基於條件隨機場的心律不整自動分類系統

Abstract

心電圖是檢查心臟功能的重要工具，可以協助醫生臨床診斷心律不整的相關疾病。現今心電圖診斷結果必須依賴醫生依據其專業知識做判斷。在病患突發狀況下可能緩不濟急，為了正確且快速地辨識心律不整病徵，本論文在不需要醫生協助的情況下提出機器學習的方式對大量的心電圖數據進行快速分析與診斷。研究方法主要分為兩個部分，前級的心跳分類，以及後級的節律偵測。心跳分類是將心電圖經過雜訊消除濾波器後分段處理，利用RR區間與離散小波轉換係數擷取其特徵參數，再分別比較決策樹、條件隨機場和最近鄰分類器。最後針對分類輸出的心跳型態所組成的節律，利用可以描述相鄰心跳間相互關係的條件隨機場進行自動分類。系統模擬顯示本文提出的心跳分類方法的正確率為98.9%，而使用人工標註心跳分類的節律，條件隨機場的分類正確率最高。

Electrocardiogram (ECG) plays an important role in physical examination of cardiac diseases. Recent studies have suggested the use of a knowledge-based deterministic automaton for automatic arrhythmic classification. To eliminate the need of prior knowledge about rhythm rule, we propose a new method based on conditional random field (CRF) which allows to capture context information between successive heartbeats. We began by classifying each single heartbeat into four categories: normal, premature ventricular contraction, ventricular flutter and heart block. Results of beat classification are then used as input of a CRF-based classifier to classify five rhythm types, including ventricular bigeminy, ventricular trigeminy, ventricular tachycardia, ventricular flutter and heart block. Experiments on the MIT-BIH arrhythmia database demonstrated the validity of the proposed method with high accuracies for arrhythmic beat and rhythm classification.