Authors :
Presenting Author: Peter Galer, PhD – Neumarker Inc.
Gregory Grecco, MD, PhD – Indiana University School of Medicine
Fan Zhang, PhD – Neumarker Inc.
Amulya Mathur, MS – Neumarker Inc.
Jonathan Halford, MD – Ralph H. Johnson VA Healthcare System
Andrew Cole, MD – Massachusetts General Hospital and Harvard Medical School
Amy Chappell, MD – Solaxa
Elizabeth Garofalo, MD – EAG Pharma Consulting
Jacqueline French, MD – Department of Neurology and Comprehensive Epilepsy Center, New York University Grossman School of Medicine, NYU Langone Health
Michael Detke, MD, PhD – Neumarker Inc.
Ken Wang, PhD – Neumarker Inc.
Qiang Li, PhD – Neumarker Inc.
Rationale:
Anti-seizure medications (ASM) are the first-line treatment for patients with epilepsy. Nevertheless, regardless of their possible epilepsy syndrome diagnosis, it is not clear which ASM will be most effective for an individual patient. Nearly half of all patients with epilepsy trial multiple ASMs, and, among this population, about half will never obtain seizure freedom. Machine learning models that accurately predict the most effective ASM for an individual could be very impactful. Leveraging clinical and EEG data collected from the Human Epilepsy Project (HEP), we developed machine learning models to predict ASM response in individuals with newly diagnosed focal epilepsy.
Methods:
We examined scalp EEG and ASM treatment outcome data from 280 individuals (range 11-65 year of age at recording) from 28 hospital systems collected for the HEP study. HEP is a multicenter prospective study which enrolled individuals with newly diagnosed focal epilepsy within 4 months of initiating ASM treatment. Separate ASM prediction models were developed for predicting response to levetiracetam from a resting-state EEG recorded while an individual was not taking any ASM (35 responders, 23 non-responders) and a model from an EEG recorded while taking levetiracetam (17 responders, 31 non-responders). We also developed a lamotrigine model (21 responders, 12 non-responders) and treatment resistant model (69 responders, 17 refractory) for EEGs recorded while an individual was not taking any ASM. Treatment response was defined as seizure free for 12 months or 3 times an individual’s longest pretreatment seizure freedom, whichever was longer. Treatment resistant was defined as failing two or more adequate trials of ASMs. For each EEG, we calculated dynamic network relationships between pairs of electrodes across different bandpowers within non-overlapping 2-second time windows (example feature set in Figure 1). We then implemented a greedy search to find the top 10 features to distinguish responders from non-responders.
Results:
After performing 10-fold cross validation, we found that the unmedicated and medicated levetiracetam model, respectively, achieved a mean AUC of 0.97±0.01 and 0.92±0.03 and a balanced accuracy of 0.90±0.03 and 0.77±0.04. The unmedicated lamotrigine model achieved a mean AUC of 0.97±0.02 and balanced accuracy of 0.87±0.04, and the unmedicated treatment resistant model achieved a mean AUC of 0.93±0.03 and balanced accuracy of 0.76±0.04.
Conclusions:
Our preliminary results suggest that resting-state scalp EEG contains information which correlates with future ASM response in new onset focal epilepsy patients. Although limited in sample size, our findings provide a proof of concept of the potential of a precision-medicine-based approach to ASM treatment based on routine clinical EEG. Future work includes plans to refine our models, expand to other ASMs, and validate our findings with EEGs from other hospital systems. Our eventual goal is to create a comprehensive ensemble model that can guide clinicians and patients to their optimal ASM after just a single clinical scalp EEG.
Funding: Neumarker Inc.