Eigenvector Biomarker for Prediction of Epileptogenic Zones and Surgical Success from Interictal Data

Sayantika Roy ^1* Armelle Varillas ² Emily A Pereira ³ Golnoosh Kamali ² Patrick Myers ² Kristin Maria Gunnarsdottir ² Nathan Earl Crone ⁴ Adam Gene Rouse ⁵ Jennifer J Cheng ⁵ Michael Kinsman ⁵ Patrick Landazuri ⁵ Utku Uysal ⁵ Carol M. Ulloa ⁵ Nathaniel Cameron ⁵ Sara K. Inati ⁶ Kareem Zaghloul ⁶ Varina Louise Boerwinkle ⁷ Sarah Wyckoff ⁷ Niravkumar Barot ⁸ Jorge Alvaro Gonzalez-Martinez ⁸ Joon Y. Kang ⁴ Sridevi V. Sarma ²

• ¹ School of Medicine and Dentistry, University of Rochester, Rochester, United States
• ² Johns Hopkins University, Baltimore, Maryland, United States
• ³ Texas Tech University, Lubbock, Texas, United States
• ⁴ Johns Hopkins Medicine, Johns Hopkins University, Baltimore, Maryland, United States
• ⁵ University of Kansas Medical Center, Kansas City, Kansas, United States
• ⁶ National Institute of Neurological Disorders and Stroke (NIH), Bethesda, Maryland, United States
• ⁷ Barrow Neurological Institute (BNI), Phoenix, Arizona, United States
• ⁸ University of Pittsburgh Medical Center, Pittsburgh, Pennsylvania, United States

More than 50 million people worldwide suffer from epilepsy. Approximately 30% of epileptic patients suffer from medically refractory epilepsy (MRE), which means that over 15 million people must seek extensive treatment. One such treatment involves surgical removal of the epileptogenic zone (EZ) of the brain. However, because there is no clinically validated biomarker of the EZ, surgical success rates vary between 30-70%. The current standard for EZ localization often requires invasive monitoring of patients for several weeks in the hospital during which intracranial EEG (iEEG) data is captured. This process is time-consuming as the clinical team must wait for seizures and visually interpret the iEEG during these events. Hence, an iEEG biomarker that does not rely on seizure observations is desirable to improve EZ localization and surgical success rates. Recently, the source-sink index (SSI) was proposed as an interictal (between seizure) biomarker of the EZ, which captures regional interactions in the brain and in particular identifies the EZ as regions being inhibited ("sinks") by neighbors ("sources") when patients are not seizing. The SSI only requires 5-minute snapshots of interictal iEEG recordings. However, one limitation of the SSI is that it is computed heuristically from the parameters of dynamical network models (DNMs).In this work, we propose a formal method for detecting sink regions from DNMs, which has a strong foundation in linear systems theory. In particular, the steady-state solution of the DNM highlights the sinks and is characterized by the leading eigenvector of the state-transition matrix of the DNM. To test this, we build patient-specific DNMs from interictal iEEG data collected from 65 patients treated across 6 centers. From each DNM, we compute the average leading eigenvectors and evaluate their potential as a biomarker to accurately predict EZ and surgical success. Our findings show the ability of the leading eigenvector to accurately predict EZ (average accuracy 66.81% ± 0.19%) and surgical success (average accuracy 71.9% ± 0.22%) with data from 65 patients across 6 centers from 5 minutes of data, which we show is comparable with the current method of localizing the EZ over several weeks. This eigenvector