- 學位論文
中醫聞診結合問診專家系統於慢性腎衰竭之應用
An Expert System with Inquiring Diagnosis and Auscultical Diagnosis of Chronic Renal Failure in Chinese Medicine
摘要
本研究主要目的是建立一套以聞、問二診資料合參之中醫慢性腎衰竭診斷專家系統。 在臨床資料的收集方面，於台北市立中醫醫院之腎臟科門診中，利用語音來量化病情。以母音/ a /、/ i /、/ u /、/ e /、/ o /為實驗語音樣本，研究分析慢性腎衰竭病患與健康人之語音樣本差異性。分別以音高，越零率及共振峰等語音參數作基本的分析，將得到的數值作統計，嘗試找出量化的依據。加上專門針對慢性腎衰竭症狀設計之量表問卷，由醫師或經訓練的研究助理，診斷時對病人進行症狀勾選；以便攜式迷你光碟錄放機（Mini-Disc, MD）收取病患語音；共收集慢性腎衰竭病患50人的語音及其症狀量表問卷，再進行類神經網路訓練。病人語音經過「中醫聞診診斷系統」之分析後，將數值與正常人作統計分析雙尾獨立T檢定。男性音高部分在五個母音P值皆小於0.05，有顯著差異；女性音高部分則只有在/ u /具有顯著差異，結果不如男性顯著。男性越零率部分在五個母音並不具統計上的差異；女性越零率部分在/ a /、/ e /兩個母音部分具顯著差異。在男性與女性共振峰部分，亦有部分差異存在。在肌酸酐與語音參數、五個母音的迴歸分析部分，相關係數R2值皆不大，可見其相關性不大。 在診斷專家系統方面，採用誤差倒傳遞（Error Back-Propagation）類神經網路的學習法則來建立系統模型。診斷系統網路共分兩層，輸入層為音高、越零率、共振峰等語音特徵數值及問卷勾選之症狀，輸出層為證型，共有5種。共蒐集病患50名作為訓練樣本，訓練5000次經兩次權值篩選後，結果顯示在診斷方面的準確率到達83.33%。網路完成後，選取權值較高之症狀39種進行規則萃取的步驟，並繪出診斷樹。
並列摘要
The purpose of this research was to establish an expert system of inquiring diagnosis and auscutical diagnosis of chronic renal failure patients in Chinese Medicine. The expert system includes Database, Server, and Client with self-learning on WWW. One of the purposes in this research was to determine the auscultical differences between chronic renal failure patients and healthy subjects. Clinical data from the Department of internal medicine in Taipei City Hospital of Traditional Chinese Medicine were collected. The patient’s condition according to vocals / a /, / i /, / u /, / e /, and / o / was evaluated. The speech parameters of this research are pitch, zero-crossing rate and formant. A mini-disc (MD) was used to record the voice from 29 healthy people and 50 patients. The “Auscutical Diagnosis System in Traditional Chinese Medicine” was used to analyze the data. The results showed that there were differences of auscultical parameters between healthy people and patients. On the other hand, this expert system was established by Error Back-Propagation neural network. The diagnostic network was separated into two layers. The input layer includes 90 symptoms, Creatinine value, and 20 auscultical parameters. There are 5 syndromes in the output layer. The 50 patients’ data were used as the training samples to train the neural network. After 5000 times of training, the results showed that the accuracy was 83.33%. Thirty-nine symptoms that had higher weights were extracted for “Rule Extraction” to setup the diagnostic trees.