Unveiling Parkinson's Disease Features from a Primate Model with Deep Neural Networks

Parkinson's Disease (PD) is a neurodegenerative disorder with increasing prevalence in the world population and is Characterised by motor and cognitive symptoms. Although cortical EEG readings from PD-affected humans have being commonly used to feed different machine learning frameworks, the directly affected areas are concentrated in a group of sub cortical nuclei and related areas, the so-called motor loop. As those areas may only be directly accessed through invasive procedures, such as Local Field Potential (LFP) measurements, most data collection must rely on animal models. To the best of our knowledge, no neural networks-based analysis centred on LFP data from the motor loop was reported so far. In this work, we trained and evaluated a set of deep neural networks on a dataset recorded from marmoset monkeys, with LFP readings from healthy and parkinsonian subjects. We analysed each trained neural network with respect to its inputs and representations from intermediate layers. CNN and ConvLSTM classifiers were applied, reaching accuracies up to 99.80%, as well as a CNN-based autoencoder, which has also shown to learn PD related representations. The results and analysis provided further insights and foster research on the correlates of Parkinson's Disease.