Convolutional neural networks to detect Parkinson’s disease based on voice recordings and transfer learning

An early detection of Parkinson’s Disease (PD) is important for patients’ quality of life and current detection methods are not optimal in this task. The search of an automatic, low-cost, non-invasive method would be appropiate to delevope a computer-aided diagnosis system. Machine learning models based on data such as neural networks are at the core of many of these systems. In this work 5 convolutional neural network architectures have been compared by their capability on classifying spectrograms generated with PCGita database. For this classification task, transfer learning using Saarbrücken Voice Database and a data augmentation technique have been addresed with 14 models from the 5 architectures. After performing 5-fold cross-validation, results show that VGG16 architecture is able to distinguish between healthy and PD spectrograms with around 85% of global accuracy. Further research is needed to explore the potential of this technology with multicondition training in medical environments.

I. Introduction

Parkinson’s disease is the second most common neurodegenerative disease that affects world population. Its patients are mainly over 50 years old and it diminishes their physical capabilities evoking muscular stiffness, slowness in movements and tremors among other symptoms. It is characterized by the loss of neurons in the part of the brain known as substantia nigra, but its causes are still unknown [1].