Application of Artificial Neural Network on Speech Signal Features for Parkinson’s Disease Classification

Parkinson’s Disease is a degenerative disorder of the central nervous system. Diagnosis is made based on patient history and physical exam findings. Since speech is affected in up to 90% of Parkinson’s Disease patients, it has drawn interest as a target symptom for diagnostic testing. Using a relatively large dataset of speech sample features compiled from Parkinson’s Disease patients and healthy individuals, a multilayer perceptron artificial neural network was trained to accurately classify between these two types of samples. By experimenting with the number of neurons in a single hidden layer network, a model was quickly found that could reach 0.937 accuracy. Using the single hidden layer results to inform a search for viable deep learning models, two hidden layer networks were found that could improve accuracy to 0.952. Finding high accuracy using a large sample size dataset indicates that multilayer perceptron is a viable method to aid in diagnosis of Parkinson’s Disease in the general population. This research demonstrates a potential search strategy for deep learning models that does not require the time and computational power of a comprehensive search of all possible models. The most accurate models found obtained 100% sensitivity. Thus, this method would be very useful as a screening test for Parkinson’s Disease as it is also non-invasive, quick, and can be done remotely.