Émile Lemoine1, Denahin Hinnoutondji Toffa1, Frédéric Lesage2, Dang Nguyen3, Elie Bou Assi1
1Neuroscience, University of Montreal, 2Biomedical Engineering, École Polytechnique de Montréal, 3Centre Hospitalier de l'Université de Montréal
Objective:
We aim to evaluate the performance of computational markers to predict seizure freedom after routine EEG.
Background:
The presence of interictal epileptiform discharges (IEDs) on routine EEG predicts seizure recurrence. However, 29–55% of routine EEGs do not show IEDs in patients with epilepsy (PWE), limiting its prognostic yield.
Design/Methods:
We selected all patients undergoing 30-minute EEG at the CHUM in 2018. Exclusion criteria were absence of follow-up after EEG, unclear epilepsy diagnosis at last follow-up, and electrical seizure(s) on EEG. Medical charts were reviewed for seizure freedom after EEG. EEGs were segmented into non-overlapping 10s windows. Several features were extracted for each electrode and each time window, each feature capturing a distinct linear or non-linear property of the EEG signal. The extracted features were: band power, peak alpha frequency, Hurst exponent, entropy (ApEn, SampEn, MsEn, SpecEn, FuzzEn, and PermEn), line length, and correlation dimension. We trained a regularized boosted trees machine learning model on multichannel EEG segments to predict seizure freedom during follow-up. We averaged predictions over windows to obtain one prediction per EEG. We tuned hyperparameters with cross-validation on a validation ensemble (80% of data) and evaluated the model on a testing set (20% of data).
Results:
656 EEGs from 534 subjects were included; 345 (65%) were from PWE. 527 EEGs (80%) did not capture IEDs. Median follow-up was 92 weeks (IQR 40–126). The classifier, using all features as input, could predict seizure freedom with above chance performance (area under the receiver-operating characteristic curve [AUC ROC] 0.68, 95%CI 0.65–0.72), even in the subset of patients with no IEDs on EEG (0.66, 0.61–0.70). Among the features, band power showed the best performance, followed by FuzzEn and line length.
Conclusions:
A computational biomarker could potentially increase the prognostic yield of routine EEG in the clinical setting.