Authors :
Presenting Author: Hmayag Partamian, PhD – The University of Texas at Arlington
Saeed Jahromi, MSc – Cook Children's Health Care System
M. Scott Perry, MD – Cook Children’s Physician Network
Eleonora Tamilia, PhD – Boston Children’s Hospital, Harvard Medical School, Boston, MA, USA
Joseph R. Madsen, MD – Boston Children's Hospital
Jeffrey Bolton, MD – Boston Childrens Hospital
Scellig Stone, MD, PhD – Boston Childrens Hospital & Harvard Medical School
Phillip Pearl, MD – Boston Children's Hospital & Harvard Medical School
Christos Papadelis, PhD – Cook Children's Health Care System
Rationale:
Accurate localization of the epileptogenic zone (EZ) is crucial for successful resective surgery in children with drug-resistant epilepsy (DRE). Several biomarkers, such as spikes, ripples, and their combinations derived from intracranial interictal electroencephalogram (iEEG) recordings, have shown promising findings, but their specificity and sensitivity vary widely across patients. Here, we build on the previous work of our group on the development of interictal iEEG biomarkers by proposing a machine learning (ML) framework that integrates multiple biomarkers into a single model, thereby automatically capturing patterns that generalize across patients and improving the localization of the EZ.
Methods:
We retrospectively analyzed iEEG from 60 children with DRE undergoing neurosurgery, classified as good (n=39, Engel I) or poor (n=21, Engel≥ II) outcome (follow-up≥ 1 year). The resection volume, defined using pre- and postoperative MRIs, was used as the gold standard to define the EZ (Fig. 1A). Five-minute interictal iEEG segments were processed to detect spikes and ripples automatically (Fig. 1B). For each channel, spike and ripple rates, propagation indices, powers, and their co-occurrence were computed (Fig. 1C). We also calculated amplitude envelope correlation connectivity across frequency bands and derived graph-theoretic metrics including eigenvector, betweenness, and degree centralities. Mean spectral powers were also computed for each channel. Individual features were evaluated using Youden’s method to determine optimal classification thresholds. All features were then used to train a support vector machine (SVM) classifier to automatically classify electrodes as epileptogenic (EP) or non-epileptogenic (NEP) (Fig. 1D). Using leave-one-patient-out cross-validation on good-outcome patients, we compared SVM model’s performance with the one of each feature. We then assessed whether resecting predicted epileptogenic regions with >50% overlap with resection could provide better surgical outcome prediction than individual features.
Results:
We found that the epileptogenic network had higher power inside (vs. outside) resection (p< 0.05) for most features in good-outcome patients but only in a few poor-outcome cases (Fig. 2A). Spike rate, spike ripple overlap, eigenvector in γ, betweenness in γ, and degree in β and γ achieved specificity of >75% and precision of >50%. The SVM outperformed individual features, reaching 93% specificity, 72% precision, and 74% accuracy (Fig. 2B). The estimated epileptogenic region in good-outcome patients overlapped more with the resection than in poor-outcome patients (p< 0.05) (Fig. 2C). Although many features predicted outcome (p< 0.05), the SVM model outperformed them achieving a 78% accuracy (p< 0.05) (Fig. 2D).
Conclusions:
Our study shows that a multi-feature ML framework improves EZ localization from interictal iEEG and predicts surgical outcome in DRE more accurately than individual interictal biomarkers. This automated, multi-biomarker approach provides clinicians a reliable tool to support presurgical decisions and improve surgical outcomes.
Funding:
RO1NS104116-01A1 & R01NS134944-01A1 by NINDS.