Use of ANN, LSTM and CNN Classifiers for the New MSCC and BSCC Methods in the Detection of Parkinson's Disease by Voice Analysis

Abstract

Parkinson's disease (PD) is a neurodegenerative condition that impacts a significant global population. The timely and precise identification of PD plays a pivotal role in facilitating early intervention and the efficient management of the condition. Recently, speech analysis has emerged as a promising noninvasive technique for the detection of PD due to its accessibility and ability to reveal subtle vocal biomarkers associated with the disease. This research introduces an innovative approach utilizing Short-Time Fourier Transform (STFT) to generate spectrograms, specifically Bark Spectrogram Cepstral Coefficients (BSCC) and Mel Spectrogram Cepstral Coefficients (MSCC). These coefficients are compared with traditional and well-known coefficients, namely Mel-Frequency Cepstral Coefficients (MFCC) and Bark Frequency Cepstral Coefficients (BFCC). To extract the most effective coefficients for Parkinson's disease detection, three robust classification techniques-Long Short-Term Memory neural networks (LSTM), Convolutional Neural Networks (CNN), and Artificial Neural Networks (ANN)-are employed. As a result, the BSCC and MSCC algorithms achieve a maximum accuracy rate of 90%, surpassing the accuracy of the traditional MFCC and BFCC coefficients. Therefore, these newly proposed coefficients prove to be more precise in diagnosing Parkinson's disease compared to the conventional MFCC and BFCC coefficients.