心電訊號動作假影雜訊抑制之深度學習演算法

心電圖是通過量測皮膚上兩點電極的電位差來記錄心臟的電生理活動，是診斷心臟疾病最基本也是最常使用的檢查項目，醫師會透過不同的檢驗方式來掌握心臟的各種狀態。而針對心臟病患在做完心導管或繞道手術後的術後監測，有越來越多的研究發現，適當活動對於心臟機能有良好的恢復。因此，隨著心電圖裝置的進步，越來越多醫院在住院期間，會以無線心電圖裝置監護病患以協助盡快恢復日常生活活動，卻也讓在量測期間，心電訊號受雜訊所影響的機率大幅增加。本論文研究重點在利用兩種深度學習模型對心電訊號的雜訊抑制進行建模，分別對兩種模型與不同的資料處理方式進行效能分析與討論。而在深度學習的架構下，我們需要有相對應的參考訊號（即真實可判讀的心電訊號）作為訓練的標準，但實際上我們無法得知當心電訊號受到動作假影干擾時，其真實可判斷的心電訊號資訊。所以我們使用來自Physio-net上的PTB診斷資料庫（PTB diagnostic database，PTBDB），在完全靜止狀態下所量測的訊號中，取出標記為Healthy Control的資料作為判讀的參考訊號來源。透過隨機加入不同強度由MIT-BIH Noise stress database（NSTDB）標記為Electrode motion artifact（EM）的資料作為動作假影的雜訊來源，將其合成具有動作假影的心電訊號資料，做為我們網路模型的數據集。同時我們也會針對日常活動狀態下量測心電圖所產生的動作假影雜訊，分析在不同的活動情況下，動作假影雜訊對於不同位置導極的影響，並使用上述兩種深度學習模型來抑制雜訊與分析其效能。

關鍵字

心電圖；動作假影；深度學習

並列摘要

The electrocardiogram records the electrical conduction system of the heart by measuring the potential difference between the two electrodes on the skin. It is the most basic and most commonly used test for diagnosing heart conditions. The doctor can understand the various states of the heart through different tests. For postoperative monitoring of heart disease patients after cardiac catheterization or bypass surgery, more studies have found that appropriate activities have a good recovery of cardiac function. Therefore, with the advancement of ECG devices, doctors will monitor patients with wireless ECG devices during their hospitalization to help restore daily activities as soon as possible. However, the probability that the ECG signal is affected by noise during the measurement period is greatly increased. The focus of this thesis is to use two deep learning models to suppress the motion artifact in ECG signals. After that, we will analyze and discuss the performance of the two models and different data processing methods. Under the framework of deep learning, we need to have a corresponding reference signal (the actual readable ECG signal) as the standard of training. In fact, we won’t know the truth readable ECG signal when the signal is interfered by motion artifacts. Therefore, we use the PTB diagnostic database (PTBDB) from the Physio-net to extract the data labeled Healthy Control as the reference source for the interpretation in the signal measured at the complete quiescent state. By randomly adding data of different level power, which marked by MIT-BIH Noise stress database (NSTDB) as Electrode motion artifact (EM), we will get the simulated noisy ECG signal as the dataset of deep learning model. At the same time, we will also measure the ECG signals in daily activities. Analyzing the effects of motion artifacts on the different positions under different activities, we will use these two deep learning models to suppress the motion artifact and evaluate its performance.