Authors :
Presenting Author: Ravinderjit Singh, MD, PhD – Icahn School of Medicine at Mount Sinai
Madeline Fields, MD – Department of Neurology, Icahn School of Medicine at Mount Sinai
Benjamin Kummer, MD – Icahn School of Medicine at Mount Sinai
Rationale:
There is currently no systematic method to distinguish individuals with well-controlled epilepsy (WCe) from those with multi-drug resistant epilepsy (MDRe) based on EEG features. Typically, MDRe is identified only after multiple failed trials of antiepileptic drugs (AEDs), delaying access to advanced therapies such as vagus nerve stimulation (VNS), responsive neurostimulation (RNS), resective surgery, and AEDs reserved for MDRe. Earlier identification of MDRe could improve clinical outcomes by expediting appropriate interventions. We conducted a pilot study to evaluate whether EEG-derived features could distinguish individuals with no Epilepsy (NE), with WCe, and with MDRe.Methods:
Since 2008, Mount Sinai has collected over 240,000 EEG records from more than 80,000 unique patients. The Mount Sinai Data Warehouse (MSDW) enables targeted queries to identify specific patient populations. We aim to leverage this resource to identify thousands of individuals with well-controlled epilepsy (WCe), multi-drug resistant epilepsy (MDRe), and no epilepsy (NE). WCe is defined as seizure control on one or no antiepileptic drugs (AEDs). MDRe refers to patients who have failed adequate trials of two AEDs, whether as monotherapy or in combination.
For a pilot analysis, we manually identified 10 individuals with NE, 10 with WCe, and 11 with MDRe. WCe and MDRe cases were sourced from MSDW; non-epileptic controls were drawn from open-source datasets. EEG data were preprocessed using a custom artifact rejection pipeline. We extracted features capturing asymmetry in canonical frequency band power across electrodes. Principal component analysis (PCA) reduced the feature space to two dimensions, and a one-vs-rest support vector machine (SVM) classifier was used to assign each sample to one of the three groups.
Results:
The classification approach achieved an overall accuracy of 81% across 31 participants. Sensitivity, specificity, and positive predictive value (PPV) for each group are summarized in the table below. A visualization of the feature space and support vector machine (SVM) classification boundaries is provided in the accompanying figure.
| Group |
Sensitivity (%) |
Specificity (%) |
PPV (%) |
| No Epilepsy |
100 |
90.5 |
83 |
| WCe |
70 |
90.5 |
78 |
| MDRe |
73 |
90 |
80 |
Conclusions:
Early results are promising but the analysis needs to be grown to a larger dataset. If a high performance is maintained on a larger dataset, EEG features that could modify clinical practice could be developed. These features can be impactful in the initial diagnosis of epilepsy and in identifying individuals with MDRe earlier in the disease course. We plan to utilize Mount Sinai’s large dataset to grow our sample size and report updated results at AES.
Funding:
This work was supported in part through the Mount Sinai Data Warehouse (MSDW) resources and staff expertise provided by Scientific Computing and Data at the Icahn School of Medicine at Mount Sinai.