Abstracts

Rationale:
Computational analysis of standard clinical EEG data is limited by abundant muscle, movement, and electrode artifacts, especially in the awake state. The mitigation of such artifacts is labor intensive and introduces methodologic risk, such as selection bias that may occur when a human electroencephalographer manually selects “artifact-free” samples for analysis. With goal of creating a stable preprocessing pipeline which can facilitate fully-automated processing of clinically-acquired EEG, we set out to design and validate a preprocessing algorithm using EEG data from children with epileptic spasms and normal controls, to obtain phase-amplitude coupling measures.
Method:
Total of 717 19-channel EEG samples recorded from both children with and without infantile spasms (cases and controls, respectively) were processed. These datasets were imported into EEGLAB14 running under Matlab 2017b. The three PAC measures were obtained as follows: (1) After applying FIR high-pass filter at 0.5 Hz and CleanLine() for removing line noise, the first PAC measure (Canolty’s modulation index, MI) was calculated (phase 3-4 Hz, amplitude 35-70 Hz). (2) clean_rawdata() was applied that included artifact subspace reconstruction (ASR), which is a solution for denoising for multivariate data using sliding-window PCA. Informax ICA was performed to obtain independent components (ICs), which were evaluated with ICLabel() to generate probabilistic labels. To evaluate the optimal level of data denoising, we prepared the three levels of cleaning: Raw, Level 1, and Level 2. For the Level 1 cleaning, IC exclusion criteria were label probability > 0.8 for either Eye, Muscle, or Heart. After rejecting those ICs, the remaining ICs were backprojected to scalp electrodes for the second PAC calculation. (3) For Level 2, IC rejection was redone using different label probability, < 0.7 for Brain, then the third PAC measure was calculated. As a validation procedure, we evaluated the number of ICs rejected and how well gamma-delta modulation indices (MI) could distinguish cases from controls.
Results:
Computation time was 3 days using 3 computers. The mean number of ICs remained were 21 (SD 0, range 21-21), 13.5 (SD 3.3, range 3-20), and 7.3 (SD 3.9, range 1-21) for Raw, Level 1, and Level 2, respectively. The PAC results revealed that scalp distribution of MI was surprisingly consistent across the denoising conditions (see Figure 1). Level 2 showed 51/717 datasets were with only < = 2 ICs which was judged to be suboptimal. The Raw condition showed MI increase at the frontal pole region, which may be due to blinks artifact.  In the validation, MI derived from unprocessed awake EEG samples did not distinguish cases from controls. In contrast, both level 1 and level 2 preprocessing greatly improved the utility to distinguish cases from controls using awake EEG data (see Figure 2).
Conclusion:
This is a successful demonstration of an automated pipeline for preprocessing scalp-recorded EEG data for phase-amplitude coupling analysis. Further study is needed to optimize these algorithms for use with other computational measures and distinct patient populations.
Funding:
:This study was accomplished with support from the Swartz Center at the University of California, San Diego, Susan S. Spencer Clinical Research Training Scholarship from the American Academy of Neurology, the Pediatric Victory Foundation, the Sudha Neelakantan & Venky Harinarayan Charitable Fund, the Elsie and Isaac Fogelman Endowment, the Mohammed F. Alibrahim Endowment, the Hughes Family Foundation, the John C. Hench Foundation, the University of California Los Angeles Children’s Discovery and Innovation Institute, and UCB Biopharma.