整合卷積雙向長短期記憶網路與納維-斯托克斯方程以評估橈動脈平均動脈壓

Abstract

根據世界衛生組織（World Health Organization, WHO）在2023年的報告，全世界每三位成年人中就有一人罹患高血壓，其中超過一半的患者尚未被確診，增加罹患心血管疾病、心臟病或中風等併發症的風險，因此日常監測血壓是非常重要的。居家保健通常使用袖套式血壓計，在量測過程中，袖套會對量測部位施加壓力，對於嬰幼兒、年長者或皮膚有傷口的患者來說，可能會造成不便。 本研究的目標是開發一種簡易裝置，利用相機與波長為525 nm的發光二極體（Light-Emitting Diode, LED），開發光學非接觸式方法，量測橈動脈（Radial Artery）的影像式光體積描記法（Imaging Photoplethysmography, iPPG）訊號，以評估平均動脈壓（Mean Artery Pressure, MAP）。然而本系統的綠光光源在人體皮膚中的穿透深度有限，僅能得到真皮層組織的訊號，不足以深入量測到直接代表橈動脈血管體積變化量的訊號。所以本研究亦架設了採用可穿透更深組織的980 nm近紅外光光源之接觸式光體積描記法（Photoplethysmography, PPG）裝置，以得到橈動脈的PPG訊號。 本研究提出三個波形指標：時域訊號的偏度（Skewness）、頻域訊號之峰度（Kurtosis）以及主諧指標（Goodness Index），作為波形品質參數，並使用支持向量機（Support Vector Machine, SVM）模型訓練波形品質分類器，從影像中提取iPPG訊號。由於真皮層組織與橈動脈相連，本研究假設真皮層與橈動脈共享相似的生理資訊，將iPPG訊號作為訓練資料，PPG訊號作為目標訊號，使用卷積雙向長短期記憶網路（Convolutional Neural Network Bidirectional Long Short-term Memory Network, CNN-BiLSTM）模型訓練波形重建器，將代表真皮層的iPPG訊號重建為橈動脈的PPG訊號。 應用本裝置探討動脈壓模型與受試者實際數據之關係，應用納維-斯托克斯方程（Navier-Stock Equations）的推導，介紹血液動力學著名的哈根-泊肅葉流（Hagen-Poiseuille's Law），用於描述血管壓降與血流量之間的關係。使用修正型布格-比爾-朗伯定律（Modify Bouguer-Beer-Lambert Law, MBLL）推導出PPG與iPPG訊號強度與血流量之間的自然對數關係。最後從生理調節的觀點出發，結合心率變異性（Heart Rate Variability, HRV）以及正副交感神經拮抗等關係，對二十位健康受試者建立兩個時域和兩個頻率域平均動脈壓迴歸模型。分析結果顯示，頻率域模型的表現優於時域模型；正副交感神經拮抗狀態的模型優於心率變異性模型。 綜觀本研究，採用相機拍攝橈動脈，並通過機器學習提升量測訊號，結合血液動力學與生理調節的影響，完成一個低成本、非接觸式且低負擔的平均動脈壓量測裝置雛形。此裝置在減少患者不適的同時，提供了一種便捷有效的血壓監測方法。參考國際協會新型血壓計開發標準，本裝置屬於C等級。由於研究過程推論出膚色是一個重要的參數，此為未來開發成商用儀器需要先進一步考量和調整的因素。

According to a 2023 report from the World Health Organization (WHO), approximately one in three adults worldwide suffers from hypertension, with over half of these cases going undiagnosed. This significantly increases the risk of complications such as cardiovascular disease, heart attacks, and strokes. Therefore, regular blood pressure monitoring is essential. Traditional cuff-based monitors, widely used in home healthcare, may be inconvenient for infants, seniors, or patients with skin wounds as they apply pressure to the measurement site. This study aims to develop a simple device that utilizes a camera and a 525 nm wavelength Light-Emitting Diode (LED) to create an optical non-contact method for measuring Imaging Photoplethysmography (iPPG) signals from the radial artery in order to evaluate Mean Artery Pressure (MAP). However, the green light only captures signals from the dermal layer of human skin due to its limited penetration depth, making it inadequate for directly measuring volume changes in the radial artery. Consequently, this study established a contact-based Photoplethysmography (PPG) device using a 980 nm near-infrared light, with deeper tissue penetration capabilities, to obtain the radial artery's PPG signals. This study introduces three waveform indicators: skewness of the time-domain signal, kurtosis of the frequency-domain signal, and the Goodness Index as parameters for assessing waveform quality. A Support Vector Machine (SVM) model was employed to train a classifier for waveform quality to extract iPPG signals from images. Based on the hypothesis that the dermal layer and the radial artery share similar physiological information, iPPG signals were utilized as training data. In contrast, PPG signals were used as target signals. A Convolutional Neural Network Bidirectional Long Short-term Memory Network (CNN-BiLSTM) model was trained to reconstruct the radial artery's PPG signals from the iPPG signals of the dermal layer. This device was applied to investigate the correlation between arterial pressure models and actual subject data. Hagen-Poiseuille's Law, derived from the Navier-Stokes Equations, explained the connection between vascular pressure drop and blood flow. Additionally, the Modified Bouguer-Beer-Lambert Law (MBLL) was utilized to establish the natural logarithmic relationship between PPG and iPPG signal intensities and blood flow. The study established two time-domain and two frequency-domain regression models for MAP based on heart rate variability (HRV) and the antagonistic relationship between the sympathetic and parasympathetic nervous systems. The analysis revealed that frequency-domain models outperformed time-domain models and that models considering the antagonistic state of the nervous systems outperformed HRV models. In conclusion, this study has designed an affordable, non-invasive, and user-friendly device for measuring MAP by utilizing machine learning to process images of the radial artery. Considering the effects of hemodynamics and physiological regulation, a prototype of this device has been created. In accordance with the International Association for the Development of New Blood Pressure Monitors standards, this device falls under Grade C. Given that skin color was identified as a significant factor during the research, further adjustments and considerations are necessary for its progression into a commercial instrument.