Comparative Evaluation of PD Detection Using Deep Learning on IMFCCs Extracted from VMD

Abstract

This paper presents a new method for extracting vocal features for the diagnosis of Parkinson’s disease (PD) via voice analysis applying variational mode decomposition (VMD). The classical method of extracting mel-frequency cepstral coefficients (MFCC) is compared to a new approach that generates coefficients named intrinsic mel-frequency cepstral coefficients (IMFCC). For this study, two audio databases were used: the SAKAR database containing 38 recordings and a PC-GITA database comprising 50 recordings. The signal preprocessing steps include frame segmentation, pre-emphasis, and filtering. The voice signal is then decomposed into intrinsic modes employing VMD. From these modes, the log-energy of specific components is calculated to extract the IMFCC. In this study, two types of classifiers were used: convolutional neural networks (CNN) and long short-term memory (LSTM). The results show that IMFCC provides a new perspective for representing vocal signals, capturing distinct features compared to classical MFCC. Notably, the IMFCC2 attained the highest accuracy of 100% adopting the CNN classifier. This approach could improve the performance of systems for identifying PD via voice analysis, offering a robust and complementary alternative to existing feature extraction methods.