Transformers-Based Deep Learning for Parkinson’s Disease Detection Using Electroencephalography and Stockwell Transform

Neurodegenerative conditions such as Parkinson’s disease (PD) have, in recent times, seen a major boost to their understanding due to the use of deep learning (DL) in the analysis and diagnosis process. DL, in particular, has been effective in unearthing many complex and intricate concepts of the EEG signal. This research is an attempt to understand the impact of the state-of-the-art DL technology, the transformer, in the diagnosis of PD from electroencephalography (EEG) signals. The time-domain EEG signals are transformed into time-frequency domain using the Stockwell transform for better insights. This transformation produces an image dataset, which can now be analyzed intricately for time and frequency features. These images are then fed to three transformer models tailored specifically for PD detection based on the response of the model. The classical vision transformer (ViT), the compact convolution transformer (CCT), and the swin transformer are selected for the purpose. The hyperparameters of each of the models are tuned by evaluating the various performance parameters at every iteration to design the most efficient models. The performance of the models is compared to each other as well as to other recent research using similar technology. The publicly available San Diego dataset is used for the analysis, and the ViT displays a commendable accuracy of 94% with the loss closing in towards zero. The CCT also performs well with fewer computational requirements.