Abstracts

Rationale:
High Frequency Oscillations (HFO) are a known biomarker of epileptogenesis[1]. Identifying pathologic HFO in traumatic brain injury patients may facilitate further research in antiepileptogenic therapies, a primary aim of the Epilepsy Bioinformatics Study for Antiepileptogenic Therapy (EpiBioS4Rx). Identifying HFO in EEG is challenging due to a variety of artifacts. Epileptologists can separate true HFO events from artifacts by visual inspection, however this procedure is time consuming and subjective. Therefore, it is necessary to develop reliable and efficient automatic methods to detect HFO. RippleLab can detect HFO automatically,[1] but there are many false positives. To better optimize the use of RippleLab we implement a new approach for a faster and more efficient independent component analysis (ICA) rejection using component correlations, as some correlations remain since the signal and artifacts are not completely independent. We compare the results from short line length (SSL HFO detection) before and after artifact removal.
Method:
Rodent EEG from EpiBioS4Rx was collected at the University of Eastern Finland[2]. The EEG consists of ten channels sampled at 5000 Hz for 5.5 hours. The signals are resampled at 1000 Hz and filtered to a bandpass FIR at 80-480 Hz. For this analysis we assume that artifact-related components are independent of EEG and artifact components are correlated with each other to some extent. We perform ICA in EEGLAB.[3] The channels are processed in two steps: whitening (decorrelation) and rotation. Under our assumption, although the components are supposed to be uncorrelated, there still remain some correlations among the components. Therefore, we reject the components with correlation sums > 0.01. Once artifact components are identified, we reconstruct the EEG signal. SSL detection is then applied on both sets.
Results:
The results from ICA show that background noise and artifacts are removed adequately. A 2-hour segment of the C3 EEG electrode in both raw and reconstructed sets is shown in Fig.1. A summary of the number of detected HFO before and after artifact removal is shown in Fig.2. The mean of the detected events has significantly decreased (p = 0.0009). By inspecting the identified events, many events falsely identified as HFO in the raw data were not detected in the reconstructed dataset at all (Fig.1).
Conclusion:
Although RippleLab is a powerful means of detecting EEG events such as HFO, a pipeline to remove signal artifacts before applying detection methods supports further research by decreasing the number of false positive HFO and processing time. We have shown that applying ICA in combination with our simple component rejection method is a sufficient way to remove artifacts from rodent EEG, reducing the need for manual detection and labeling. References: [1] Navarrete M, Alvarado-Rojas C, Le Van Quyen M, et al. RIPPLELAB: A Comprehensive Application for the Detection, Analysis and Classification of High Frequency Oscillations in EEG Signals. 2016. PLOS ONE 11(6) [2] Duncan, D., Vespa, P., Pitkänen, A., Braimah, A., Lapinlampi, N., & Toga, A. W. (2019). Big data sharing and analysis to advance research in post-traumatic epilepsy. Neurobiology of Disease, 123, 127–136. https://doi.org/10.1016/j.nbd.2018.05.026 [3] Delorme A, Makeig S. EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics. 2004. J Neurosci. Methods (134):9-21  
Funding:
:This research was supported by the National Institutes of Health under award numbers R01NS111744 and U54NS100064 (EpiBioS4Rx).