| 摘要: | 利用穿戴式裝置量測生理訊號並估測血壓的方式，相對於目前普遍使用的壓脈帶式血壓計，具有非侵入性、便利性與可長時間連續監測等優點，能有效提供高血壓等心血管疾病的早期發現與長期管理的可能性。醫學研究指出，高血壓為心血管疾病的主要危險因子之一，若未能及時控制，可能導致中風、心肌梗塞與腎功能衰竭等嚴重後果。因此，發展準確且易於日常使用的血壓估測技術，對於提升整體公共健康具有重要意義。本論文旨在發展一套通用化且無需個人校正的血壓估測方法，藉由感測裝置量測手指端的生理訊號，包括光體積描記訊號(Photoplethysmography, PPG)與心電訊號(Electrocardiography, ECG)。透過一系列的訊號前處理與特徵萃取步驟，擷取可能與血壓變化高度相關的特徵(Hand-Crafted Features)。本研究導入基於注意力機制(Attention)的深度學習模型進行血壓估測，藉由編碼器(Encoder)捕捉PPG波形時間序列中的重要關聯性，並且透過Hand-Crafted Features穩定模型收斂並且提高模型的泛化能力。
血壓估測參考了國際量測血壓標準，以符合資料分布要求，總共85人共255筆量測資料作為測試資料集，實驗結果顯示，通用化收縮壓迴歸方式在不需個人化校正的前提下，整體收縮壓平均誤差與標準差為0.15±9.09 mmHg，另外通用化舒張壓迴歸方式，整體平均誤差與標準差為0.49±7.84 mmHg。
由於通用化血壓迴歸方法可能受到訓練資料集多樣性的限制，因此在處理特殊血壓族群時可能會發生過度自信以及估測較不準確的問題，本研究藉由貝氏神經網路量化模型的不確定性，探討資料不確定性以及模型不確定性。此外，本研究也透過生成式資料改善女性訓練資料量較不足的問題，男性則透過殘差網路模型，改善全域收縮壓估測準確度。實驗結果顯示改善後的男性收縮壓平均誤差與標準差為0.45±8.11 mmHg，女性收縮壓平均誤差與標準差為0.92± 8.68 mmHg。
;Compared to traditional cuff-based blood pressure monitors, using wearable devices to measure physiological signals offers advantages such as non-invasiveness, convenience, and the ability for continuous long-term monitoring. These benefits enhance the early detection and long-term management of hypertension and other cardiovascular diseases. Since hypertension is a major risk factor for conditions like stroke and heart attack, developing an accurate and easy-to-use blood pressure estimation method is crucial for healthcare of cardiovascular disease.
This study proposes a generalized, calibration-free blood pressure estimation approach using fingertip physiological signals, including photoplethysmography (PPG) and electrocardiography (ECG). After signal preprocessing and hand-crafted feature extraction, a deep learning model based on attention mechanisms is applied. The encoder captures temporal dependencies in the PPG waveform, while the extracted hand-crafted features help stabilize training and improve generalization.
Blood pressure estimation follows international standards, using 255 measurement samples from 85 individuals. Without personal calibration, the mean error and standard deviation for systolic blood pressure estimation are 0.15 ± 9.09 mmHg, and for diastolic blood pressure are 0.49 ± 7.84 mmHg.
To address the potential limitations of generalized models in special populations, Bayesian neural networks are used to quantify aleatoric and epistemic uncertainty. Additionally, synthetic data from a generative model is used to augment female training samples, while a residual network is applied for male subjects to enhance global systolic blood pressure accuracy. After improvement, the estimation systolic blood pressure errors are 0.45 ± 8.11 mmHg for male subjects, and 0.92 ± 8.68 mmHg for female subjects, respectively. |
