Authors :
Presenting Author: Hmayag Partamian, PhD – The University of Texas at Arlington
Saeed Jahromi, M.Sc. – Department of Bioengineering – The University of Texas at Arlington; Scott Perry, MD – Cook Children's Health Care System; Eleonora Tamilia, PhD – Harvard Medical School; Joseph Madsen, MD – Harvard Medical School; Jeffrey Bolton, MD – Harvard Medical School; Scellig Stone, MD – Harvard Medical School; Phillip Pearl, MD – Harvard Medical School; Christos Papadelis, PhD – Cook Children's Health Care System
Rationale:
Accurately delineating the epileptogenic zone (EZ) in children with drug resistant epilepsy (DRE) can help control or free patients from seizures and the associated symptoms. Traditionally, human experts visually analyze intracranial EEG (iEEG) to identify the seizure onset, which serves as the best EZ approximator. However, signals contain hidden information that is not generally visually evident. Here, we use an unsupervised machine learning (ML) technique to decompose iEEG signals into dominant components and automatically identify channels that contain epileptogenic activity from interictal recordings. We hypothesize that the proposed biomarker can delineate
the EZ with high precision and predict the surgical outcome of children with DRE.Methods:
We analyzed interictal iEEG data recorded from 26 good (Engel 1) and 14 poor outcome patients (Engel 2-4) with DRE. For each patient, 0.25-second-long windows were used as input to
an unsupervised ML method, the Dynamic Mode Decomposition (DMD), to learn a finite number of features per iEEG contact (Fig. 1A). To fully reconstruct a spike, we note it depends on higher frequency components more compared to non-spike (Fig. 1B). For each patient, we used a data-driven threshold, the clinical information, and imaging data to project the biomarker values onto the brain (Fig. 1C). The identified regions were then analyzed in relation to the resected areas and the seizure onset zone (SOZ). The biomarkers were computed in: (1) all modes; (2) modes with frequency values below 30 Hz (≤β); and (3) above 30 Hz (>β) and comparatively analyzed. We then evaluated the significance of the proposed biomarker inside and outside the resection volume and in SOZ and non-SOZ channels that served as the gold standards of the EZ. We finally assessed the ability of the proposed biomarker to predict surgical outcome.Results:
We found that the proposed biomarker identifies clusters of channels close to the resection volumes in good-outcome patients, but they are scattered in poor outcome patients (Figure 2A). Our analysis shows that > β modes show significance (Wilcoxon rank sum) in specificity (p=
0.0106) and precision (p=0.00076) between good and poor outcome patients (Fig. 2B). The normalized values of the biomarker showed significance between inside and outside resection (<10 mm) in good (p=0.049) but not in poor outcome patients (p=0.26) (Figure 2C). Figure 2D shows the confusion matrix and the performance metrics of the three cases (all modes, ≤β, >β) when the resection region (<