A comprehensive, physician-trained algorithm to remove artifactual false positive high frequency oscillations in...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Epilepsy Research

Background:

• High frequency oscillations (HFOs) are sensitive biomarkers for epilepsy.
• Automated HFO detection methods often yield false positives due to EEG artifacts.
• Clinicians can visually distinguish artifactual HFOs, but this is not scalable for automated systems.

Purpose of the Study:

• To develop and validate an automated artifact detector for high frequency oscillations (HFOs).
• To create a new gold standard for HFO labeling using expert clinician markings.
• To improve the accuracy and reliability of HFO detection in epilepsy patients.

Main Methods:

• Clinicians labeled 8,000 intracranial EEG events from 35 patients as brain-derived or artifactual.
• Supervised learning classifiers were trained using extracted features from intracranial EEG data.
• The Michigan Intracranial Artifact Filter (MIAF) was developed using binary logistic regression on intracranial data.

Main Results:

• The MIAF significantly increased positive predictive value for HFO detection from 86% to 98%.
• Achieved high performance with Area Under the Curve (AUC) of 99% (Precision-Recall) and 92% (ROC).
• Improved correlation between HFOs and seizure onset zones/resected volumes in a majority of patients.

Conclusions:

• The MIAF effectively filters false positive HFO detections while preserving true positives.
• The filter's reliance on raw intracranial data allows for easy integration with other HFO detection systems.
• This tool enhances the clinical utility of HFOs as epilepsy biomarkers.