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

Abstract

<i>Objective.</i>High frequency oscillations (HFOs) are a promising biomarker of epilepsy, but automated detectors have significant risk for false positives due to diverse EEG artifacts. Many of these artifacts, previously uncharacterized in HFO research, are readily visible to clinicians under standard EEG viewing conditions. We present an artifact detector trained on clinician markings that identify when HFOs were produced by artifacts.<i>Approach.</i>Clinicians read standard resolution (10 sec per screen, all channels visible, 0-30 Hz), intracranial EEG with 8000 HFO events marked in 35 patients. They labeled each event as brain-derived or artifactual based upon their expert interpretation of the EEG at the time of the event, creating a new gold standard of HFO labeling. We used 4000 events for training/validation and 4000 for held-out prospective testing. We extracted features at the time of the HFOs from the single intracranial HFO channel and the scalp and intracranial common average reference, then trained candidate supervised learning classifiers to distinguish artifacts and non-artifactual HFOs (naHFOs).<i>Main results.</i>The resulting Michigan Intracranial Artifact Filter (MIAF) uses binary logistic regression on just intracranial data at the time of the HFO detection to remove false positives. The MIAF applied on held-out patient data significantly increased positive predictive value from 86% to 98% and resulted in an area under the precision recall curve and receiver operating characteristic curve of 99% and 92% respectively. It improved the correlation between HFOs and the seizure onset zone and resected volume in 76.5% and 88.9% of patients respectively, outperformed alternative artifact detectors, and allowed HFO analysis from all states of vigilance.<i>Significance.</i>MIAF effectively removes false positives HFO detections while retaining sufficient naHFOs for downstream analysis. Because it relies only on raw intracranial data during HFO detections, it can be easily ported to other HFO detectors and recording environments.