Abstracts

Rationale: Studies have shown that the rate of high frequency oscillations (HFOs), measured as the number of HFOs occurring per minute per channel, is generally higher in the seizure onset zone (SOZ), and it has therefore been suggested as a potential tool for planning of epilepsy surgery. However, the rate is not a robust metric, as relative rates between individual electrodes change over time, and the measurement is negatively impacted by false positive detections. As an alternative, differences in HFO amplitude between SOZ and non-SOZ channels have also been reported, but these differences are often small. This could be due to the choice of HFO detection algorithm and its associated parameters; for example, the use of an amplitude threshold to detect HFOs may bias the estimate of event amplitude. Therefore, we measured HFO amplitude using two different automated detection methods and evaluated it as a potential biomarker of the SOZ. Methods: We detected HFOs in intracranial EEG from 11 patients undergoing evaluation for epilepsy surgery using a novel machine learning approach (anomaly detection algorithm, ADA) and a standard amplitude-based approach (root-mean-square, RMS). Multiple independent 3-minute segments of intracranial EEG were analyzed for each subject (range: 7-17), and both SOZ and non-SOZ channels were analyzed; in total, we detected HFOs in 1,958 single-channel segments of data. We first compared the events detected by ADA and the RMS detector by measuring event rate, amplitude, and duration. We also calculated the number of events detected only by ADA, the number detected only by RMS, and the percentage of events detected by both methods. Then, for each detection algorithm, we used the rate and amplitude of high frequency events to classify SOZ and non-SOZ channels, and we compared the performance using receiver operating characteristic (ROC) curves. Results: Consistent with prior literature, the rate of HFOs was higher in SOZ channels than non-SOZ channels when the RMS algorithm was used for detection (p<0.05, Wilcoxon rank sum test). In contrast, the rate of ADA events did not exhibit robust differences between SOZ and non-SOZ channels. However, for both ADA and RMS, the amplitude of detected events was significantly higher in SOZ than non-SOZ channels (p<0.05, Wilcoxon rank sum test). Across individual subjects, amplitude more accurately classified SOZ and non-SOZ channels than HFO rate, as indicated by significantly higher values of area under the ROC curve and sensitivity, as well as lower false positive rates (n=11, p<0.05, Wilcoxon signed-rank test). The values of both rate and amplitude fluctuated over time, but the amplitude measurement was more consistent across segments of data (indicated by a lower coefficient of variation, p<0.05, Wilcoxon rank sum test), and it exhibited lower noise levels in non-SOZ channels. Moreover, the classification accuracy based on amplitude remained high over a range of detection parameters, while the performance of rate suffered when the detector was not properly optimized. Conclusions: As a biomarker of the SOZ, HFO amplitude offers three advantages over the standard metric of HFO rate: (1) It enables more accurate classification of SOZ and non-SOZ channels, (2) The measurement of amplitude is more consistent over time, and (3) The performance of amplitude as a biomarker is robust to changes in detection parameters. Such improvements have the potential to increase the generalizability of HFOs and facilitate clinical implementation as a tool for localization of the SOZ during surgical planning. Funding: Royal Thai Government fellowship awarded to K. Charupanit