- 學位論文
使用光體積變化描記圖估測血壓之研究
Blood Pressure Estimation Using Photoplethysmography
摘要
本研究提出一血壓估測系統，使用光體積變化描記圖(Photoplethysmography, PPG)為生理訊號，估測收縮壓與舒張壓。 本研究資料取自於民國102年國立中正大學心理系翁嘉英教授所主持之科技部研究計畫(NSC102-2410-H194-022)，系統流程依序為資料分類、生理訊號擷取、特徵點偵測、特徵擷取以及血壓估測五個部分。在資料分類方面，將資料分為四大類(健康大學生、高血壓吃藥、高血壓為吃藥、血壓正常但可能有其他疾病)，且每類有四期(基線期、回憶期、描述期、恢復期)。在生理訊號擷取上，依不同類別作訊號擷取並偵測重要特徵點，接著計算各類特徵，包含時間特徵、振幅特徵、波寬特徵、心律變異(HRV)特徵、性別以及年齡，總共65個。在估測方法上，使用倒傳遞類神經網路(Back propagation neural network, BPNN)，經實驗比較後，參數設定如下：隱藏層層數為2層，轉移函數為對數雙彎曲，隱藏層的神經元數目為130。 在結果方面，All train all test驗證在各情況誤差皆能在3mmHg以下；而在兩類(健康大學生、血壓正常但可能有其他疾病)取三期(無描述期)時，使用Leave one out交叉驗證，收縮壓及舒張壓誤差分別為8.598 mmHg、6.89 mmHg；兩類皆只取基線期時，收縮壓及舒張壓誤差分別為7.746 mmHg、6.81 mmHg。
並列摘要
In this study, we proposed a blood pressure estimation system. We estimated systolic and diastolic blood pressure by using Photoplethysmography (PPG). The database collected by the projects supported by the grants from Ministry of Science and Technology (NSC 102-2410-H-194-022) to Chia-Ying Weng. The system contained five parts, namely data classification, physiological signal acquisition, feature points detection, feature extraction, and blood pressure estimation. In the data classification part, data were divided into four categories including healthy students, hypertensive patients with medication, and hypertensive patients without medication and people but might have other diseases with normal blood pressure. Each of the categories had four stages: baseline, recall, description, and recovery. In the physiological signal acquisition part, we acquired physiological signal and detected important feature points from different categories of signals. And then totally 65 features, including the features of time, amplitude, waveform, HRV, gender, and age were calculated. In the estimation part, we used the back-propagation neural network (BPNN). The parameters were empirically set as follows: 2 hidden layers, transfer function of log-sigmoid, and 130 neurons in the hidden layer. The results showed that the error in each case was always limited within 3 mmHg by using all-train-all-test validation. When the signals were acquired from healthy students and people with normal blood pressure but might have other diseases and only the more stable three stages, including baseline, recall, and recovery were considered the errors in systolic and diastolic blood pressure were 8.598 mmHg and 6.89 mmHg, respectively, by using leave one out cross-validation. On the other hand, when the data contained the same two categories mentioned above and only the baseline stage was considered, the systolic and diastolic blood pressure errors were reduced to 7.746 mmHg, 6.81 mmHg, respectively.