Abstracts

Rationale: Variations in methods to detect High Frequency Oscillations (HFOs) are a potential confound in HFO analysis and interpretation. Previous comparisons between automated HFO detectors were generally limited by using recordings of only a few hours and inconsistencies in the preprocessing and handling of artifacts. Our objective was to compare HFO detectors using the full-duration recording, with a standard preprocessing pipeline and our previously published methods for redacting artifacts.

Methods: : Data were processed in 10-minute epochs. An automated detector of data quality redacted channels with poor data quality per epoch before computing the common average reference. An 80-500 Hz elliptical band-pass filter was followed by a putative HFO detector. As of date, we have analyzed Staba et al. (2002), denoted method A, or Charupanit et al. (2020), denoted method B. Analysis of additional detectors is on-going. Fast transient, muscle, and background artifacts were removed using the methods of Gliske et al. (2016), Ren et al. (2019), and Gliske et al. (2020). Association with the clinically identified Seizure Onset Zone (SOZ) was quantified using the asymmetry measure. Method B has a free parameter which controls the overall sensitivity. We used a value of 0.001 based on analysis of the first subject.

Results: We analyzed 68 sequential patients from the University of Michigan Intracranial EEG database, including 5844 channels and spanning 11,108 hours, resulting in over 12 million HFO detections (method A) and 38 million HFO detections (method B). Method B detected more HFOs at both the population level (p=0.0095, Wilcoxon Rank Sum) and patient level (p=7x10^-9; Paired Wilcoxon Sign Rank), with the ratio (Method B / Method A) being positively correlated with the number of HFOs detected in Method A (rho=0.33, p=0.007, Spearman Correlation Coefficient) and Method B (rho=0.69, p< 10^‑15, Spearman Correlation Coefficient). Method A had a median of 55% of events not detected by Method A, while Method B had a median of 72% of events not detected by Method A. The overall association with SOZ was not statistically different between the two methods (p=0.09, Paired Wilcoxon Sign Rank). A positive correlation was observed between the total HFO counts and the difference in the asymmetry in the two methods (rho=0.3, p=0.02, Spearman Correlation Coefficient). Restricting to the 25% of patients with the highest number of HFOs lead to a statistically significant difference in the association with SOZ regardless of whether the cut was based on Method A (p=0.015, Paired Wilcoxon Sign Rank) or Method B (p=0.010, Paired Wilcoxon Sign Rank).


Conclusions: HFO methods analyzed included a sizeable overlap with each other as well as a sizeable number of unique events. However, among the full population, neither was inferior nor superior in their association with SOZ in the full population, though among subjects with the highest HFO counts, Method B had higher association. The correlation with number of counts suggests the need for a patient-dependent tuning of the sensitivity factor.


Funding: R01 NS094399