Voice-Based Detection of Parkinson’s Disease Using Empirical Mode Decomposition, IMFCC, MFCC, and Deep Learning

Parkinson’s disease, a chronic and progressive neurological disorder that mainly affects the elderly, has recently drawn research attention to its manifestation in speech disorders as early indicators. In this particular study, the use of features based on Empirical Mode Decomposition (EMD) was employed to effectively capture the distinct characteristics present in speech affected by this disease. To this end, a comparison between two types of features, namely intrinsic mode function cepstral coefficients (IMFCC) and Mel frequency cepstral coefficients (MFCC) with the deep learning algorithms ANN, LSTM and CNN, was proposed as a means of accurately representing the unique vocal attributes exhibited by people with Parkinson's disease. The performance of the proposed features was meticulously evaluated using a Sakar dataset comprising 18 non-Parkinson's subjects and 20 Parkinson's individuals. The results obtained demonstrate unequivocally that the application of MFCC features in conjunction with the LSTM classifier produces classification accuracy superior to the use of intrinsic cepstral coefficient features. This is demonstrated by a significant 18.18% increase in accuracy.