Data-mining for selection of interictal electrocorticographic activities based on seizure outcome in neocortical epilepsy

Abstract

Introduction: We introduce a new data-mining method to select interictal pathologic activities based on the outcome of resective epilepsy surgery defined as the presence/absence of seizures in neocortical epilepsy (NE). Methods: We analyzed electrocorticographies from 39 patients with medically intractable NE. We separately analyzed 37 frequency-bins from 0.9 to 600 Hz to select the bands related to the seizure outcome. An automatic detector using amplitude-duration-number thresholds was used. The two different interictal electrocorticography datasets containing epileptiform activities were selected. In the first training dataset, the automatic detector was optimized to best differentiate the seizure-free group from the not-seizure-free-group based on the ranks of resection percentages of the activities detected using a genetic algorithm. We optimized in a patient group with 20 patients and validated optimized threshold in a different patient group with 19 patients to evaluate stability of results in a different patient group. Significant reproducibility was determined from expected numbers of significant results from the binomial distribution. The differences in the resection percentage of the detected activities between the seizure outcome groups (Dif-R) in the validation dataset were measured. Results: There were 16 seizure-free (41%) of 39 patients. The mean follow-up duration was 21 ± 11 months (13 to 44 months). In the validation dataset from the different 19 patient group, delta in 2.0 – 2.3 Hz were significantly reproducible. Low-frequency activities (LFAs) between 4.9 – 43 Hz including theta, alpha, beta and low-gamma were significantly reproducible. High-gamma in 62 and 75 Hz and high-frequency activities (HFAs) in 108 and 322 Hz were reproducibly related to seizure outcome. Dif-Rs in the different patient group was about mean 10 – 20 % in reproducible frequency-bins. In LFAs, the resection of detected activities were positively related with better seizure outcome as intended. However, high-gamma activities are paradoxically negatively related with seizure outcome. Conclusion: Using the presented method, in a different interictal segment validation, we achieved Dif-Rs that were higher than the best manual and automatic HFA detections described in the literature using only training dataset (17 to 27 %). In a different patient group validation, our results, 10 – 20 % Dif-Rs were comparable to literature analyzing only training dataset. A new method selecting pathologic activities based on seizure outcome can be potentially useful finding pathologic activities to be resected.