Authors :
Presenting Author: Naoto Kuroda, MD – Tohoku University Graduate School of Medicine
Keiki Inoue, MD – NHO Sendai Medical Center; Hiroshi Uda, MD, PhD – Wayne State University; Yu Kitazawa, MD, PhD – Wayne State University; Kanako Tsuchimoto, MD student – Kagoshima University; Kotaro Saito, DMD, PhD – Kagoshima University; Riri Kimura, MD student – Yokohama City University; Risa Hikino, MD student – Yokohama City University; Ayame Yamagishi, MD student – Yokohama City University; Shin-ichiro Osawa, MD, PhD – Tohoku University Graduate School of Medicine; Hitoshi Nemoto, RT – Tohoku University Hospital; Makoto Ishida, PhD – Tohoku University Graduate School of Medicine; Kazushi Ukishiro, MD, PhD – Tohoku University Graduate School of Medicine; Carlos Makoto Miyauchi, PhD – Tohoku University Graduate School of Medicine; Eichi Takaya, PhD – Tohoku University Graduate School of Medicine; Shinya Sonobe, MD, PhD – Tohoku University Graduate School of Medicine; Keiya Iijima, MD, PhD – National Center of Neurology and Psychiatry; Yutaro Takayama, MD, PhD – Yokohama City University; Masaki Sonoda, MD, PhD – Yokohama City University; Sandeep Sood, MD – Wayne State University; Aimee Luat, MD – Wayne State University; Masaki Iwasaki, MD, PhD – National Center of Neurology and Psychiatry; Nobukazu Nakasato, MD, PhD – Tohoku University Graduate School of Medicine; Eishi Asano, MD, PhD – Wayne State University
Rationale:
The goal of intracranial EEG (iEEG) is to localize the epileptogenic zone, which leads to seizure control if removed completely. Promising interictal iEEG biomarkers include the rate of high-frequency oscillation (HFO) and the phase-amplitude coupling between HFO and delta wave. Given that nonepileptic areas consistently generate HFO at rates specific to different brain regions, we hypothesized that consideration of a deviation from the normative mean and a difference from adjacent sites or epochs could enhance the accuracy of HFO-related biomarkers to localize the epileptogenic zone. To test this hypothesis, we have developed an ensemble model that incorporates iEEG biomarkers beyond HFO rate and MRI data. We then assessed the model's accuracy in predicting postoperative seizure outcomes using a test dataset.
Methods:
The study included patients aged ≥4 years who underwent chronic iEEG monitoring followed by curative resection surgery at Wayne State University (WSU) in the United States, Tohoku University (TU) in Japan, and the National Center of Neurology and Psychiatry (NCNP) in Japan. The training dataset consisted of 135 consecutive WSU cases undergoing surgery until May 2018. The test dataset included nine cases from WSU since June 2018, thirty one from TU, and eight from NCNP. We developed a machine-learning ensemble model derived from 12 logistic regression models, each incorporating clinical, ictal iEEG, MRI, and one of the 12 interictal iEEG biomarkers. The considered interictal iEEG biomarkers included HFO rate and phase-amplitude coupling and their statistical deviations from the normative mean. For each of the iEEG biomarkers mentioned above, we calculated spatial volatility as the difference from adjacent sites, and temporal volatility as the difference from immediately neighboring epochs. The ensemble model weighted each of the 12 logistic regression models based on their accuracies of classifying the postoperative outcomes in the training dataset. Receiver operating characteristic (ROC) analysis determined the accuracy of the ensemble model in predicting the postoperative seizure outcomes in the test dataset.
Results:
Patient demographics were comparable between the training and test datasets. An ILAE class 1 outcome was achieved in 94 cases in the training dataset and 28 in the testing dataset. The analysis included 14607 iEEG electrode sites in the training and 3623 in the test dataset. The ensemble model assigned the highest weight to the spatial volatility of the statistical deviation of phase-amplitude coupling (weight: 9.2%). The ensemble model demonstrated an area under the ROC curve (AUC) of 0.814 in the training dataset and 0.784 in the test dataset. The AUC, reflecting the predictive performance, remained high in the test datasets across all three facilities: 0.786 in WSU, 0.748 in TU, and 0.867 in NCNP.
Conclusions:
Using datasets acquired internationally, this study validates our machine learning model, providing a potentially effective tool for predicting postoperative seizure outcomes in patients with drug-resistant focal epilepsy.
Funding:
JSPS JP22J23281 (to N.K.), JP22K16664 (to M.S.), JSPS JP19K09494 (to M.I.).
NIH R01 NS064033 (to E.A.).