可攜式智慧型心音聽診系統

Abstract

本研究針對上一代無線聽診器進行效能提升，並且開發新一代的可攜式智慧型心音聽診系統。此系統包含一個無線聽診裝置與一套心音訊號處理軟體，主要目的為瓣膜性心臟病的自動診斷。無線聽診裝置以MICAz無線感測網路模組為核心，搭配其他自製的模組所構成，包括心音感測模組和雙模式(充電/執行)電源管理模組。相較於前一代裝置，本裝置增加了心音即時監聽功能，也提高了訊號處理電路之頻寬，並將無線傳輸模組之參數最佳化。量測到的心音訊號，可以透過ZigBee無線通訊協定傳輸到個人電腦進行處理。在後端診斷系統中，利用C++ Builder撰寫使用者介面，具有心音訊號時頻域之即時顯示、多種檔案格式之儲存與讀取等功能。此外我們蒐集了網路上的心音資料庫進行分析，在MATLAB程式環境下建立異常心音偵測及心音分類演算法，使用滑動視窗之正規化平均夏儂能量擷取出心音包絡線，並執行心動週期分離以自動化閾值將連續的心音訊號切割出數個單一週期，分割出之心動週期具有93.37%的正確率。接著將心音包絡線轉換至頻域，使用正常心音的頻率特徵建立一個比對樣板，透過樣板比對與相關係數運算來偵測異常心音。在以接受者操作特性曲線(receiver operating characteristic curve, ROC curve)分析兩種方法的整體辨識能力下，兩者差異並不大，都可以用來偵測異常心音，在ROC曲線上選擇一適當閾值當作診斷標準，兩種偵測方法的偵測率分別為82.46%與81.20%。另外，結合心音包絡線特徵和基於短時傅立葉轉換的時頻域統計特徵，透過多層感知器倒傳遞類神經網路進行正常心音、舒張期心雜音及收縮期心雜音分類，其正確率分別為96.67%、94.19%、95.56%。

The objective of this research is to improve the previously developed wireless stethoscope and to develop a newer portable intelligent auscultation system for heart sound signals. The system which includes a wireless auscultation device and a heart sound signal processing software is designed and implemented to diagnose valvular heart disease automatically. The auscultation device is based on a MICAz wireless sensor network module, a heart sound sensing module, and a power management module. For the new auscultation system, real-time heart sound auscultation is added, the bandwidth of signal processing is improved, and the parameters of wireless transmission are also optimized. Heart sound signals can be sent to personal computer via ZigBee communication protocol. The diagnosis system witch is developed with Borland C++ Builder provides some functions, such as displaying real-time heart sound signal in time-frequency domain, and saving and loading different file formats. Besides, two algorithms for abnormal heart sound detection and heart sound classification are constructed in MATLAB by analyzing heart sound samples collected from internet databases. Normalized average Shannon energy with sliding window is applied to extract heart sound envelope, and then the continuous heart sound signal is segmented to individual heart cycles by heart cycle segmentation based on auto-threshold with accuracy of 93.37%. After above procedures, those envelopes are transferred to frequency domain and compared with a pattern constructed from frequency feature of normal heart sounds by pattern matching and correlation coefficient operation for detecting abnormal heart sounds. In this study, receiver operating characteristic curve (ROC curve) is used to analyze the identification ability of two methods. The correlation coefficient operation and pattern matching are similar in the identification capability, and accuracies of abnormal detection are 82.46% and 81.20%, respectively, based on suitable threshold recommended by the ROC curve. Furthermore, the multiple layer perceptron back-propagation (MLP-BP) neural network is used as a heart sound classification algorithm. The input of this neural network is heart sound envelope feature which combines statistic information of time-frequency domain based on short-time Fourier transform (STFT). The neural network can classify normal heart sound, diastolic murmur, and systolic murmur, with accuracies of 96.67%, 94.19%, and 95.56%, respectively.