Abstracts

Rationale: The diagnosis of epilepsy is challenging, with a ~20% misdiagnosis rate. While the presence of interictal epileptiform discharges (IEDs) on an electroencephalogram (EEG) can support the diagnosis in the right clinical context, the EEG is limited by poor sensitivity and a risk of overinterpretation. Deep learning (DL) has shown promise to model high-dimensional EEG data. Our objective is to use DL to automatically detect epilepsy on EEG.

Methods: We selected a retrospective cohort of consecutive patients who underwent an EEG at the University of Montreal Hospital Center (train/validation set: Jan 2018–July 2019; test set: July–Sep 2019, with no overlap of patients). We reviewed their clinical records longitudinally for the diagnosis of epilepsy until the end of the available follow up period, as per their treating physician’s notes. We trained DL models to classify EEGs according to the diagnosis of epilepsy: a deep convolutional neural network (CNN) based on the ConvNeXt model, and a Vision Transformer (ViT) with two different tokenizers (linear and convolutional). We tested different configurations for each model (small, large, and huge). EEGs were segmented into 10s or 30s segments, and input directly into the models. The predictions were aggregated at the EEG level using the median of the predicted probabilities. We also investigated the addition of random data augmentations during training. We compared our performances in terms of area under the ROC curve to two previously used methods: the ShallowConvNet (Schirrmeister et al. Hum. Brain Mapp. 2017 [38], 5391-5420) and extraction of linear and nonlinear features input into a boosted trees model (Lemoine et al. Sci Reports 2023 [13], 12650–12666). The train/validation set was used to select hyperparameters (80/20% random split). The temporally shifted testing cohort was used only to evaluate the final performances and compute statistical intervals.

Results: The train/validation set contained 820 EEGs from 728 patients. The temporally shifted testing cohort included 128 EEGs from 118 patients, none of whom were in the train/validation set. In the testing cohort, 72 patients had epilepsy (61%). The median follow-up times in each group were 115 weeks (IQR: 52–151) and 89 weeks (48–121), respectively. The large ViT with convolutional tokenizer had the highest performance on the test set, with an AUROC of 0.75 (95% CI: 0.68–0.83). All DL models had above-chance performance and exceeded previous methods (Figure 1). The model performed best with 30s segments, and data augmentation improved most DL models. In the subgroup of patients whose EEG did not capture IEDs (n = 110), performances were similar (AUROC: 0.76 [95%CI: 0.66–0.84]).

Conclusions: This work suggests that DL could be used to uncover biomarkers of epilepsy on EEG without relying on IEDs, increasing its diagnostic yield. It also provides practical information on the adaptability of DL models to routine EEG. Further work is needed to explore the representation learned from the EEGs by the DL models.

Funding: This work is supported by the Canadian Institutes of Health Research, IVADO, the Brain Canada Foundation, and the Fonds de Recherche Santé-Québec.