Deep Transfer Learning for Early Parkinson's Disease Detection

Abstract

The main reason for Parkinson's Disease (PD) is unspecified. No permanent cure is available for this disease. Only medication can mitigate its effect. At present PD can be diagnosed through gait characteristics, voice recording or by handwriting. These methods share the same pipelines to detect the infancy level of PD. But to detect the early stage of PD is very challenging. In our study we have used deep convolutional neural networks to detect early stages of PD through patients' handwriting images. To increase the performance, we have combined four datasets of PD handwriting images without the additional signals and used an ensemble method of transfer learning technique. High handwriting sample variability presents a difficulty that is tackled by the transfer learning approach. We have used accuracy, loss, precision, recall, AUC and F1 score as measure metrics to evaluate the models. Our approach shows that the proposed ensemble model shows 95.5% accuracy