| 摘要: | 根據世界衛生組織統計，心血管疾病為全球一大主要死因，其急性症狀發作如心肌梗塞和腦中風都是危及生命的緊急情況，因此平時的日常照護和預防相當重要，盡早就醫能夠減少死亡率以及病發後遺症的嚴重程度。血壓與脈波傳導速率皆為心血管疾病的重要指標，脈波傳導速率需要於醫院使用儀器進行測量，而血壓的量測通常使用脈壓式血壓計，除了較為不適，也無法長時間監測，不利於睡眠時期血壓的監控。本論文利用穿戴式裝置收集生理訊號，分別以手腕的光體積描記訊號和心電訊號來估測脈波傳導速率，以及只用手腕的光體積描記訊號來估測血壓。對訊號會先進行前處理，去除市電60Hz的干擾及基線飄移，並對訊號做預篩選以濾除品質不良的波形，接著進到特徵提取與加權式波形拆解得到所需的特徵，而訊號品質檢測會去除不良波形所提取的特徵。在資料處理程序，取得每一特徵序列的中位數，並移除數值上的離群值以確保特徵具有代表性。將資料分為訓練資料集、驗證資料集與測試資料集，而為了資料區間的平衡，另外對訓練資料集進行複製。脈波傳導速率的估測以使用隨機森林的分類器搭配極限梯度提升回歸器來進行分層回歸模型的結果為最佳，男性平均的均方根誤差為145.0 cm/s，女性平均的均方根誤差為141.4 cm/s。而血壓的估測為校正式之個人化模型，以不拘束受試者的活動下能提供日夜間包含睡眠時期的血壓，使用模型無關之元學習演算法中的少樣本學習概念，僅需要少量的校正數據就能使模型快速收斂，提升了使用者的便利性。在使用5筆校正資料下，平均15人的收縮壓之平均誤差與標準差為0.06 ± 6.90 mmHg，均方根誤差為6.76 mmHg，舒張壓之平均誤差與標準差為0.01 ± 6.54 mmHg，均方根誤差為6.82 mmHg。而夜間血壓下降之平均絕對誤差為4.80 mmHg。而減少到使用3筆校正資料，收縮壓之平均誤差與標準差為-0.62 ± 7.62 mmHg，均方根誤差為7.57 mmHg，舒張壓之平均誤差與標準差為-0.32 ± 6.96 mmHg，均方根誤差為7.00 mmHg。而夜間血壓下降之平均絕對誤差為5.77 mmHg。;According to statistics from the World Health Organization, cardiovascular diseases are a major cause of death globally. Acute symptoms such as myocardial infarction and stroke are life-threatening emergencies, making daily care and prevention crucial. Early medical intervention can reduce mortality rates and the severity of post-incident complications. Both blood pressure (BP) and pulse wave velocity (PWV) are important indicators of cardiovascular disease. PWV measurements require hospital equipment, while BP is typically measured using a sphygmomanometer, which is uncomfortable and unsuitable for long-term monitoring, particularly during sleep. This study uses a wearable device to collect physiological signals. PWV is estimated using photoplethysmography (PPG) and electrocardiogram (ECG) signals from the wrist, and BP is estimated using only PPG signals. At signal preprocessing stage, 60Hz power line interference and baseline wandering are removed, and poor-quality waveforms are filtered out through pre-screening. Feature extraction and weighted waveform decomposition are then performed to obtain the features. Features extracted from poor-quality waveforms can be dispelled according to signal quality index. In the data processing step, the median of each feature sequence is obtained, and outliers are removed to ensure representativeness. The data is divided into training, validation, and test datasets. To balance the data range, the training dataset is additionally duplicated. The best result for PWV estimation uses a hierarchical regression model combining a random forest classifier and an extreme gradient boosting (XGBoost) regressor. The average root mean square error (RMSE) for males is 145.0 cm/s and 141.4 cm/s for females. Additionally, a calibrated personalized model provides daytime and nighttime BP estimates, including sleep period, in daily life. Using a model-agnostic meta-learning (MAML) algorithm, the concept of few-shot learning allows the model to converge quickly with only a small amount of calibration data, enhancing user convenience. With 5 calibration data, the mean error and standard deviation for SBP across 15 individuals are 0.06 ± 6.90 mmHg, with an RMSE of 6.76 mmHg. For DBP, the mean error and standard deviation are 0.01 ± 6.54 mmHg, with an RMSE of 6.82 mmHg. The mean absolute error for nighttime SBP dip is 4.80 mmHg. When reduced to 3 calibration data, the mean error and standard deviation for SBP are -0.62 ± 7.62 mmHg, with an RMSE of 7.57 mmHg. For DBP, the mean error and standard deviation are -0.32 ± 6.96 mmHg, with an RMSE of 7.00 mmHg. The mean absolute error for nighttime SBP dip is 5.77 mmHg. |
