Authors :
Presenting Author: Mark Hays, PhD – Neurologic Solutions
Golnoosh Kamali, PhD – NeuroLogic Solutions
Mary Rose Streett, BS – Johns Hopkins School of Medicine
Sarah Johnson, BS – Thomas Jefferson University
Dale Wyeth, MBA – Thomas Jefferson University
Ronald Miranda, BS – University of Maryland Medical Center
Guang Ren, BS – University of Maryland Medical Center
Anant Walia, MD – University of Maryland School of Medicine
Andrew Gotshalk, BS – NeuroLogic Solutions
Michael R. Sperling, MD – Jefferson Comprehensive Epilepsy Center, Thomas Jefferson University
Khalil Husari, MD – Johns Hopkins School of Medicine
Jennifer Hopp, MD – University of Maryland School of Medicine
Jorge González-Martínez, MD,PhD – University of Pittsburgh Medical School
Sridevi Sarma, PhD – Johns Hopkins University
Rationale:
50 million people worldwide are diagnosed with epilepsy, but around 30% may be misdiagnosed. High misdiagnosis rates stem from lack of a simple, reliable diagnostic process, instead relying on a comprehensive clinical evaluation. Visual inspection of a routine scalp EEG can confirm an epilepsy diagnosis if interictal epileptiform discharges (IEDs) are present, but the rarity of IEDs often leads to inconclusive diagnoses. Patients may undergo further evaluation in the Epilepsy Monitoring Unit (EMU) for days to weeks to capture a seizure or IEDs on EEG while off anti-seizure medications and sleep-deprived to provoke seizures. Despite this costly process, some patients have insufficient or no seizures, leading to an inconclusive or possibly incorrect diagnosis. By characterizing imbalances within brain networks that can’t be visually detected from EEG, dynamical network models (DNMs) offer a potential solution to this clinical need for accurate diagnoses from inconclusive EEG. While EpiScalp™, our patent-protected algorithm that applies these DNMs to predict epilepsy diagnoses, was originally validated in routine EEG, we sought to investigate its performance in the noisier EMU setting.
Methods:
We gathered interictal scalp EEG data from the Johns Hopkins Hospital and University of Maryland Medical Center for 62 patients undergoing evaluation in the EMU. DNMs, derived from approximating short windows of EEG as linear time invariant models, were built to characterize the network’s N to N channel interactions across time, for the first hour of the EMU recordings. We hypothesized that neural fragility and source-sink index, two DNM-derived metrics previously used to localize the epileptogenic zone (EZ) in intracranial EEG recordings, can be applied to interictal scalp EEG to distinguish between epileptic and non-epileptic brains. The EZ nodes in epileptic brains would have high fragility and behave like strong sinks, while these dynamics would be absent in non-epileptic brains. We trained a logistic regression model to predict epilepsy and non-epilepsy diagnoses using these features, treating each 20 minute segment of the hour of EMU recording as a separate sample. 62 patients were split into training (39) and testing (23) sets. The predicted probabilities were converted to epilepsy risk scores ranging from 0-100, and scores were averaged over the three 20 minute segments for each patient to compute an overall risk score.
Results:
EpiScalp™ was tested on 23 EMU patients with either normal or abnormal scalp EEG. EpiScalp™ made a predictive definitive diagnosis for 17 of those patients with 83% accuracy, 83% sensitivity and 82% specificity.
Conclusions:
These findings are significant because they show that EpiScalp™ can serve as a valuable tool to assist in diagnosing inconclusive EEG in the EMU within the first hour, possibly removing the need for lengthy stays in the EMU and decreasing the financial burden on the healthcare system.
Funding:
This study was funded by the Maryland Innovation Initiative and NINDS SBIR Phase 2 grant.