Identification of Pathological High Frequency Oscillations through Multi-modal Feature Classification
Abstract number :
1.038
Submission category :
1. Basic Mechanisms / 1C. Electrophysiology/High frequency oscillations
Year :
2025
Submission ID :
518
Source :
www.aesnet.org
Presentation date :
12/6/2025 12:00:00 AM
Published date :
Authors :
Presenting Author: Jack Lin, Phd – University of Michigan
Stephen Gliske, PhD – University of Nebraska
William Stacey, MD, PhD – University of Michigan
Rationale: High-frequency oscillations (HFOs) have shown promise as biomarkers for epileptogenic brain tissue. However, their clinical application remains hindered by the inability to reliably distinguish pathological HFOs from physiological ones. In the absence of a definitive labeling ground truth, a pragmatic and outcome-based approach is needed to define and identify pathological HFOs.
Methods: We introduced a classification framework that defines pathological HFOs as those localized within the resected seizure onset zone (ResSOZ) of patients who achieved seizure freedom (Engel Class I) following surgery. Conversely, non-pathological HFOs were defined as those occurring in areas neither resected nor implicated in seizure generation (norResSOZ). Using multi-day/week intracranial depth and grid recordings from these patients, we trained several machine learning classifiers—including deep neural networks and logistic regression—across three feature domains: raw filtered EEG, power spectral density (PSD), and a curated set of spectral, temporal, and morphological features (DerivedFeatures).
Results: Classifiers trained on all three feature sets successfully distinguished pathological from non-pathological HFOs, achieving AUCs between 0.73 and 0.81. Importantly, these models generalized to patients with poor surgical outcomes, indicating the presence of cross-patient features reliably associated with pathological HFOs. Additionally, we found that fast ripples (250–500 Hz), while previously hypothesized to be more specific for epileptogenic tissue, were not superior to broadband HFOs (80–500 Hz) in either ripple rate analysis or feature-based classification.
Conclusions: Our findings demonstrate that separating pathological from non-pathological HFOs enhances the specificity of HFOs as biomarkers. The continued mixing of these two types of HFOs in clinical and research contexts undermines the diagnostic value of the signal. By offering a principled, outcome-based framework, this work provides a clinically grounded foundation for transforming HFOs into a reliable tool for surgical planning in epilepsy.
Funding: R01NS094399, K01ES026839
Basic Mechanisms