Classifying Emotions of Parkinsonian Patients from Electroencephalogram Signals Using Efficient Attention Capsule Network

Detecting Parkinson’s disease (PD) and other neurodegenerative disorders holds significant importance for early and automated intervention using non-invasive modalities like EEG. Although PD detection using the Electroencephalogram (EEG) signal has been done previously, identifying subtle differences in a wider spectrum of EEG signals from persons with PD and neurotypical (HC) individuals remains an open and challenging problem. PD patients often exhibit emotional dysregulation; identifying this is vital for their appropriate treatment. Though complex neural networks, such as Capsule Networks (CapsNet) and graph convolutional neural networks, have been applied to do this task, they are constrained by available computational resources. To address this issue, we proposed an efficient Capsule Network that leverages dynamic convolutional feature extraction and self-attention to mitigate CapsNet complexity in PD classification. By incorporating parallel fully connected neurons with CapsNet, regularisation and normalisation performance both in terms of types of predictions and on testing set is increased in the model. In this paper, binary classification of PD vs HC and categorical classification of emotions using machine learning techniques are explored. The proposed model archives 98.71% test accuracy in PD vs HC classification and 92.35% test accuracy in PD emotion classification. This method can serve the purpose of easy identification of emotions in persons with PD for better management of their day-to-day lives.