Spectral entropy variability of intraoperative electrocorticography predicts outcome after epilepsy surgery in people with focal cortical dysplasia.

OBJECTIVE

Epilepsy surgery in people with focal cortical dysplasia (FCD) requires accurate removal of all epileptogenic tissue, and outcome is difficult to predict. We explored whether spectral entropy, a fast computable electroencephalographic (EEG) feature, could estimate epileptic activity in intraoperative electrocorticography (ioECoG) and forecast postsurgical outcomes.

METHODS

We included people with FCD pathology who underwent ioECoG-assisted resective surgery. We analyzed ioECoG recordings acquired before and after resection. We computed spectral entropy across eight frequency bands (1-500, 1-4, 4-8, 8-12, 12-20, 20-80, 80-250, 250-500 Hz) in 2-s epochs during 1 min, and the mean and SD per electrode. The preresection features were the input for a random forest machine learning model to classify channels covering resected from nonresected tissue. Explainable artificial intelligence was used to select features that positively influenced the model predicting resected tissue. We then related these markers measured after the resection to postsurgical outcome using trend analysis (Jonckheere-Terpstra) across outcome groups Engel IA without medication, Engel IA-D, Engel II, Engel III, and Engel IV.

RESULTS

We analyzed ioECoG data from 37 patients (age range = 0-61 years, 21 patients were ≤16 years), including 2270 preresection and 1278 postresection channels. The random forest model discriminated between resected and nonresected cortex (validation fold area under the curve = .84, 95% confidence interval =.74-.95). Features with the strongest positive Shapley Additive Explanations values included high spectral entropy variability (SEV) in the 80-500-Hz bands. In postresection recordings, higher mean SEV and greater spatial variability of SEV were associated with poorer Engel outcome across several frequency bands. After multiple comparison correction, the positive relationship of high mean SEV (p = .004) and high spatial variability (p = .001) at 250-500 Hz with poor seizure outcome remained statistically significant.

SIGNIFICANCE: SEV is a real-time computable invasive EEG marker that represents persisting epileptic activity after resection. SEV may be used to evaluate resection adequacy directly or may reflect residual epileptic activity. This would enable epilepsy surgery guidance and postsurgical counseling and decision-making.