兩種心電圖分析系統：低功耗隨機森林模型及資訊瓶頸深度學習模型

Abstract

心電圖為現代非常重要的醫學量測儀器，其種類也非常多元。 這篇論文提出了兩種心電圖的系統，分別針對不同的用途而做出不同導向的設計：第一種使用了低運算資源的系統設計以及隨機森林分類器，為低功耗導向設計；第二種使用了變異資訊瓶頸理論的深度學習模型，為高準確度導向設計。

在第一種系統中，我們使用了非常少量的運算資源、能量、記憶體來去抽取心電圖的必要特徵，也提出了一個即時的心電圖分類器可以運用這些特徵來分類各種不同種類的心律不整。 這套系統是建立在兩個delta-sigma調變器以及三套波偵測演算法上，我們的系統直接在調變器輸出位元流時直接做特徵擷取，如此可以大幅精簡系統架構，而系統的取樣頻率也設計成較低的250赫茲，且每一個心跳的特徵也會被壓縮成68位元的資料，約只佔其他方法的1.8%。 我們使用MITBIH的資料來進行訓練與驗證，並採用AAMI的黃金標準來去分類兩種主要的心律不整心跳：SVEB、VEB。 針對QRS波偵測演算法的效能，我們獲得了99.11%的f1 score，與其他先進方法相當，且過程中節省了大量的系統資源。 而對於心律不整的分類，我們的分類器分別在SVEB與VEB心律不整上各獲得了 79.48%與 95.51%的f1 score，也與其他先進的方法相當。 這套系統提供了可靠且準確方法來分析心電圖，而其最大的優勢在於它精簡的架構、低功耗、低記憶體、低複雜度的優點。 因此我們相信這套系統會在資源有限的裝置中具有很大的潛力，例如植入式裝置、穿戴式裝置等。

在第二種系統中，我們直接使用原始的心電圖信號搭配其標籤做監督式的深度學習來去分析心電圖。 一般的心電圖信號會因很多因素如體質、心情、環境而造成變異，我們將這些變異稱為病患本質資訊，並假設每一個心電圖信號的資訊可以分割成心律不整資訊、病患本質資訊等等，我們目標便是剔除病患本質資訊， 如此來讓神經網路免於受到病患的一些變異而導致的過擬合。 我們採用殘差捲積神經網路，並且搭配資訊瓶頸的理論來完成上述這項任務。 資訊瓶頸會層層分離與心律不整不相關的資訊，而其誤差函數與傳統交叉熵比起來，他更能著墨於搜尋心律不整的資訊，並排除其他不相關的雜訊。 因此這個方法的編碼器會輸出與心律不整最為相關的資訊，而編碼的結果會再透過一層的全連接層來完成心跳的分類。 優化的過程中我們使用了變異的上下界來去優化資訊瓶頸的目標函數，如此可以節省大量訓練時間，此外我們也做了很多視覺化的步驟來去確立編碼器的效能。 這個深度學習的系統分別在SVEB與VEB的f1 score上獲得了84.26%與96.45%的成效，優於許多文獻，因此我們相信這套系統可以在已經紀錄且儲存心電圖資料的設備上有很大的優勢，例如Holter裝置、12導程設備等。

Electrocardiogram (ECG) is a crucial medical instrument, and it is used in various medical applications. This paper presents two ECG systems, which are aim for different applications. One is the resource-saving architecture with random forest classifiers, which is resource-saving oriented. The other is a deep learning model with the variational information bottleneck, which is high accuracy oriented.

For the first system, we extract a few important features of electrocardiogram (ECG) signals using concise system architecture and a little power and memory. In addition, real-time classifiers are proposed as well to classify different types of arrhythmias via these features. The proposed feature extraction system is based on two delta-sigma modulators adopting 250 Hz sampling rate and three wave detection algorithms to analyze outputs of the modulators, extracting the features in the bit stream domain directly. It extracts essential details of each heartbeat, and the details are encoded into 68 bits data that is only 1.8% of the other comparable methods. To evaluate our classification, we use a novel patient-specific training protocol in conjunction with the MIT-BIH database and the recommendation of the AAMI to train the classifiers. The classifiers are random forests that are designed to recognize two major types of arrhythmias. They are supraventricular ectopic beats (SVEB) and ventricular ectopic beats (VEB). The QRS detection rate of the proposed wave detection algorithms achieves 99.11% F1 score, which is comparable to many state-of-the-art methods. Furthermore, the performance of the arrhythmia classification reaches to the F1 scores of 79.48% for SVEB and 95.51% for VEB, which are also comparable to the state-of-the-art methods. The method provides a reliable and accurate approach to analyze ECG signals. Additionally, it also possesses time-efficient, low-complexity, and low-memory-usage advantages. Benefiting from these advantages, the method can be applied to practical ECG applications, especially wearable healthcare devices and implanted medical devices, for wave detection and arrhythmia classification.

For the second system, we use deep learning to analyze the ECG signal. The supervised learning is based on the raw signal and the corresponding labels. Usually, there are lots of variances in an ECG signal, such as physique, mood, and environment. We define these variances as 'intrinsic difference', and we assume that an ECG signal is composed of intrinsic difference information, arrhythmia information, and so on. In order to to remove the intrinsic difference information and to focus on the arrhythmia information, we propose residual convolutional neural network and information bottleneck method. The design can make the neural network devote itself to investigate arrhythmia information and erase the redundant information, avoiding overfitting. Specifically, the information bottleneck squeezes the ECG signal into layered neural networks to extract the arrhythmia information. That is, the encoder will output the significant arrhythmia information, which is also the input of a-single-layer fully connection layer that output the classification results. Furthermore, we adopt variational upper and lower bound to optimize the objective function, and this variational approximation saves lots of time during training. In addition, the performance of the information bottleneck encoder is verified by scrutinizing the high dimension visualization plot. This deep learning model reaches to the F1 scores of 84.26% for SVEB and 96.45% for VEB, which are better than the state-of-the-art methods. Thus, this method can be applied to the recorded and saved ECG equipment, such as 12-leads ECG instruments, Holter devices, and so on.