Development of non-invasive wearable continuous blood pressure estimation system using multimodal pulsatile signals = 멀티 모달 맥동 신호를 이용한 비침습 웨어러블 연속 혈압 모니터링 시스템의 실증적 연구

다국어 초록 (Multilingual Abstract)

Although continuous blood pressure (cBP) is a useful physiological indicator that can accurately and immediately notify changes in blood pressure, it is not utilized in daily life due to the limitations of the invasive measurement method and the portability of the measuring device. The purpose of this study is to propose a portable wearable continuous blood pressure measurement technology that can be used in daily life by convergence of the tonometry technique and the multi-wavelength photoplethysmogram (PPG) technique, and to validate its feasibility through clinical trials.
A multi-model wearable system was developed, and it was equipped with a multi-wavelength PPG for measuring depth-dependent blood volume and a multi-channel pressure sensor to measure changes in pressure over the radial artery. The blood pressure estimation algorithm was developed based on the bidirectional long short-term memory (Bi-LSTM) model. For validation feasibility of the developed system, a clinical trial was performed on a total of 26 patients, and signals were simultaneously acquired using the commercial invasive blood pressure measuring device (IBP) and the developed wearable device, and the resulting blood pressure values were compared. Among the 49 data repeatedly obtained from participants, 33 data with the highest quality were used for algorithm development.
In this study, two model types were trained and evaluated: a participant-specific Bi-LSTM model and a participant-independent (generalized) Bi-LSTM model. The acquired data was segmented into lengths of 2, 3, 4, and 5 seconds8. The participant-specific model demonstrated the best performance with the 2-second segmentation. The performance of the participant-specific Bi-LSTM model was evaluated by comparing data processing methods and validation strategies. The application of PCA preprocessing significantly enhanced the model's precision; under 10-fold cross-validation, the PCA-based model achieved an estimation error of (2.4 ± 0.8) mmHg for SBP, (1.1 ± 0.3) mmHg for DBP, and (1.8 ± 0.5) mmHg for cBP (coefficient of determination with 0.97, 0.98, and 0.99), which was a marked improvement over the non-PCA model's error of (2.9 ± 1.0) mmHg for SBP, (1.3 ± 0.3) mmHg for DBP, and (1.9 ± 0.5) mmHg for cBP. Furthermore, in chronological validation, the PCA-based model demonstrated robust performance with an error of (5.4 ± 4.0) mmHg for SBP, (2.2 ± 1.2) mmHg for DBP, and (3.6 ± 2.0) mmHg for cBP (coefficient of determination with 0.84, 0.91, and 0.94), consistently outperforming the non-PCA approach, which showed an error of (5.9 ± 3.8) mmHg for SBP, (2.2 ± 1.2) mmHg for DBP, and (3.7 ± 1.8) mmHg for DBP.
Based on the segmentation length findings, a participant-independent Bi-LSTM model applicable to all participants was developed using PCA-processed data from the entire cohort. The estimation error showed significant improvement, with (3.5 ± 3.6) mmHg for SBP and (1.6 ± 1.6) mmHg for DBP, as compared to IBP. The coefficients of determination were 0.92 and 0.95, respectively. The error for the continuous blood pressure waveform across all sample points was (2.6 ± 2.9) mmHg, with a coefficient of determination of 0.97. It was confirmed that the participant-independent Bi-LSTM model developed with PCA-processed data demonstrates a performance equivalent to Grade A of the British Hypertension Society (BHS) protocol, which is a certification standard for intermittent non-invasive blood pressure devices.
To ensure the usability of the system developed in this study in both clinical and daily life settings, further improvements are necessary. These include enhancing the reproducibility of device placement, increasing robustness against external noise such as motion artifacts, and conducting additional validation on a more diverse group of participants.

