| 摘要: | 得益於公開生物醫學數據、漸進式可穿戴傳感器、電子健康記錄、機器學習技術和強大的計算機資源的普適性，基於人工智能的醫療保健發展激起人們的興趣。於本論文中，我們使用諸如全連接網絡 (FCNN)、長短期記憶神經網路 (LSTM) 和 卷積神經網絡 (CNN)等各種深度學習技術來進行血壓估計 (BP)；並且探討了深度學習本位的個人化 BP 預測系統於多種設置上的效果。首先，我們開發了一個LSTM 模型從心電圖 (ECG) 和光電容積描記圖 (PPG) 訊號中提取特徵，並用於預測收縮壓 (SBP) 和舒張壓 (DBP) 的數值。儘管該模型可以較傳統的機器學習模型有更好的成效，但我們提出了一種更簡單的架構── FCNN 模型，用於從PPG 訊號中提取的特徵集來預測 BP，它不僅具有更好的性能，並且證實了系統的實用性 (無需ECG訊號)。其次，我們利用 CNN 作為模型中的自動特徵提取來實現一種無需特徵工程的 BP預測系統；並採用這種方法來解決噪聲訊號所致特徵提取失敗的問題。此外，我們訓練了一個自監督式LSTM 自動編碼器模型，不僅可以預測 SBP 和 DBP 的離散值，還可以預測整個動脈 BP (ABP) 的波形。本論文呈現出基於深度學習的模型進行無需特徵工程的BP 估計，不僅避免了耗費時間成本的特徵提取和篩選過程，並揭示了BP訊號中非預期的週期模式；本文也提出了一種實時BP預測的移動裝置應用，為其開發提供了可行性。總體而言，本論文的研究結果顯示出無袖帶血壓估計 (cuffless BP estimation) 的深度學習技術可以增強和擴展現有回歸模型、開闢新的研究機會並有助於構建實時健康監測系統。;The wide availability of public biomedical data, progressive wearable sensors, electronic health records, the advancement of machine learning techniques, and powerful computer resources contribute to the significant interest rise in artificial intelligence-based healthcare development. In this dissertation, we focus on blood pressure (BP) estimation using various deep learning techniques, including a Fully Connected Neural Network (FCNN), a Long Short-Term Memory (LSTM), and a Convolutional Neural Network (CNN). This dissertation explores the efficacy of personalized deep learning-based BP estimation systems in several settings. First, we develop an LSTM model for estimating systolic BP (SBP) and diastolic BP (DBP) values from a feature set extracted from electrocardiogram (ECG) and photoplethysmogram (PPG) signals. Although this model could outperform the traditional machine learning models, we came up with a simpler architecture, an FCNN model, to estimate BP from a feature set extracted from PPG signal only, which not only performs better but also justify the practicality of an ECG-free system. Second, we introduce a feature engineering-free BP estimation system by utilizing CNN as the automatic feature extraction in our model. This approach is taken to tackle the problem of feature extraction failure due to heavy noises on the signal. Furthermore, we train a self-supervised LSTM autoencoder model to not only predict the discrete value of the SBP and DBP but the whole arterial BP (ABP) waveform. This dissertation demonstrates how using deep learning-based models for featureless BP estimation not only avoids the expensive feature extraction and selection procedure but also reveals unexpected and intuitive patterns in the input data. Lastly, a real-time BP estimation mobile and watch application is presented in this dissertation, providing feasibility on its development. Overall, this dissertation′s findings supported the notion that deep learning techniques used for cuffless BP estimation could enhance and expand existing regression models, open up new research opportunities, and contribute in the building of real-time health monitoring applications. |
