Assistant Professor Ghent University, Epilog NV, Belgium
Rationale: In clinical practice, the choice of antiseizure drug (ASD) is often not guided by efficacy considerations but rather by ease of use and side effect profile related to the specific patient. This inevitably leads to patients having new seizures during the establishment of treatment and is not only a burden on the quality of life of patients but also holds a significant risk of injury from seizures or even sudden unexpected death. Therefore, a biomarker that allows assessing whether or not a patient will respond to a specific ASD, without the need for waiting for the next seizure, would allow to decide more quickly how to continue therapy. Methods: In this study, 76 patients who were treated with Levetiracetam as mono-therapy after a first seizure from the Neurology Clinic in Marburg were included. Clinical information such as seizure type, drug abuse, familial history, number of risk factors, etc. and baseline EEG recorded before treatment initialization were collected for each patient. Spectral and functional connectivity analysis was performed to extract features from the routine EEG. Machine learning was used to build a classifier that predicts whether or not a patient would respond to the treatment. A responder was defined as a patient who was completely seizure-free after treatment. A classifier was built separately based on the clinical EEG data and on the EEG features. Five-fold cross-validation was performed to assess the performance of the classifier. As outcome measures the accuracy, sensitivity, specificity, positive predictive value and negative predictive were calculated. Results: The classifier of the clinical data led to an error rate of 46%, indicating that based on clinical data no difference was found between responders and non-responders. The EEG data led to an error rate of 15%. The accuracy was 81% with a 95% confidence interval from 70% to 89%. The sensitivity was 86%, specificity 74%, PPV 82% and NPV 79%. Most distinguishing features were FCA features and delta energy of electrodes of the frontal lobe. Conclusions: This retrospective study shows the potential of an EEG biomarker indicative of the efficacy of ASD treatment with high accuracy. Nevertheless, prospective studies are needed to assess the added value of the developed EEG biomarker in a clinical setting. Funding: Please list any funding that was received in support of this abstract.: Pieter van Mierlo is funded by the Swiss National Science Foundation grant No. CRSII5_180365 and CRSII5-170873.
