Abstracts

Rationale: Intracranial electroencephalography (iEEG) presents a unique opportunity to extend or complement findings from cognitive research studies utilizing neuroimaging modalities, such as fMRI, MEG, and scalp EEG. However, most iEEG data are recorded from patients diagnosed with drug resistant epilepsy who display EEG data with 1) physiological activity, 2) non-cerebral artifacts, and 3) transient pathological activity known as interictal epileptiform discharges (IEDs) or high frequency oscillations (HFOs). Outside of the clinical utilization of IEDs and HFOs for localizing epileptogenic tissue, IEDs have been shown to disrupt proper cognitive processes, potentially confounding results in cognitive iEEG studies and making it a major concern in preprocessing. In this study, we employed multiple annotated iEEG datasets to develop an automated classifier capable of marking and classifying events from continuous task-based iEEG as physiology, artifact, or pathology.

Methods: In this study, we utilized two public datasets of iEEG data recorded from depth and subdural iEEG electrodes during different brain states. Dataset 1 (DS1) was collected at two institutions: St. Anne’s University Hospital (Brno, Czech Republic) and Mayo Clinic (Rochester, Minnesota) and contains 3-second data segments (fs =5,000Hz) classified as non-cerebral (n=73,902), pathological (n=67,797), or physiological (n=151,290). Dataset 2 (DS2) was collected at Motol University Hospital (Prague, Czech Republic) and contains 5-min of continuous recordings (fs =1,000Hz) with expert IED annotations. Features were extracted from DS1 to train a random forest classifier (RFC) and gradient boost classifier (GBC) to distinguish events as non-cerebral, pathological, or physiological. Segments were filtered at 0.5Hz and 50/60Hz to remove drift and power line noise, down sampled to 1000Hz, and processed to extract features. Extracted features included (1) frequency specific components related to beta/gamma (20-80Hz), ripples (80-200Hz), fast ripples (200-500Hz), and wicket spikes (6-11Hz); (2) morphological features related to line length, number of crossings, peaks/troughs, max amplitude, peak-to-peak amplitude; (3) anatomical location; (4) brain state; and (5) electrode type.

Results: On DS1, our RFC had an overall accuracy 88.6% (GBC: 86.2%) with class specific accuracies of 89.4% (physiology), 94.9% (pathology), and 93.0% (non-cerebral artifacts). The sensitivity, specificity, and precision values for all three classes > 0.85 (see Table 1). On DS2, our RFC had an area under the curve (AUC) = 0.83 (Figure 1), sensitivity =0.85 and specificity of 0.84 for annotated IEDs.

Conclusions: Our RF classifier successfully classified iEEG data into three distinct classes with high levels of accuracy, precision, sensitivity, and specificity on a larger multi-institutional dataset. Additionally, we used the same classifier on a dataset collected at a lower sampling rate and achieved similar performance levels.

Funding: Please list any funding that was received in support of this abstract.: Research was supported by the National Center for Advancing Translational Sciences (TL1 TR001428).