목차 (Table of Contents)

• 1. Introduction 1
• A. Hypertension 1
• B. Continuous Blood Pressure Measurement 2
• 1) Diagnostic Errors in Hypertension 3
• 2) High Blood Pressure Variability 4
• 3) Disease Prevention and Management through Continuous Blood Pressure Measurement 4
• C. Invasive Blood Pressure Measurement 5
• D. Non-invasive Continuous Blood Pressure Measurement 6
• 1) Arterial Tonometry Method 7
• 2) Volume Clamp Method 8
• E. PPG-based Blood Pressure Estimation Method 10
• F. Related Research 11
• G. Objectives 13
• 2. Wearable Multimodal Signal Acquisition System 15
• A. Wristwatch-type Wearable System 15
• B. Multi-Wavelength PPG for Continuous Blood Pressure Estimation 18
• 1) Single-Wavelength PPG for Blood Pressure Estimation 18
• 2) Multi-Wavelength PPG for Blood Pressure Estimation 18
• 3) Advantages of Multi-Wavelength PPG in Blood Pressure Estimation 19
• 4) Development of Multi-Wavelength PPG Sensor 21
• C. Multi-Channel Pressure Sensor and Wrist-Worn Wearable Design 24
• 1) Structural of the Hybrid Sensing Platform 24
• 2) Signal Quality Optimization and Wearability Improvement 26
• 3) Development of Multi-Channel Pressure Sensor 27
• 4) Development of Wrist-Worn Wearable Design 28
• D. Data Acquisition Method 31
• E. Data Acquired 34
• F. Result of Multimodal Signal Acquisition 36
• 3. Machine Learning Based Blood Pressure Estimatin Using Multi-Channel PPG and Radial Arterial Pressure Signals 38
• A. Preprocessing 38
• B. Bidirectional Long Short-term Memory Model 45
• C. Validation and Evaluation 47
• D. Statistical Analysis 48
• E. Blood Pressure Estimation Performance of Participant-Specific Model 49
• 1) 10-Fold Cross-Validation Approach 49
• a) SBP and DBP Estimation of Participant-Specific Model Trained Using Multi-Channel PPG and Radial Arterial Pressure Data 54
• b) cBP Waveform Estimation of Participant-Specific Model Trained Using Multi-Channel PPG and Radial Arterial Pressure Data 55
• 2) Chronological Validation Approach 55
• a) SBP and DBP Estimation of Participant-Specific Model Trained Using Multi-Channel PPG and Radial Arterial Pressure Data 59
• b) cBP Waveform Estimation of Participant-Specific Model Trained Using Multi-Channel PPG and Radial Arterial Pressure Data 60
• 4. Machine Learning Based Blood Pressure Estimation Using PCA-Processed PPG and Radial Arterial Pressure Signals 61
• A. Principal Component Analysis 61
• B. Preprocessing 62
• C. Two BP Estimation Models: Participant-Specific and Participant-Independent 65
• D. Bidirectional Long Short-term Memory Model 66
• E. Validation and Evaluation 69
• F. Statistical Analysis 70
• G. Blood Pressure Estimation Performance of Participant-Specific Model 71
• 1) Comparative Analysis of Estimation Performance According to the Number of Bi-LSTM Layers 71
• 2) 10-Fold Cross-Validation Approach 74
• a) SBP and DBP Estimation of Participant-Specific Model Trained Using Multi-Channel PPG and Radial Arterial Pressure Data 77
• b) cBP Waveform Estimation of Participant-Specific Model Trained Using Multi-Channel PPG and Radial Arterial Pressure Data 78
• 3) Chronological Validation Approach 78
• a) SBP and DBP Estimation of Participant-Specific Model Trained Using PCA Preprocessed Data 82
• b) cBP Waveform Estimation of Participant-Specific Model Trained Using PCA Preprocessed Data 83
• H. Blood Pressure Estimation Performance of Participant-Independent Model 83
• 1) SBP and DBP Estimation of Participant-Independent Model Trained Using PCA Preprocessed Data 87
• 2) cBP Waveform Estimation of Participant-Independent Model Trained Using PCA Preprocessed Data 87
• 5. Discussion 89
• A. Wearable Multimodal Signal Acquisition System 89
• B. Blood Pressure Estimation from Multi-channel PPG and Radial Arterial Pressure Signals 90
• C. Blood Pressure Estimation with PCA-processed PPG and Radial Arterial Pressure 92
• D. Blood Pressure Estimation Performance in Chronological Validation Approach 94
• E. Comparison of Blood Pressure Estimation Model Performance 96
• F. Limitations 99
• 1) Data Quality and Measurement Failure 99
• 2) Time Delay in Blood Pressure Estimation using PAC-Processed Data 100
• 3) Verification Environment and Guideline for cBP Monitoring System 102
• 6. Conclusion and Future Plans 103
• References 104
• 국문초록 118