以深度卷積神經網路結合心電圖與光體積變化描記圖訊號在心律不整的判別應用

Abstract

隨著醫療產業的進步，心臟相關疾病的預防以及治療都有明顯改善，但在面對心律不整(Arrhythmia)時其仍然是臨床和公衛方面的一大挑戰。由於節律的不規則使血液無法完全地從心臟輸出，很可能會形成血栓而導致最終造成中風。其中最常見的心律不整為心房震顫(Atrial fibrillation, AF)，在臨床上的特徵表現為心房快速且不規律的異常收縮，估計至2050年患有心房震顫的人數將增加2到3倍。 有鑑於光體積變化描記圖的量測裝置相較於心電圖裝置在穿戴式裝置上的便利性，也可以進一步提供在血流動力學上的生理訊息，因此本研究嘗試利用兩者之間的同步訊號，結合深度學習演算法在一維卷積神經網路上進行心律不整的辨識訓練，希望能在特定心血管疾病種類上實現以PPG單獨判別心律不整的可行性。 在現行的研究中，已經有許多演算法應用在偵測、預測或是分類心房震顫在心電圖訊號上，以此為基礎，本研究收集不同演算法分析心房震顫相關的文獻，選擇最合適的卷積神經網路作為模型架構，透過臨床試驗實際收集心律不整患者之心電圖(Electrocardiogram)與光體積變化描記圖(Photoplethysmography)資料庫，將訊號經過Moving Average與Difference等前處理，最後分割片段以投入進行5-fold Nested Cross Validation訓練，並進一步將訓練好的模型嘗試讀取PPG訊號達到實際偵測心律不整。 最終在訓練結果顯示出，透過心電圖與光體積變化描記圖兩種訊號之下，共同訓練的模型可以達到98.5 %的辨識準確率，實現在症狀發作前或是沒有任何明顯異常的表現當中，以PPG單獨判別心律不整的可行性，並在不同種類的辨識情況下同樣具備高準確率，初步證實透過多種心臟生理資訊有助於患者達到更精確的判別。 未來可以將辨識模型結合穿戴式裝置應用在需要長時間監測的患者上，並進一步持續加強辨識模型，納入更多實際病患的訓練資料以及不同類型的異常心律種類，達到連續監測並且方便有效的心律不整演算法系統。

With advancements in the medical industry, there have been significant improvements in the prevention and treatment of cardiovascular diseases. However, dealing with arrhythmias remains a major challenge in clinical and public health settings. The irregular heart rhythms can prevent blood from being fully pumped out of the heart, potentially leading to the formation of blood clots and eventual stroke. Atrial fibrillation (AF) is the most common type of arrhythmia, clinically characterized by rapid and irregular contractions of the atria. It has been estimated that the number of individuals with atrial fibrillation will double or triple by the year 2050. Considering the convenience of wearable devices for photoplethysmography (PPG) measurements compared to electrocardiography devices, as well as the additional physiological information they provide on hemodynamics, this study attempts to use synchronized signals from both modalities and apply deep learning algorithms for arrhythmia recognition training. The aim is to achieve the feasibility of solely detecting arrhythmias using PPG in specific cardiovascular conditions. In current research, various algorithms have been applied to detect, predict, or classify atrial fibrillation in electrocardiogram (ECG) signals. Based on this, our study collects and analyzes literature on different algorithms related to atrial fibrillation. We select the most suitable convolutional neural network (CNN) as the model architecture. Through clinical trials, we collect ECG and PPG data from patients with arrhythmias. The signals are preprocessed with techniques such as moving average and difference, segmented, and inputted for training using 5-fold nested cross-validation. Furthermore, the trained models are attempted to be applied to PPG signals for the actual detection of arrhythmias. The results show that the model jointly trained with ECG and PPG signals achieved a recognition accuracy of 98.5%. This demonstrates the feasibility of using PPG alone to identify arrhythmias, even before the onset of symptoms or in the absence of any obvious abnormalities. Additionally, the model demonstrated high accuracy across different recognition scenarios, confirming that using multiple types of cardiac physiological information can help patients achieve more precise diagnoses. In the future, the recognition model can be integrated with wearable devices for continuous monitoring of patients requiring long-term monitoring, and further efforts can be made to continuously enhance the recognition model by incorporating more training data from actual patients and different types of abnormal rhythms, aiming for a continuous monitoring and convenient and effective arrhythmia algorithm system.