“Electroencephalograms are a means to detect pathologies such as neurological disorders or tumors. However, determining whether EEG recordings are normal or abnormal is a difficult and subjective task. So what if Artificial Intelligence could help neurologists decode EEG efficiently?
In this context, the Temple University Hospital (TUH) built a public database with 30,000 clinical EEG recordings among which 3,000 recordings were labeled either “pathological” or “normal”. This dataset constituted the TUH Abnormal EEG Corpus. The initiative was followed by an effort to find an efficient automatic interpretation of pathological EEG recordings. A TUH thesis student explored several methods to classify this data: non-parametric models, Hidden Markov Chains and Neural Networks.
This methodological project focuses on the research work of R. Schirrmeister, K. Eggensperger, F. Hutter, L. Gemein, and T. Ball through their article Deep Learning with convolutional neural networks for decoding and visualization of EEG pathology. This paper, published in January 2018, presents the results the authors got using their own Deep Convolutional Neural Network (ConvNet) architecture to distinguish whether an EEG recording was pathological or not. They based their analysis on the TUH Abnormal EEG Corpus (an ongoing data collection effort), and they claimed to have the best results that had been published on this particular issue and dataset so far.
In this report, we shall go through the article of interest point by point in order to highlight its methods, breakthroughs and main conclusions. We will also put it into perspective and anchor it into context by evaluating related works.
The analysis leads to wonder: what are the particularities of the ConvNets’ architectures presented in this article? Does the paper actually submit the best forecasting results as the authors wrote?
To answer these questions, we have in the first place, presented the dataset. Then, we studied how the data is processed and introduced the methods laid out in this paper and what particular choices the authors have made against the general architecture of ConvNets. After presenting the results, the next step led us to bring to light the coherence the article has with its related work. Finally, we concluded on the model’s performance and we will put it into perspective”.