2.4. EEG Feature Extraction

Profiles of electrophysiological features were derived from recorded EEG data using custom MatLab scripts as described in Keogh et al. 2018 [23]. All scripts used in this analysis are available at https://github.com/conorkeogh/IGF_Response/.

Power spectra were calculated using Fourier Transforms of preprocessed EEG data. Overall power, measured in decibels, was assessed using the mean of individual channel spectra. Individual frequency bands were assessed by isolating these bands and calculating relative power normalised to overall power [42].

Distribution of activity was assessed by evaluating the relative activity at each hemisphere. This was calculated by subtracting spectral power measures in the right hemisphere from those in the left. Detailed profiles of power distribution were derived by analysing power across the whole hemisphere, at each individual electrode and within frequency bands.

Network function was analysed by evaluating interactions between electrodes [43]. Measures of interelectrode coherence were derived for each electrode pair through calculation of the cross-spectrum of the channels normalised by the power spectra of both channels, repeated for each unique pair:

Dimensionality reduction of connectivity measures was performed via principal component analysis [44]. Network architectures were assessed and compared between groups by examining the covariance of interelectrode coherence measures and associated loadings of network measures (i.e., the principal components of the coherence measures) [45]. This allowed the overall architecture of cortical networks to be compared between groups while limiting the number of statistical tests required. Pairwise comparisons of individual electrode pairs were performed on a post hoc, exploratory basis to better characterise the nature of any differences seen in comparisons of principal components.

Patient recruitment and clinical evaluations were performed by personnel who were not involved in EEG analysis. Processing and feature extraction algorithms were applied to all recordings without consideration of clinical features, disease severity, treatment status, genetics or any other clinical parameter.

All data processing was performed blinded to patient characteristics.