/BHS Grade A criteria. Comparative evaluations confirmed MPIM’s superior performance over other ML and DL models, while ablation studies highlighted the critical contributions of physiological-informed loss components and sparse regression. Feature importance and correlation analyses identified IPG- and ST-derived features as key contributors to BP estimation accuracy, consistent with findings from Chapters 4 and 5. These results validate the effectiveness of integrating physiological-based constraints to improve BP estimation. By leveraging AI to extract high-dimensional features and incorporating mechanism constraints, MPIM effectively captures complex nonlinear relationships between multimodal signals and BP while preserving an interpretable physiological framework, demonstrating significant potential for improving beat-to-beat BP estimation accuracy and reliability.

Additionally, the appendix summarizes original contributions to the IEEE Standard for Wearable Cuffless Blood Pressure Measuring Devices (IEEE 1708/D3) regarding grading criteria, providing a comparative analysis against established standards including AAMI, BHS, ESH, and ISO. Notably, IEEE 1708/D3 Grade A remains the strictest while closely aligning with the widely accepted AAMI standard for sphygmomanometers. Updates to grades B and C incorporate |ME| ≤ 5 mmHg alongside MAE thresholds as refined criteria. A comparison of predicted and experimental results demonstrates that the simplified MAE-SD relationships can effectively characterize BP error distributions without complicated calculation or testing, enabling high-accuracy preliminary assessments in models with small MEs. Furthermore, these findings support that BP error data is more accurately represented by a t4 distribution rather than a normal distribution, particularly considering the recommendation of 85 testing subjects in most standards.

In summary, this thesis addresses key challenges in wearable BP monitoring by developing three multimodal signal-based models that enhance accuracy under dynamic conditions. The proposed MPM establishes a foundation for integrating physiological mechanisms, while McBP-Net demonstrates the effectiveness of multimodal DL for BP estimation. Finally, MPIM bridges physiological principles with AI-driven modeling, achieving significant accuracy improvements particularly in population-level calibration. Collectively, these advancements contribute to the development of medical-grade wearable continuous BP monitoring technologies, supporting proactive healthcare and early intervention in CVD management.

Date of Award	8 Jul 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Xinge YU (Supervisor), Ho Man CHAN (Co-supervisor) & Yuanting Zhang (External Co-Supervisor)