Abstracts

Rationale: High frequency oscillations (HFOs) are a promising biomarker of epileptic tissue. The identification of these electrographic events remains a challenge because both manual and automated methods rely on empirical, rather than physiological, definitions of an HFO. HFOs are required to reach thresholds for amplitude and duration, measured as a minimum number of oscillations. These thresholds are subjective, as human reviewers often disagree about whether an event qualifies as an HFO, especially when recordings contain almost constant high frequency activity. Moreover, the amplitude of each HFO will depend on the distance between the neural generator and the recording electrode, so a strict amplitude threshold may cause smaller events to be missed. Therefore, we studied high frequency activity using a data mining procedure to identify unique patterns in human intracranial data. This method is based on techniques for anomaly detection and pattern recognition in time series data and does not require any prior assumptions about the shape or amplitude of the events. Methods: We analyzed 36 one-minute intracranial EEG (iEEG) recordings from four patients. First, an 80 Hz high pass filter was applied, and the filtered signal was divided into 30 ms windows with 80% overlap. Then the distance (similarity) between each pair of windows was calculated using dynamic time warping. The most distinct events were identified based on the highest total distance from all other windows. We then computed features of each event, including amplitude envelope, line length, duration, peak frequency, and power spectrum. The results were compared to HFOs in the data that were visually marked by two reviewers. Results: The data mining procedure identified 718 anomalous events, 2-48 for each channel. The number of events found in each channel was proportional to the number of visually marked HFOs; 383 events (53.3%) corresponded to HFOs visually marked by two reviewers; 93 (13.0%) were HFOs marked by one reviewer; 28 (4.0%) appeared in the visually marked baseline; and 214 (29.8%) were neither marked as baseline nor HFOs. The algorithm identified these 214 events because their oscillation patterns were unique relative to the background. Multiple events showed a dominant frequency of 90-110 Hz; however, their amplitude was lower than the visually marked events. We hypothesize that these low amplitude events are HFOs with neural generators that are further from the recording electrode. For all events, the mean duration was 67 ms (range 33-387 ms) and peak frequency was 93±16 Hz. Conclusions: Data mining does not require prior assumptions about event amplitude or duration, yet it can identify HFOs that correspond to visually marked events. Moreover, it can find unique low amplitude events that human reviewers may miss. This technique can be used to obtain unbiased, data-driven characteristics of HFOs, which will inform parameter selection and design of automated detection algorithms and provide insight into a physiological definition for these events. Funding: This research was financially supported by Royal Thai Government Fellowship awarded to K. Charupanit.