| 摘要: | 在近幾年的研究中，語音分析被認為可以客觀且有效的診斷帕金森氏症(Parkinson′s disease, PD)，然而語音分析工具大部分都須依靠特定儀器或電腦運作，這些設備不利於攜帶或移動，若採用行動裝置能有效的解決攜帶的問題，因此我們開發了一款語音分析的Android行動裝置軟體，並測試五種分類器，從中尋找合適的分類器對PD進行診斷。
在實驗設計使用了74位帕金森患者的語音與50位健康者的語音，這些語音樣本為連續母音/a/，在實驗中測試了聲學參數對PD的相關性，包含了19個多面向音聲分析系統(Multidimensional Voice Program, MDVP)參數、歸一化噪音能量(Normalized Noise Energy, NNE)、平滑倒頻譜的峰值(Cepstral Peak Prominence Smoothed, CPPS)、長時間平均頻譜(Long-Term Average Spectrum, LTAS)、梅爾倒頻譜係數(Mel Frequency Cepstral Coefficients, MFCC)和可調Q因子小波轉換(Tunable Q-Factor Wavelet Transform, TQWT)。
在過去使用TQWT診斷PD的研究中擁有432個參數，而當參數過於龐大時容易導致分類器過度擬合，因此須對TQWT進行降維，首先在實驗中我們測試Principal Component Analysis (PCA)、Linear Discriminant Analysis (LDA)和Hellinger Linear Discriminant Analysis (HLDA)對TQWT的降維能力，其中HLDA獲得最好效果且解決LDA無法調整參數的問題。
在分類器中，選擇了最近鄰居法(K Nearest Neighbor, KNN)、多層感知器(Multi-Layer perceptron, MLP)、支持向量機(Support Vector Machine, SVM)、梯度提升決策樹(Gradient Boosting Decision Tree, GBDT)和多類海靈格線性判斷決策樹(Multi-class Hellinger Linear Discriminant decision tree, MHLDT)。
共5組進行參數的比較，在實驗中將參數依照1)時域測量、2)噪音測量與3)MFCC分成3組，再加上4)全部的參數與5)海靈格距離(Hellinger distance, HD)挑選的10個參數，測試參數混和的效果。
在結果中顯示噪音測量與MFCC的參數各自在不同的分類器中表現優於時域測量，與使用HD挑選的參數都為噪音測量與MFCC的結果一致，結合選中參數的特性與過去研究的結果發現測量聲帶受損導致的氣聲能有效的診斷PD。
在分類器與參數的比較結果中，當使用SVM與HD所挑選的參數能獲得最高的準確度最高為97.5%，最終將選中的分類器與參數製作成Android 軟體，軟體中可以錄製語音並診斷PD。
;In recent years of research, voice analysis was believed to be objective and effective in the diagnosis of Parkinson′s disease (PD), but most voice analysis tools today still need to work with specialized equipment or computers, which are not convenient for carrying or moving. Therefore, using of mobile devices could effectively solve the problem of carrying.
In this study, we developed an Android app for mobile devices to perform voice analysis, and tested 5 distinct classifiers, from which to find a suitable classifier to diagnose PD.
In experimental design we used voice samples of 74 PD patients and 50 healthy speakers, and these voice samples were sustained vowels /a/. In the experiment, we tested the correlation between PD and various voice parameters, including 19 Multidimensional Voice Program (MDVP) parameters, Normalized Noise Energy (NNE), Cepstral Peak Prominence Smoothed (CPPS), Long-Term Average Spectrum (LTAS), Mel Frequency Cepstral Coefficients (MFCC) and Tunable Q-Factor Wavelet Transform (TQWT).
In the past studies, there are 432 parameters using TQWT to diagnose PD. If the number of parameters is high, it is easy to cause classifier overfitting, so TQWT has to be reduced in dimensionality. Two experiments were conducted in this study.
In the first experiment, we tested the dimensionality reduction techniques based on the performance of Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA) and Hellinger Linear Discriminant Analysis (HLDA) on TQWT, where HLDA performed optimally and resolved the parameter adjust issue for LDA.
The classifiers, K Nearest Neighbor (KNN), Multi-Layer perceptron (MLP), Support Vector Machine (SVM), Gradient Boosting Decision Tree (GBDT) and Multi-class Hellinger Linear Discriminant decision tree (MHLDT) were used to determine if the voice belonged to a PD patient.
A total of 5 groups of parameters a were compared, the parameters were divided into three groups according to 1) time-domain measurement, 2) noise measurement, and 3) MFCC to test the performance of different characteristics. In addition, 4) all the parameters and 5) 10 parameters selected by Hellinger distance (HD) were also used to test the performance of parameter mixing.
The results showed that the parameters of noise measurement and MFCC outperform those of time-domain measurement in different classifiers. The results are consistent with the parameters selected using HD for noise measurements and MFCC.
Combining the characteristics of the selected parameters and the results of previous studies, it was found that measuring the breathy voice caused by the abnormal vocal cord can effectively diagnose PD.
In the comparison of parameters and classifiers, the highest performance was observed using SVM and the 10 parameters selected by HD, and the accuracy was 97.5%.
Finally, the selected classifier and parameters were implemented as an Android app, which could record voice and diagnose PD. |
