Elimination of pseudo-HFOs in iEEG using sparse representation and Random Forest classifier

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• High-Frequency Oscillations (HFOs) are key biomarkers for identifying the epileptogenic zone.
• Conventional HFO detectors struggle with sharp artifacts, leading to inaccurate seizure onset zone (SOZ) localization.
• Distinguishing true HFOs from pseudo-HFOs (artifacts) is crucial for precise epilepsy diagnosis.

Purpose of the Study:

• To develop a novel classification method for differentiating true HFOs from pseudo-HFOs using sparse representation.
• To improve the accuracy of seizure onset zone (SOZ) localization in epilepsy patients.
• To classify detected HFOs into Ripples (R) and Fast Ripples (FR) subcategories.

Main Methods:

• Implemented a sparse representation framework using Orthogonal Matching Pursuit (OMP) and a Gabor dictionary.
• Candidate HFO events were sparsely represented over 30 iterations, forming a 30-dimensional feature vector.
• A random forest classifier was trained on these feature vectors, with expert visual inspection of 2075 events from 5 subjects.

Main Results:

• Achieved 90.22% classification accuracy for distinguishing true HFOs from pseudo-HFOs.
• Demonstrated a 21.16% improvement in SOZ localization compared to conventional amplitude-threshold detectors.
• Attained 91.24% SOZ accuracy when classifying HFOs into R and FR subcategories.

Conclusions:

• The developed sparse representation framework offers a robust method for identifying true HFOs in long-term iEEG recordings.
• This approach enhances SOZ identification reliability without necessitating the exclusion of artifact-containing segments.
• The method provides a significant advancement in HFO analysis for epilepsy research and clinical practice.