Authors :
Presenting Author: Shi Bei Tan, BS – University of Michigan
Stephen Gliske, PhD – University of Nebraska Medical Center
Neha John, MS – University of Michigan
Temenuzhka Mihaylova, MD PhD – University of Michigan
Garnett Smith, MD – University of Michigan
Nancy McNamara, MD – University of Michigan
Nicholas Beimer, MD, PhD – University of Michigan
Erin Romanowski, DO – University of Michigan
William Stacey, MD, PhD – University of Michigan
Rationale:
High Frequency Oscillations (HFOs) are a promising biomarker of the epileptogenic zone. Automatic HFO detectors reduce manual effort but often yield false positives due to artifacts, many of which are not recognized in single-channel EEG. Clinicians are trained to identify and ignore artifacts while reading multichannel EEG, highlighting limitations of single-channel-trained HFO detectors. We developed an artifact detector trained on clinician-labeled multichannel EEG to reduce false positives and ensure compatibility with any HFO detection pipeline.Methods:
EEG data from 35 patients yielded 40 files covering all sleep and wakefulness stages. We used the qHFO detector to identify HFOs, randomly selecting 200 events per file across the whole EEG duration with a 1:4 ratio inside vs. outside the seizure onset zone (SOZ). Six clinicians (three pediatric, three adult epileptologists) reviewed EEGs using the Persyst viewer in standard clinical resolution, where the HFO events were marked as ‘comments’ at the precise time of HFO detection. They labeled the events as either Artifact, Brain, or Uncertain based on clinical interpretation of the background activity, not the filtered HFO signal. Each event was labeled independently by three clinicians. Events labeled as Artifact with consensus were considered artifacts; others were non-artifactual HFOs (naHFOs). The first 20 files were used for training/validation, and the remaining 20 for testing.
We computed 27 features using the HFO channel and multichannel scalp and intracranial common average reference (CAR). Models were used to distinguish between artifacts and naHFOs. Models tested included logistic regression, neural networks, and support vector machines. Models were evaluated using area under precision recall curve (AUPRC), area under receiver operating characteristic curve (AUROC), precision, specificity, and sensitivity. An asymmetry measure assessed the relationship between HFOs and epileptic tissue post-artifact removal. Feature importance analysis identified discriminative characteristics of artifacts and naHFOs.
Results:
8,000 HFO events were labeled; 52.0% were Brain signals and 16.7% Artifacts by consensus. Artifact rejection was effective even without scalp EEG. The selected logistic regression model using only intracranial EEG features achieved an AUPRC of 0.97, AUROC of 0.90, precision of 0.98, specificity of 0.93, and sensitivity of 0.72 on the held-out test set. Applying the detector improved HFO asymmetry inside vs. outside SOZ and resected volume (RV) in 82.4% and 83.3% of patients, respectively. Features linked to artifacts included higher mean HFO curvature, power, and peak HFO amplitude; features associated with naHFOs included higher median HFO frequency, ripple-to-fast ripple ratio, and peak intracranial CAR amplitude.
Conclusions:
We developed an artifact detector, trained on clinician-labeled full EEG, that effectively removes false positives, enhancing the clinical utility of HFO analysis, and is compatible with any HFO pipeline.Funding:
NIH R01 NS949399