DIME: Data-driven Importance MEtric Guides Localization of the Seizure Onset Zone from Intracranial EEG Data

Abstract

Epilepsy affects over 50 million individuals, many of whom require surgical treatment that is dependent on accurate localization of the seizure onset zone (SOZ). Conventional SOZ biomarkers are based on strictly defined intracranial electroen-cephalography (iEEG) phenomena and cannot benefit from increased datasets. The scarcity of SOZ-labeled iEEG data impedes biomarker development. We introduce the Data-driven Importance MEtric (DIME) to guide SOZ localization in an interpretable pipeline that improves with ictal-labeled iEEG data. We apply DIME to an open-source dataset (n=21; 13 successful; 8 failed) for SOZ localization and surgical outcome prediction. The highest DIME-ranked electrode belonged to the clinically annotated SOZ for 69.2% of patients with successful surgery (p < 0.001). DIME scores were significantly higher in SOZ electrodes than nonSOZ electrodes in both successful and failed surgeries (p < 0.001), though the DIME distribution for successful cases differed from failed cases (p = 0.002). DIME predicted surgical outcome with 92.3% recall and 66.7% accuracy.