心電圖作為非侵入的心臟電信號量測方法,目前已經被廣泛地應用在心臟疾病的診斷上。在這篇論文中將基於心電圖和K-SVD演算法,提出一種心臟疾病的診斷方式,並採用MIT-BIH心律不整資料庫中的記錄來驗證所提出的方法,方法主要包含兩個部分,其中一個是使用足夠地病徵(正常)信號樣本進行特徵擷取的工作,另一部分是將偵測信號與每個病徵特徵做比較,並以最接近的特徵做為病徵偵測結果輸出。無論在特徵擷取亦或病徵偵測部分,輸入信號皆需經過去除基線飄移和消除高頻雜訊的前置信號處理,而且兩部分所採用的方式必須一致,本論文使用中值濾波器和小波轉換達成。特徵擷取部分使用K-SVD演算法訓練稀疏表示中的辭典,病徵偵測部分使用OMP方法得到之係數達成,藉由稀疏表示的概念,比較原信號與先前得到的辭典乘以係數間的平均誤差,並選擇誤差最小的辭典所代表的病徵當成輸出,最後藉由比較判斷結果與MIT-BIH心律不整資料庫所提供的病徵註解,得到病徵診斷的正確數目與誤判數目。我們所提出的方法與傳統基於時域分析與頻域分析的方法相比較,可以有效降低演算法的複雜度,且可適用於多數病徵,在複合性病徵的處理上,亦不需要額外的運算資源。
The electrocardiogram (ECG) is a non-invasive method of measuring the electrical properties of the heart, and it is widely used for diagnosis of heart disease in practice. In this study we present a method for heart disease detection based on ECG using the K-mean with Singular Value Decomposition (K-SVD) algorithm. The method we present includes two parts: feature extraction and detection for various types of heart diseases. Conventional wavelet on signal processing is used to combine with the K-SVD algorithm to perform the feature extraction and detection. In our study, the MIT-BIH arrhythmia database as a benchmark to test and verify our proposed method. In this study, we also performed pre-processing on ECG signals to achieve high accuracy on heart disease pattern detection. The pre-processing performed on ECG signals include baseline drift removal, 60-Hz interference de-nosing, median filter and wavelet transform. The K-SVD algorithm is used to construct an over-complete dictionaries and the OMP method is used to compute the coefficients with the dictionaries we derived. The disease pattern can be tested by comparing the difference among the normal signals and those heart disease patterns. Finally, we evaluated the performance in terms of the false positive and false alarm on patterns in MIT-BIH heart disease patterns. We believed that once the feature dictionary is constructed, the K-SVD algorithm is a high potential algorithm which can be implemented in portable devices to achieve reasonable processing speed as well as high accuracy on heart disease detection. Compared to those conventional frequency-domain and time-domain based analysis methods, the proposed scheme can greatly reduce the complexity of online detection by training heart disease patterns offline. As such, the proposed scheme is very suitable to be used in practice. In addition, the proposed scheme can deal with composite diseases without additional computing resource.