Graduate Researcher University of California, Irvine
Rationale: High frequency oscillations (HFOs) are a promising new biomarker of epileptogenicity, as they occur more frequently in the seizure onset zone (SOZ) and may aid demarcation of the epileptogenic zone. Development of reliable, automatic HFO detection algorithms is necessary for translation into clinical practice. While existing algorithms have sufficient levels of sensitivity and specificity when applied to individual data sets, there are currently no standards for their broad application. It is not uncommon for a previously validated algorithm to work poorly when applied to a new data set, and there is no consensus on whether parameter optimization should be done. Here, we evaluate the impact of detector optimization on two independent datasets using a widely cited automatic HFO detector based on the root-mean-square (RMS) amplitude. Methods: Twenty medically refractory epilepsy patients who were seizure-free after resective surgery were included in this study. Data for seven subjects were recorded at the University of California, Irvine Medical Center, and data for 13 subjects were obtained from the freely available dataset associated with Fedele et al. 2017. We used the RMS algorithm to detect HFOs across a wide range of detection parameters, and we subsequently applied two validated automatic HFO rejection methods to reduce false positives. To measure the accuracy of SOZ localization based on elevated HFO rates, we computed the precision-recall (PR) and receiver-operating-characteristic (ROC) curves for all parameter sets, and we assessed the variance in SOZ localization results across patients. The optimal parameters for each patient were determined based on the optimal F1 score of the PR curve and the area-under-the-curve (AUC) of the ROC curve. Results: While varying parameters in the RMS detection algorithm, we found that the duration of the RMS window and the detection threshold significantly impacted the localization results, while changing the minimum event duration had little influence. Automatic rejection of artifacts had a small, but positive effect on the results. Across all 20 patients, the optimal detection parameters were patient-specific, and in some cases, the most accurate localization resulted from detection with unconventional parameters. This suggests that the standard configurations are not suited for all patients, and testing of a wide range of detection parameters is necessary. Conclusions: This study demonstrates the significance of patient-variability and suggests the need to optimize parameters for individual patients before applying automatic detection algorithms. Funding: Please list any funding that was received in support of this abstract.: None.