| 摘要: | 在現代社會中，心血管疾病是導致人類死亡的主要原因，而血壓的監測在預防、診斷、評估和治療疾病的過程中至關重要。傳統的血壓量測方法以示波法為主，此方法無法進行高頻率的血壓監測，更因為在量測期間須依靠壓脈帶對動脈施加壓力，進一步造成使用者的不適感。為了克服這些問題，並實現前所未有的便利性，許多研究利用穿戴式裝置中可取得的光體積變化描記圖（photoplethysmography，PPG）與心電圖訊號（electrocardiogram，ECG）進行血壓估測，實現無需壓脈帶的血壓監測方法。
本研究以深度學習方法中的Transformer模型為基礎，對光體積變化描記圖與心電圖訊號進行分析，以提取與血壓相關的特徵。為了更準確地呈現訊號中的局部資訊，我們將訊號切分成部分重疊的區塊，作為個別的Token輸入到模型中。同時，考慮到人與人之間生理訊號的差異，我們引入了校正程序，並額外納入年齡、性別、身高等資訊作為模型的參考依據。本研究在VitalDB資料集上進行實驗，此資料集包含1,437名病患的資料，我們提出的演算法在VitalDB測試集上的收縮壓平均誤差為 -0.11 ± 6.44 mmHg，舒張壓平均誤差為 -0.10 ± 4.07 mmHg，這些結果符合AAMI標準對於血壓量測設備的要求，進一步驗證了我們所提出演算法的有效性。
;In modern society, cardiovascular disease is the leading cause of human deaths, and blood pressure monitoring is essential for prevention, diagnosis, assessment, and treatment. Traditional blood pressure measurement mainly relies on oscillometric method, which cannot achieve high-frequency blood pressure monitoring. Furthermore, these methods require the application of pressure on the artery using a cuff, which can cause discomfort to the user. In order to address these challenges and achieve unparalleled convenience, numerous studies have employed the use of photoplethysmography (PPG) and electrocardiogram (ECG) signals, which can be acquired from wearable devices, to estimate blood pressure. This advancement allows for cuffless blood pressure monitoring.
This study is based on the Transformer model, a deep learning method, to analyze signals from PPG and ECG for extracting blood pressure-related features. To accurately capture local semantic information in the signals, we partitioned the signals into partially overlapping patches, which were treated as individual tokens inputted into the model. Moreover, considering the variabilities observed among individuals in physiological signals, we incorporated a calibration procedure and additional information such as age, gender, and height as reference factors for the model. The effectiveness of the proposed algorithm was evaluated on VitalDB datasets, which consists of data from 1,437 patients. In the VitalDB test set, our algorithm achieved mean errors of -0.71 ± 6.91 mmHg for systolic blood pressure and -0.69 ± 4.22 mmHg for diastolic blood pressure, meeting the standards set by the Association for the Advancement of Medical Instrumentation (AAMI) for blood pressure measurement devices. These results further validate the efficacy of the algorithm proposed in this study. |
