| 摘要: | 心血管疾病對於人類健康造成的威脅不容小覷，而血壓作為心血管疾病的重要指標，臨床上通常以血壓計配合壓脈帶進行量測，然而該方式除了有諸多不便之處，且容易受到白袍效應而無法得到客觀的數值外，亦無法連續地長時間追蹤數值。近年來受惠於穿戴式裝置的發展，許多研究紛紛投入無脈壓帶的血壓估測方法，卻在準確度、應用環境及實用性等層面面臨挑戰。本研究利用穿戴式手錶可量測的光體積描記訊號來估測血壓數值，並依照校正與否及應用範圍分為通用化模型及個人化模型。從裝置量測的訊號處理階段開始，先對波形進行前處理並去除基線飄移及市電訊號的干擾，接著設定預篩選條件過濾品質不良的訊號，再透過特徵提取與波形拆解來解析訊號特性與標記特徵值，最後利用訊號品質檢測確保提取的特徵具有代表性。在資料處理階段，對於無須校正的通用化模型，我們將資料分成訓練、驗證及測試資料，並對訓練資料進行資料擴充，避免模型因不均勻的資料分布導致估測偏差，而測試資料則依照美國醫療儀器促進協會(AAMI)的標準，選擇85人255筆且符合血壓區間規範之資料集，確保效能估測的客觀性。我們嘗試了三種機器學型模型，包含神經網路、卷積神經網路及殘差網路，經由驗證的估測標準差，在收縮壓上達到11.35(mmHg)、舒張壓達到8.75(mmHg)的準確度，在相同條件且符合AAMI驗證標準下優於其他文獻。對於有校正之個人化模型，我們利用遷移式學習的概念來輔助少樣本學習的應用條件，並利用模型無關之元學習演算法建立個人化模型，在受試者可自由活動及運動的前提估測其日間及夜間血壓，收縮壓之標準差達到 7.78(mmHg)，舒張壓之標準差達到6.99(mmHg)，夜間血壓下降之絕對誤差達4.45(mmHg)，我們在量測訊號種類較少、裝置較簡便的前提優於文獻上其他也是自由活動情境下之個人化模型。;The threat of cardiovascular disease to human health cannot be underestimated. As an important indicator of cardiovascular disease, blood pressure is usually measured clinically with a sphygmomanometer and a cuff. In addition to being unable to obtain objective value due to white-coat effect, it is also impossible to continuously track the value for a long time in daily life. In recent years, benefiting from the development of wearable device, many studies have invested in cuffless blood pressure estimation methods, but they face challenges in terms of accuracy, application environment and practicability. This study uses the photoplethysmography from the smart watch to estimate the blood pressure. A general model and a personal model are constructed respectively according to whether it is calibrated or not and the application. Starting from the signal processing stage, the waveform is pre-processed to remove baseline wandering and the interference, then the pre-screening conditions are set to remove the signal with poor quality. The charactersitics of PPG morphology are acquired through feature extraction and waveform decomposition, and finally signal quality assessment is used to ensure that the extracted features are representative. In the data processing stage, for calibration-free general model, we divide the data into training, validation and testing sets, and augment the training data to avoid estimation bias. The testing data is in accordance with with the standard of the Association for the Advancement of Medical Instrumentation (AAMI), selecting 85 people with 255 data that meet the blood pressure interval specifications of required distribution. We tried three machine learning models, including neural network, convolutional neural network and residual network, the estimated standard deviation reaches 11.35 (mmHg) and 8.75 (mmHg) for systolic and diastolic blood pressure. The accuracy is better than other literature with wrist PPG signals under the same conditions and in compliance with AAMI verification standards. For personal model with calibration, we use the concept of transfer learning to assist the application conditions of few-shot learning, and use the model-agnostic meta-learning algorithm to establish initial weights of the personal model, the standard deviation reached 7.78 (mmHg) and 6.99 (mmHg) for systolic and diastolic blood pressure respectively, and the mean absolute error of nocturnal blood pressure dip reached 4.45 (mmHg). The results outperform the 24-hour free-living personal model in the literature. |
