Optimized Deep Learning for the Classification of Parkinson's Disease Based on Voice Features

Abstract

Parkinson's disease (PD) is a neurodegenerative disorder. Hence, there is a tremendous demand for adapting vocal features to determine PD in an earlier stage. This paper devises a technique to diagnose PD using voice signals. Initially, the voice signals are considered an input. The signal is fed to pre-processing wherein the filtering is adapted to remove noise. Thereafter, feature extraction is done that includes fluctuation index, spectral flux, spectral centroid, Mel frequency Cepstral coefficient (MFCC), spectral spread, tonal power ratio, spectral kurtosis and the proposed Exponential delta-Amplitude modulation signal (delta-AMS). Here, exponential delta-amplitude modulation spectrogram (Exponential-delta AMS) is devised by combining delta-amplitude modulation spectrogram (delta-AMS) and exponential weighted moving average (EWMA). The feature selection is done considering the extracted features using the proposed squirrel search water algorithm (SSWA), which is devised by combining Squirrel search algorithm (SSA) and water cycle algorithm (WCA). The fitness is newly devised considering Canberra distance. Finally, selected features are fed to attention-based long short-term memory (attention-based LSTM) in order to identify the existence of PD. Here, the training of attention-based LSTM is performed with developed SSWA. The proposed SSWA-based attention-based LSTM offered enhanced performance with 92.5% accuracy, 95.4% sensitivity and 91.4% specificity.