Abstracts

Rationale: Epilepsy patients have a higher risk of traffic accidents due to seizures while driving. Interictal epileptiform activity (IEA) cannot be seen by an observer and is not realized by patients. IEA can lead to transient cognitive impairment but its impact on driving is unclear. A tight temporal correlation between IEA and the task to be performed is needed for optimal analysis of its impact. We automated real-time IEA detection,1whose sensitivity and specificity we aim to improve with easy to use deep learning frameworks. Representing time series as images enables the use of techniques from computer vision. We use a pretrained residual neural network (ResNet), which utilizes skip-connections building on constructs known from pyramidal cells in the cerebral cortex, for real-time EEG time series classification.  Methods: Model training of the pretrained ResNet was done using raw EEG from our own dataset consisting of presurgical phase-1 telemetry surface recordings of 1100 h in total, verified to contain frequent IEA. The EEG was preprocessed in the time domain and then encoded as recurrence plots, a Gramian Angular Field (GAF) and Markov Transition Field (MTF).2 GAF and MTF are 2-dimensional representations of that time series and preserve temporal dependency. For the real-time classification task, the EEG was cut into frames and then processed in the same way as training data before fed into the prediction network.  Results: As a proof of principle, we tested our algorithm with 50 pre-annotated IEA images and 62 images of normal EEG, corresponding to periods of 500ms duration prior to annotated IEA. 70% of data were taken for training and 30% for validation. Images were generated for classification (Figure 1). Classification sensitivity was 100% and specificity was 90%. Total time for IEA detection was a mean 318.1 ms, consisting of 195.3 ms for sampling 100 EEG datapoints with 50% overlap, 75.4 ms for conversion into MTF images, and 47.4 ms for image classification. Conclusions: Our model's performance was close to state-of-the-art results and had acceptable processing time. Transformed images from running EEG recordings can now be fed into this model. Next, we will expand the training data. Deep learning techniques from computer vision for time series classification show a promising path. Moreover, transfer learning can help to achieve new results in a very short period of time and with minimal effort.References[1] von Allmen A, Krestel H. Method for automated detection of epileptic and atypic epileptiform discharges in the EEG and triggering a stimulus in real time. Patent No 709839.[2] Wang Z, Oates T. Imaging Time-Series to Improve Classification and Imputation. arXiv:1506.00327v1 2015. Figure 1.Conversion of EEG windows into Markov Transition fields for real-time image classification.Row A) as example for normal EEG. Row B) as example for IEA. Funding: This project has received funding from the European Union’s Framework Programme for Research and Innovation Horizon 2020 (2014-2020) under the Marie Sklodowska-Curie Grant Agreement No. 799791.