Abstracts

Simplifying Automated HFO Detection: A Method to Increase Accuracy While Circumventing Threshold Optimization

Abstract number : 3.093
Submission category : 2. Translational Research / 2C. Biomarkers
Year : 2023
Submission ID : 713
Source : www.aesnet.org
Presentation date : 12/4/2023 12:00:00 AM
Published date :

Authors :
Presenting Author: Kavyakantha Remakanthakurup Sindhu, M.Tech – University of California, Irvine

Hernando Ombao, Ph.D – King Abdullah University of Science and Technology; Hiroki Nariai, MD, PhD – University of California, Los Angeles; Daniel Shrey, MD – Children's Hospital of Orange County (CHOC); Beth Lopour, Ph.D – University of California, Irvine

Rationale:

Many algorithms have been proposed for the automatic detection of high-frequency oscillations (HFOs), but they all require selection of an energy-based threshold for the EEG signal. The threshold may be selected from the literature or optimized for the current dataset, sometimes relying on visually marked HFOs as ground truth. However, there is no consensus on the most accurate approach. Because changes to the threshold will significantly impact the resulting localization of the seizure onset zone (SOZ), this may weaken results and impedes comparisons across studies. We propose the Feature Integration Across Thresholds (FIT) method to circumvent the need for optimization by using the rate and amplitude of HFOs calculated using a range of thresholds.



Methods:

FIT Method: We hypothesized that events detected consistently across a range of thresholds would most robustly identify channels in the SOZ. After HFO detection using thresholds ranging from two to eight standard deviations above the mean, the average HFO rate and amplitude were calculated for each threshold. This resulted in a rate-amplitude curve for each channel (Figure 1). The area under this curve (AUC) was used as an indicator of the SOZ. Given that high rates and amplitudes are associated with the SOZ, we hypothesized that AUC values would be higher inside than outside the SOZ.

Data Analysis: First, the FIT method was tested on simulated data consisting of pink noise with added HFOs. Then data from 35 human subjects from two centers were analyzed, consisting of 15 minutes of sleep intracranial EEG per subject. HFOs were detected in the data using the method from Staba et al. 2002, and AUC values were calculated as described above for all channels. Channels with outlier values of AUC were identified using Tukey’s upper fence. Positive predictive values (PPV) were calculated for each subject, with resected channels that were high outlier values taken as true positives. To compare to the standard method of HFO rate, outlier values of rate were also identified, and PPV values were calculated as above.



Results:

The simulation results demonstrated that the FIT method was superior to HFO rate based on a fixed threshold, especially for cases with low rate or amplitude of HFOs. In the first dataset (20 subjects), the FIT method had significantly higher PPVs (p< 0.05) compared to the rate-based method for all thresholds in seizure-free subjects (13/20). The second dataset (15 subjects) showed similar results, with the FIT method having significantly higher PPVs (p< 0.05) in seizure-free subjects (10/15) for all thresholds except one. In subjects that were not seizure-free, no significant difference was found between the two methods.

Translational Research