Deep Feature Extraction From the Vocal Vectors Using Sparse Autoencoders for Parkinson’s Classification

Abstract

Parkinson’s disease (PD) is a chronic, neurodegenerative disorder that mainly affects the motor system of the body. The progression rate is very slow but, as the condition worsens it stimulates the non-motor characteristics, especially mood disorders, hallucinations, delusions, and other observable cognitive changes. Genetic and environmental factors greatly influence the risk of Parkinson’s development. The early diagnosis of PD improves the condition through proper treatment. During the initial stage of PD, vocal impairments become more common among the affected individuals. Advanced studies on vocal disorders give more assistance for precise PD detection. In this paper, the vocal features of PD affected individuals are analyzed with sophisticated computational models. Initially, the samples are pre-processed, as it contains more missing values. Then the predictor candidate subset is identified from the processed vocal features using the Adaptive Grey Wolf Optimization Algorithm, a meta-heuristic global search optimization technique. Furthermore, the latent representation of the candidate features is extracted through sparse autoencoders for effective discrimination between the PD affected and control cases. For classification, six supervised machine learning algorithms were employed. The proposed model is trained with the data, divided through 10-fold cross-validation scheme and the performance is evaluated based on validation metrics. The PD dataset is accessed from University of California (UCI), Irvine Machine Learning repository to conduct the experimental analysis. The result shows that the proposed algorithm outperformed the benchmarked models, exhibits its efficacy in distinguishing the affected and healthy samples of PD. The outcome of this study emphasizes the significance of intelligent learning models in complex disease diagnosis.