20 likes | 149 Views
心雜音之自動分析模式建立. 中文摘要 心雜音是臨床診斷上的重要參考,可對心臟結構或病理問題做早期診斷,而今,心臟聽診仍是確認心雜音的重要技術,且需具備相當經驗,因此,我們將評估聲音訊號處理及統計方法在心雜音分析上的應用。 在心音訊號的蒐集上,本研究共蒐集了 31 個有心雜音的樣本 ( 年齡從 42 到 89 歲;平均年齡 71.1 歲 ) 及 32 個正常樣本 ( 年齡從 19 到 36 歲;平均年齡 22.3 歲 ) ,透過對收錄訊號的處理與分析,及使用兩群體之 Student’s T 檢定後,發現有心雜音的群體及正常心音群體間,在某些頻率區間其能量比值有顯著差異。
E N D
心雜音之自動分析模式建立 • 中文摘要 • 心雜音是臨床診斷上的重要參考,可對心臟結構或病理問題做早期診斷,而今,心臟聽診仍是確認心雜音的重要技術,且需具備相當經驗,因此,我們將評估聲音訊號處理及統計方法在心雜音分析上的應用。 • 在心音訊號的蒐集上,本研究共蒐集了31個有心雜音的樣本(年齡從42到89歲;平均年齡71.1歲)及32個正常樣本(年齡從19到36歲;平均年齡22.3歲),透過對收錄訊號的處理與分析,及使用兩群體之Student’s T檢定後,發現有心雜音的群體及正常心音群體間,在某些頻率區間其能量比值有顯著差異。 • 接著,我們對每一頻率區間之能量比值予以排序,並取兩相鄰樣本其值的平均為候選鑑別值(threshold),依醫學決策中ROC(Receiver Operating Characteristic) Curve的定義繪製出ROC 曲線,結果發現雖然在部份頻率區間其最佳鑑別值的TPR(True-Positive Rate)值可達80%以上,但此時其FPR(False-Positive Rate)值卻超過30%,因此我們另引用類神經網路(Artificial Neural Networks)的方法以改善效能。 • STATISTICA Neural Networks是一套綜合且快速的類神經網路分析套裝軟體,我們用其Intelligent Problem Solver功能以找出合適的類神經網路架構,先使用48個樣本來訓練與建立類神經網路模組,然後再利用其餘的15個樣本來驗證該模組,經多方的試算與分析,我們得到的最佳模組為多層感知器(Multilayer Perceptrons; MLP)架構:輸入變數40個(所有頻率區間能量比)、隱藏層具14個神經元及1個輸出變數,對15個測試樣本驗證結果中,正確率為100% (TPR為100%;FPR為0%)。 • 在本研究中,我們已對所蒐集的樣本建立了自動分析模式,希望將來能在此基礎上,隨著樣本的擴增及技術的精進,發展出判別心雜音的自動系統。
Establishing an Automatic Analysis Model for Cardiac Murmurs • 英文摘要 • Cardiac murmur is significant in clinical diagnosis, and can be used to diagnose the physiological problems (physiology program) of heart earlier. (Up to now, auscultation of the heart is still a critical and definite skill required to identify and diagnose these murmurs). Auscultation is still the key method to identify cardiac murmurs but relies much on experience. Therefore, we evaluated the application of acoustic signal processing technique and statistical methods for cardiac murmur analysis. • Thirty-one cases (aged 42 to 89 years; mean 71.1 years) who were diagnosed to have cardiac murmurs and 32 cases (aged 19 to 36 years; mean 22.3 years) were included in this study. The recorded data have been processed and analyzed. Using the Student’s t-test, we found that there were significant differences between the normal and abnormal cases at specific frequency intervals. • Furthermore, we sorted the ratio of energy for every frequency interval, and picked every mean value of the neighboring pairs for candidate threshold. By the definition of Receiver Operating Characteristic (ROC) curves in medicine decision, we got the ROC curves and found that the true positive rate was more then 80% at some specific frequency intervals, but the false positive rate was above 30%. We expected to improve the performance more by using Artificial Neural Networks. • STATISTICA Neural Networks is a comprehensive and rapid neural network data analysis package. We use its Intelligent Problem Solver to get an initial idea of right neural network architecture. We picked and built networks, trained them in 48 cases, and tested them in another 15 cases later. After many times of computing, we got a Multilayer Perceptorns model, which used 40 input parameters, 14 neural units in hidden layer and 1 output parameter. That model was 100% correct for 15 study cases. (The true positive rate was 100% and false positive rate was 0%.) • In this study, we demonstrated an automatic analysis model for collected cases. Base on this foundation, we plan to develop an automatic determine system of cardiac murmurs in the future.