Authors :
Presenting Author: Wen Shi, PhD – Massachusetts General Hospital; Harvard Medical School
Dana Shaw, / – Boston University; Katherine Walsh, / – Mass General; Dhinakaran Chinappen, / – Boston University; Mass General; Xue Han, PhD – Boston University; Uri Eden, PhD – Boston University; Robert Richardson, MD – Mass General; Harvard Medical School; Julia Jacobs, PhD – University of Freiburg; Benjamin Brinkmann, PhD – Mayo Clinic; Gregory Worrell, PhD – Mayo Clinic; William Stacey, PhD – University of Michigan; Mark Kramer, PhD – Boston University; Catherine Chu, MD – Mass General; Harvard Medical School
Rationale:
Accurate identification of the epileptogenic zone (EZ) is crucial for successful neurosurgical intervention in drug-refractory epilepsy. We recently found that automatically detected spike ripples (SR) had improved specificity for the EZ over other interictal biomarkers in a multicenter retrospective dataset (Shi et al., medRxiv, 2023). Here we sought to develop and test a prospective tool to estimate whether a channel is in the EZ using automatically detected SR rates from interictal intracranial recordings.
Methods:
To train the method, we utilized an international dataset of 42 subjects from four epilepsy centers (21F, mean age 31, range 8-65yrs) with interictal intracranial recordings and ILAE 1 outcomes with > 1 year follow up. For each subject, bipolar channels were categorized as resected or not-resected based on co-registration with post-operative MRI/CT images, where resected channels were designated as within the EZ and not-resected channels as outside of the EZ. Spike ripple rates, spike rates, and high frequency oscillations (HFO) rates were computed for each channel using validated detectors. Using absolute rate thresholds across subjects, the area under the curve (AUC) was calculated for the receiver operator characteristic (ROC) curve for each biomarker to evaluate prediction accuracy. To assess classifier performance, we performed leave-one-subject-out cross-validation (LOOCV). To do so, the rate thresholds yielding the highest F1 score were determined from all subjects but one and applied to the left-out subject, and performance was computed for the left-out subject. To validate the classifier’s performance, we applied the optimal threshold from the entire training data to 10 naïve subjects (2hr recording from 1-3AM per subject) from a fifth center with ILAE 1 outcome at > 1 year (4F, mean age 33, range 8-56yrs). The channel-wise accuracy, sensitivity, and specificity of the classifier were computed across and within subjects. Channel-wise probabilities of being in the EZ were also computed from SR rate and data duration using empirical distributions from the training data. Results:
Spike ripples outperformed spikes and HFOs in classifying channels in the EZ (AUC 0.75, 0.65, and 0.69, respectively, Figure 1). Our tool provides by-channel probabilities estimates of a channel being in the EZ and classification. Accuracy statistics for example probability estimates of a channel being in the EZ based on training data are provided (Table 1A). On LOOCV analysis across the 42 subjects, channel-wise classification based on SR rate had an accuracy =79%, with a specificity =83% and sensitivity =57%. When applied to a naïve dataset of 10 subjects, the absolute threshold (SR rate >= 7.0/hr) had an average accuracy =77%, a specificity =77% and sensitivity =70% to accurately classify channels in the EZ (Table 1B).
Conclusions:
We introduce an automated tool to prospectively predict which intracranial channels are in EZ that performs with good sensitivity and specificity. Future work will include increasing the testing dataset to further evaluate this new clinical tool to aid in accurate identification of the EZ in patients undergoing surgical evaluation.
Funding:
R01NS119483, R01NS110669