Authors :
Presenting Author: Varun Reddy Bajepally Chandra Varun Reddy, Meng – University of Missouri
Stephanie Bustros, MD – University of Missouri
Patrick Hackett, BS – University of Missouri
Sudeepa Bhattacharyya, PhD – University of Missouri-Kansas City
Kousalya Velagapudi, MS – University of Missouri
Komal Ashraf, DO – University of Missouri
Sean Lanigar, MD – University of Missouri
Tyler Davis, MD, PhD – University of Utah
Elliot Smith, PhD – University of Utah
John Rolston, MD, PhD – Brigham and Women's hospital, Harvard Medical School
Bornali Kundu, MD, PhD – University of Missouri
Rationale: Patients with drug-resistant focal epilepsy often require invasive monitoring to localize the seizure onset zone (SOZ). Our lab has previously demonstrated that Single-pulse electrical stimulation (SPES) holds promise in aiding SOZ localization (Kundu et al). Here we applied a machine learning (ML) approach to SPES data to predict the SOZ location. Prior studies include a 1d-CNN applied to stereo-EEG(SEEG) data from 10 patients (78.1 % sensitivity; 74.6 % specificity, Johnson et al) and a 1D-CNN-Transformer model on ECoG data from 35 patients (AUROC = 0.730, Norris et al).
Methods: Single pulses of stimulation were administered to a subgroup of electrodes, including those within the clinically identified SOZ. Evoked responses (10-100 ms post stimulation) were recorded from 16 patients (7 Utah, 9 Missouri). Four static descriptors were extracted for each stimulation channel: center (Hospital), channel ID, brain region, and mean root mean square (RMS) amplitude. To analyze this data, we developed a hybrid deep learning model. A Residual CNN–Transformer was used to process the time series data, while a small fully connected neural network (multi-layer perceptron, or MLP) was used to incorporate static inputs. The model was trained and evaluated using a patient-stratified 5 × 4-fold cross-validation. The same architecture was retrained on a larger combined dataset incorporating Utah, Missouri, and the publicly available BIDS formatted data yielding a total of 51 patients.
Results:
In the 16-patient SEEG cohort, the model achieved a mean sensitivity = 0.70, specificity = 0.75, and AUC = 0.75.
In the 35-patient ECoG cohort, the model achieved a sensitivity = 0.63, specificity =0.61, AUC = 0.63,
In the full 51 patient dataset: the model achieved mean specificity = 0.62, sensitivity = 0.67, and AUC = 0.66.
Conclusions: A transformer-based ML model applied to SPES-evoked responses shows moderate accuracy in identifying the SOZ. While these results are promising, generalizability to other centers and i-EEG data types remain challenging. More patient data is likely needed to build such a model. Static features defining the underlying connectivity likely will help improve performance.
Funding: None