基於機器學習分析帕金森氏症患者之語音;Voice Analysis of Patient with Parkinson’s Disease Based on Machine Learning Approach

NCU Institutional Repository > 資訊電機學院 > 電機工程研究所 > 博碩士論文 > Item 987654321/84236

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/84236

題名:	基於機器學習分析帕金森氏症患者之語音;Voice Analysis of Patient with Parkinson’s Disease Based on Machine Learning Approach
作者:	戴妮莎;Tandjung, Mirna Danisa
貢獻者:	電機工程學系
關鍵詞:	帕金森氏病（PD）;多維語音程序（MDVP）;機器學習（ML）;優化梯度提升;Parkinson′s disease (PD);Multidimensional Voice Program (MDVP);Machine Learning (ML);Optimized Gradient Boosted
日期:	2020-08-20
上傳時間:	2020-09-02 18:31:45 (UTC+8)
出版者:	國立中央大學
摘要:	語音分析是協助帕金森氏症(Parkinson’s disease, PD)患者早期發現疾病、鑑別診斷，以及監測疾病進程的重要工具。帕氏症患者會有獨特的運動言語障礙。多向度語音分析軟體(The Kay Multidimensional Voice Program, MDVP) 是一個自動化的語音分析系統，此系統可以快速計算達33種的語音功能測試，並以圖形化的形式呈現語音的功能，藉此可辨別在臨床上可能潛在的語音功能差異。MDVP分析中的錯誤主要的來源有兩種，分別是來自電腦程式執行計算的錯誤以及用戶在分離訊號時的錯誤。本研究共收集了55個健康者的語音樣本，55個帕氏症控制組語音樣本以及145個帕氏症患者語音樣本。語音樣本是一個持續三秒鐘的母音如/a/及/u/。在機器學習的分類技術中，本研究使用了最佳梯度提升 (Optimized Gradient Boosted) 模型來確認此分類機制的表現。共有五個參數，分別是APQ, PPQ, Jitta, ShdB, Jitt，其最佳的表現分別達到94.6%精確率(precision)、96.7%召回率、93%正確率，操作者特徵 (Receiver operating Characteristic, ROC) 曲線為87.9%。馬修斯相關係數 (Matthews Correlation Coefficient, MCC) 為83.1%。在此實驗中觀察到帕氏症患者的母音/a/，相較於母音/u/以及帕氏症控制組，母音/a/有較高的預測表現；另外健康者的母音/u/，相較於健康者的母音/a/，健康者的母音/u/有較高的預測表現。帕氏症患者及帕氏症控制組的母音/a/可以達到95.7%的精確率、95.7%的召回率、91.3%的準確率，以及100%的ROC曲線。健康者的母音/u/可達到95.6%的精確率、86%的召回率、91.3%的準確率，以及100%的ROC曲線。根據五個參數以及母音的結果，本研究可以有效的區分健康者及帕氏症患者的語音樣本。;Voice analysis of a patient with Parkinson′s disease (PD) could be an important tool in early detection, differential diagnosis, and monitoring of disease progression. Patients with PD develop distinctive motor speech disorders. The Kay Multidimensional Voice Program™ (MDVP) is an automatic voice analysis. This system rapidly calculates up to 33 measures of vocal function and displays them on a graph that incorporates normative values for the identification of potentially important clinical differences. Discrepancies in the MDVP analysis arise from two major sources: errors in the computations performed by the program and errors made by the user in an attempt to isolate a given portion of the signal. This study collected 55 healthy voice samples, 55 PD control samples, and 145 samples of PD patients. Voice samples of a 3-second sustained vowel sound /a/ and /u/. Among the machine learning classification technique, we used Optimized Gradient Boosted model to verify the performance of the classification mechanisms. The best performance with 5 parameters (APQ, PPQ, Jitta, ShdB, Jitt) achieved a precision 94.6%, recall (sensitivity) 96.7%, accuracy 93%, Receiver Operating Characteristic (ROC) curve 87.9%, and Matthews Correlation Coefficient (MCC) is 83.1%. It is observed vowel /a/ PD has a higher prediction than vowel /u/ PD and PD control, and vowel /u/ Healthy have a higher prediction than vowel /a/ Healthy. Vowel /a/ for PD and PD control voice achieved a precision 95.7%%, recall 95.7%, accuracy 91.3% and ROC curve 100%. The vowel /u/ of Healthy voice achieved a precision 95.6%, recall 86%, accuracy 91.3%, and ROC curve 100%. According to these 5 parameters and vowel results, we can efficiently differentiate between healthy and PD voice samples.
顯示於類別:	[電機工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	221	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....