A Deep Learning Technique to Clinch the Detection of Parkinson’s Disease using Speech and Voice Attributes

Abstract

Among the neurodegenerative diseases Parkinson’s Disease ranks second only to Alzheimer’s disease. Though extensive research is carried out in this area there have been no biomarker suggested. At present the diagnosis and monitoring of the disease progression is possible only through clinical examination and function symptoms observation. Voice impairment has been identified as an early marker for Parkinson’s Disease and hence the research in this field is gaining popularity. Machine Learning algorithms have proved useful in analyzing the enormous data with high dimensionality. But this has not been successful in extricating features that will have a strong correlation in predicting the disease accurately. This calls for a more effective and powerful technique like Deep Learning that uses deep neural networks that can select the optimal features and can contribute in the identification of the disease. In this paper an initial step was made by designing an Artificial Neural Network model. This yielded a train and test accuracy more than ninety-nine percentage and seventy-five percentage respectively for classifying the disease but showed overfitting problem which resulted in a decrease in the performance. Hence, the Artificial Neural Network model was hyper-tuned to reduce this problem and there was a slight improvement in the performance. Two methods were employed for optimization – a regularization method early stop and another validation method called Stratified K -Fold Cross Validation. Among these the second approach showed better results by slightly reducing the overfitting issue and it yielded a train and test accuracy score of approximately ninety-nine percentage and ninety-seven percentage with K-fold as five and Stochastic Gradient Descent as the optimizer. Even though the results were promising it was unable to unravel the prime attributes that would eventually identify the disease.