Abstracts

Rationale: 30-50% of patients who undergo brain resection for drug resistant epilepsy (DRE) fail to achieve sustained seizure freedom. Accurate outcome prediction tools are necessary to address this patient selection problem. Predictive models built on scalp EEG could be valuable as this technology is inexpensive, non-invasive, and already ubiquitous in routine care. Based on results from animal studies and implanted EEG in humans, we hypothesized that “peri-ictal” EEG (i.e. minutes immediately before and after a seizure) may be optimal for creating an EEG-based surgical outcome prediction framework.


Methods: We analyzed data from 294 patients who had undergone temporal lobe resection for DRE and been followed longitudinally. All patients had undergone preoperative inpatient video EEG evaluation with witnessed seizures. We captured 5 minutes of “peri-ictal” EEG from a single pre-operative seizure for each patient. A validated machine-learning (ML) artifact detector was applied to raw EEG data. Artifact-free data were then transformed into the frequency domain (Fourier transformation) and power-spectral density (PSD) was calculated across a range of physiological frequencies (1-40Hz). Iterative ML model-building experiments were conducted to find classifier algorithms to predict post-operative seizure control based on pre-operative peri-ictal EEG features (Figure A).


Results: Of the 294 patients, 170 were seizure-free at last follow-up while 124 had recurrent seizures. Average age was 37.3 years, mean follow-up was 3.4 years, and 78% had abnormal MRI. There were no statistically significant differences between outcome groups in terms of age, sex, duration of epilepsy, baseline seizure frequency, etiology of seizures, presence of non-localizable seizures, or interictal discharges on EEG. We were able to build multiple machine learning models that could predict seizure outcome on the basis of peri-ictal EEG features (Figure B), underscoring the value of peri-ictal data in this context. The winning model was a Light Gradient Boost Machine (LGBM) classifier; when tested on a hold-out dataset which the model had not previously encountered, accuracy was 91.9% (AUC 0.977). The winning model was constructed with 34 EEG features (each ‘feature’ was the normalized PSD captured by an electrode at a specific frequency e.g. FT10-25Hz). 70% of features were from the pre-ictal period and 30% were from post-ictal period. 62% of features were from temporal electrodes while 38% were from extra-temporal electrodes. 56% of features were from the beta band.


Conclusions: There is a persistent need for accurate outcome prediction models for epilepsy surgery. Prior efforts have relied on simple clinical variables with limited accuracy or on data inputs that are invasive (intra-cranial EEG) and absent in routine clinical care (connectomic data). We demonstrate that using machine learning methods, 5 minutes of peri-ictal scalp EEG can be used for postoperative seizure outcome prediction with accuracies >90%. Once validated on a larger multicenter dataset, this approach will be implementable for all temporal DRE patients using data that is captured in routine clinical care.


Funding: None