開發以卷積神經網路為基礎之心律不整偵測系統並分析資料長度的效應

心律不整為心臟十分常見之疾病，以心房顫動為例，全球每年約三百萬人被診斷出患有心房顫動，但由於心房顫動可能並無症狀，所以在數字是被低估的。由於心律不整並非一直存在，有時患者不舒服時去醫院檢查，待檢查時又恢復正常，現在可透過穿戴裝置長時間追蹤心電圖（Electrocardiography，ECG），在心律不整發生時能即時發現，如此病患無須在醫院就能初步檢測出有無心律不整。由於心律不整中有些相似波形但並非同樣類型的心律，目前大多數文章並沒有探討資料長度與相似波形對於卷積神經網路模型的影響。本文以六個波形相似的心律不整與兩個需要立即救治的心律不整為基礎以卷積神經網路開發一個心律不整偵測系統，並測試心電圖資料時間長度的影響。本實驗從Physio Net中WFDB抓取MITDB、AFDB與VFDB三個資料庫中八種心律資料，以資料庫內建的標記做為分類依據，切割後取得五秒的訓練集有5234個、驗證集有2250個；十秒的訓練集有5434個、驗證集有2250個並五秒與十秒的訓練資料以卷積神經網路做訓練並測試，最後在五秒的神經網路取得99.5%的準確度，十秒取得99.4%的準確率。本研究發現，五秒與十秒的心電圖皆可準確分類心律不整，因此真實應用時，五秒的時間已足以用以偵測心律不整的發生。

關鍵字

心律不整；卷積神經網路；心電圖

並列摘要

Arrhythmia is a very common disease of the heart. Take atrial fibrillation as an example. About 3 million people are diagnosed with atrial fibrillation every year. However, because atrial fibrillation may be asymptomatic, the number is underestimated. Since arrhythmia does not always exist, sometimes patients go to the hospital for checkups when they are uncomfortable, but their electrocardiography (ECG) turn back to normal when they are checked. Now the ECG can be tracked for a long duration through the wearable device, and the arrhythmia can be detected immediately instead of checking ECG in the hospital. There are some similar ECG waveforms in different types of arrhythmia, most of the current published articles did not discuss the effect of data length on the development of arrhythmia detection model. This study attempted to develop a convolution neural network for detecting six similar types of arrhythmia and two of needed instant treat arrhythmia. And the performances of different ECG length was compared. This study got eight types of rhythm data from MITDB, AFDB and VFDB by WFDB from Physio net, the arrhythmia classify depend on the database note, after cutting we have 5234 training set and 2250 validation set in the data set of five seconds, in data set of ten seconds we have 5435 training set and 2250 validation set, here has five seconds and ten seconds data for input to training validation by convolution neural network, after validation we got 99.5% accuracy in five seconds network and 99.4% accuracy in ten seconds network. The results revealed that accuracy of our proposed models were good enough for arrhythmia detection and performance for the 5 seconds and ten seconds data length were similar. Thus the five seconds ECG is enough for using in real application.

並列關鍵字

Arrhythmia ； Convolution Neural Network ； Electrocardiogram

主題瀏覽