Early running the network on the testing data. We

Early detection of severity of a disease is conducive to
treating the patient sooner. This paper intended to verify the effectiveness of
the application of Deep Learning for predicting the severity of Parkinson’s
disease in a patient using his or her voice characteristics. The dataset used
was UCI’s Parkinson’s Telemonitoring Dataset, comprising of 16 attributes or
biomedical voice measurements with various range of values from 42 people with
early-stage Parkinson’s disease. It was first pre-processed by applying
normalisation. Then the segmentation of the normalised dataset was done to
create training dataset and testing dataset. Deep Neural Networks were trained
on the training data, and finally the accuracy of severity prediction was
obtained by running the network on the testing data. We were able to
successfully implement deep neural network in predicting the severity of
Parkinson’s disease, achieving an accuracy of 81.6 % and 62.7 % in the case of
motor-UPDRS and total-UPDRS scores respectively. In order to analyse the
dataset and make an attempt to understand the trend of these severity scores,
an analysis of the normalised dataset was performed on the basis of gender and
age of patients. The results indicate that accurate prediction of severity of
Parkinson’s disease can be done using deep learning. This implies that Deep
Learning can be used for severity prediction and medical analysis for other
diseases of similar types as well. Although we have used a dataset of 5875
instances, the accuracy of our approach can be further improved by implementing
it on a larger dataset, having more number of instances of each severity class.
Moreover, more number of patient attributes like- gait and handwriting
features- can be added to make the model more reliable. Also, more powerful
computing resources(i.e. GPUs with better processing capabilities) can be used
to improve the time complexity of our approach.