根據世界衛生組織統計，心血管疾病長期位居全球主要死因之一，因此透過 日常監測進行早期預防具有重要意義。血壓與脈波傳導速率皆為評估心血管健康 狀態的重要指標，然而傳統量測方式多需仰賴專業設備或間歇式血壓計，不利於 長時間且連續性的健康監測。本研究利用穿戴式裝置所量測之生理訊號進行非侵 入式心血管指標估測。為降低特徵冗餘與模型複雜度，並提升估測模型之效能， 本研究進一步採用特徵選擇技術篩選具代表性之生理訊號特徵。在脈波傳導速率 估測方面，以手腕光體積描記訊號及心電訊號作為輸入；在血壓估測方面，則僅 使用手腕光體積描記訊號。首先對訊號進行前處理及品質檢測，以排除品質不佳 之波形，接著透過波形特徵提取、加權式波形拆解及特徵選擇建立脈波傳導速率 估測模型；而血壓估測則以模型無關式元學習為基礎，結合多步損失與多步梯度 下降策略進行個人化校正。本研究分別採用 最小冗餘最大相關性（minimum Redundancy Maximum Relevance， mRMR）與二元粒子群優化（ binary particle swarm optimization， BPSO） 進行分類與回歸模型之特徵選擇，並在特徵數量大幅減少的 情況下進行脈波傳導速率估測。實驗結果顯示，以每位受試者平均結果計算，男 性受試者之均方根誤差為 136.1 cm/s，女性受試者之均方根誤差為 141.5 cm/s， 顯示在降低特徵維度後仍可維持良好的估測效能。在血壓估測方面，當校正資料 為 5 筆時，15 位受試者平均收縮壓及舒張壓均方根誤差分別為 6.63 mmHg 及 6.45 mmHg，夜間收縮壓下降量之平均絕對誤差為4.82 mmHg；當校正資料減少 至3筆時，收縮壓及舒張壓均方根誤差分別為7.44 mmHg及7.56 mmHg，夜間 收縮壓下降量平均絕對誤差為4.68 mmHg。研究結果顯示，所提出方法能有效利 用穿戴式裝置量測之生理訊號進行脈波傳導速率與血壓估測，其中個人化血壓模 型僅需少量校正資料即可維持良好估測效能，具備應用於居家照護、睡眠監測及 個人化健康管理之潛力。;Cardiovascular diseases remain one of the leading causes of death worldwide. Blood pressure (BP) and pulse wave velocity (PWV) are important indicators of cardiovascular health; however， conventional measurement methods often require specialized equipment or cuff-based devices， limiting long-term continuous monitoring. This study investigates non-invasive cardiovascular parameter estimation using physiological signals acquired from wearable devices. For PWV estimation， wrist photoplethysmography (PPG) and electrocardiography (ECG) signals were utilized. Signal preprocessing， quality assessment， waveform feature extraction， weighted pulse decomposition， and feature selection using minimum redundancy maximum relevance (mRMR) for classification and binary particle swarm optimization (BPSO) for regression were performed to establish the estimation model. For BP estimation， a personalized calibration framework based on Model-Agnostic Meta-Learning (MAML)， Multi-Step Loss (MSL)， and multi-step gradient descent was developed using only wrist PPG signals. For PWV estimation， the average root mean square error (RMSE) was 136.1 cm/s for male subjects and 141.5 cm/s for female subjects. For BP estimation， when five calibration samples were available， the average RMSEs of systolic and diastolic blood pressure were 6.63 mmHg and 6.45 mmHg， respectively， with a mean absolute error (MAE) of 4.82 mmHg for nocturnal systolic blood pressure dipping. When only three calibration samples were used， the corresponding RMSEs were 7.44 mmHg and 7.56 mmHg， with a dipping MAE of 4.68 mmHg. The results demonstrate the feasibility of wearable-device-based PWV and BP estimation， while showing that accurate personalized BP estimation can be achieved using only a small number of calibration samples.
