Human and automated detection of high-frequency oscillations in clinical intracranial EEG recordings

Abstract

Objective: Recent studies indicate that pathologic high-frequency oscillations (HFOs) are signatures of epileptogenic brain. Automated tools are required to characterize these events. We present a new algorithm tuned to detect HFOs from 30 to 85 Hz, and validate it against human expert electroencephalographers.

Methods: We randomly selected 28 3-min single-channel epochs of intracranial EEG (IEEG) from two patients. Three human reviewers and three automated detectors marked all records to identify candidate HFOs. Subsequently, human reviewers verified all markings.

Results: A total of 1330 events were collectively identified. The new method presented here achieved 89.7% accuracy against a consensus set of human expert markings. A one-way ANOVA determined no difference between the mean F-measures of the human reviewers and automated algorithm. Human kappa statistics (mean kappa=0.38) demonstrated marginal identification consistency, primarily due to false negative errors.

Conclusions: We present an HFO detector that improves upon existing algorithms, and performs as well as human experts on our test data set. Validation of detector performance must be compared to more than one expert because of interrater variability.

Significance: This algorithm will be useful for analyzing large EEG databases to determine the pathophysiological significance of HFO events in human epileptic networks.

Figure 1

Representative EEGs with HFO events showing: (top) multichannel recordings, (middle) single channel with event zoom, (bottom) single channel high-pass (fc = 35 Hz) zoom. (a) Pediatric patient, recorded from four neighboring frontal grid electrodes (marked event frequency ∼ 48 Hz). (b) Adult patient, recorded from four neighboring frontal grid electrodes (marked event frequencies ∼ 72 Hz, 73Hz). (Dashed lines) HFO event onset/offset times. Note the heterogeneous EEG amplitude and background characteristics between the patients.

Figure 2

Block diagram of basic EEG processing for bandpass energy-based HFO detection. The EEG signal, x(t), is analyzed first by preprocessing, then by detection, to produce an indicator, y(t), of HFO events. The short-time energy, E* (t;T), may be replaced by similar measures, e.g., line length. (Gray) Equalization may be applied during preprocessing to compensate for EEG spectral rolloff.

Figure 3

(Top row) Distribution of line length and energy values for patient 1 (solid blue lines), and difference between distributions for patients 1 and 2 (dashed red lines). (Bottom row) The cumulative distributions of line length and energy values for patient 1 (solid blue lines) and patient 2 (dashed red lines) used to select a threshold for the automated detection algorithm. Thresholds are shown for patient 1 for the 0.975 quantile.

Figure 4

Representative HFO identifications for both patients for each detector. (Top-to-bottom in each plot) Detection sequences correspond to A (red), B (green), C (blue), X (magenta), Y₁ (black), Y₂ (charcoal). Gray background highlights detections by automated detectors.

Figure 5

Comparison of automated detections (X, Y₁, Y₂) with the ground truth set of events identified by all three human reviewers. (Red) Number of ground truth events undetected (false negatives). (Green) Number of ground truth events successfully detected (true positives). (Gray) Number of unlabeled detections—these events include an unknown mixture of false positive detections made by the automated detectors, and false negative detections made by the human reviewers.