Machine learning of brain-specific biomarkers from EEG

Abstract

Electroencephalography (EEG) has a long history as a clinical tool to study brain function, and its potential to derive biomarkers for various applications is far from exhausted. Machine learning (ML) can guide future innovation by harnessing the wealth of complex EEG signals to isolate relevant brain activity. Yet, ML studies in EEG tend to ignore physiological artifacts, which may cause problems for deriving biomarkers specific to the central nervous system (CNS). We present a framework for conceptualizing machine learning from CNS versus peripheral signals measured with EEG. A common signal representation across the frequency spectrum based on Morlet wavelets allowed us to define traditional brain activity features (e.g. log power) and alternative inputs used by state-of-the-art ML approaches (covariance matrices). Using more than 2600 EEG recordings from large public databases (TUAB, TDBRAIN), we studied the impact of peripheral signals and artifact removal techniques on ML models in exemplary age and sex prediction analyses. Across benchmarks, basic artifact rejection improved model performance whereas further removal of peripheral signals using ICA decreased performance. Our analyses revealed that peripheral signals enable age and sex prediction. However, they explained only a fraction of the performance provided by brain signals. We show that brain signals and body signals, both reflected in the EEG, allow for prediction of personal characteristics. While these results may depend on specific prediction problems, our work suggests that great care is needed to separate these signals when the goal is to develop CNS-specific biomarkers using ML.

Exploring the neuromagnetic signatures of cognitive decline from mild cognitive impairment to Alzheimer’s disease dementia

“…We computed spectral features with Morlet wavelets (Hipp et al, 2012) using the meeglet library (Bomatter et al, 2023). This wavelet approach implements Morlet wavelets spaced on a logarithmic grid such that the spacing between wavelets and their spectral smoothness increase log-linearly with frequency (Bomatter et al, 2023;Hipp et al, 2012).…”

“…We computed spectral features with Morlet wavelets (Hipp et al, 2012) using the meeglet library (Bomatter et al, 2023). This wavelet approach implements Morlet wavelets spaced on a logarithmic grid such that the spacing between wavelets and their spectral smoothness increase log-linearly with frequency (Bomatter et al, 2023;Hipp et al, 2012). Such wavelets are well suited for capturing the log-dynamic frequency scaling of brain activity (Buzsáki & Mizuseki, 2014) and have been proven useful in multiple EEG-biomarker applications (Frohlich et al, 2019;Hawellek et al, 2022;Hipp et al, 2021).…”

“…On the other hand, the dwPLI metric captures changes in phasesynchronization with reduced sensitivity to uncorrelated noise sources and increased statistical power to detect changes in phase-synchronization compared to PLI (Vinck et al, 2011). Power envelope correlation can detect non-instantaneous power correlationsregardless of their sign -which has been used to study synchronised signal amplitude changes between distant brain regions (Hipp et al, 2012), The power envelope was defined as the log of the rectified wavelet-convoluted signal (Bomatter et al, 2023;Hipp et al, 2012).…”

“…This procedure conducts non-parametric distance MANOVA to test for group differences (Anderson, 2001). As a distance, we used the Riemannian affine invariant distance which has been used with great success in different machine learning applications for MEG and EEG including brain computer interfaces, brain-age prediction and sex classification (Barachant et al, 2012;Bomatter et al, 2023;Congedo et al, 2017;Sabbagh et al, 2020). Theoretical analysis and empirical benchmarks have shown that Riemannian metrics mitigate distortions due to electromagnetic field spread and provide latent representations well suited to statistically isolate information related to cortical current generators (Sabbagh et al, 2019(Sabbagh et al, , 2020.…”

“…Furthermore, pathology and medical treatments may alter neural dynamics at frequencies that are not well represented by standard frequency bands (Jobert & Wilson, 2015). This motivates approaches that analyze the frequency spectrum continuously (Hawellek et al, 2022;Hipp et al, 2012), particularly since different pathological conditions can modulate electrophysiological signals at different spatial scales (Bomatter et al, 2023). For example, oscillatory activity in the alpha band (~10Hz) tends to be maximal over posterior visual cortices, while that in the beta band (~20Hz) tends to be maximal over motor cortex (Bourguignon et al, 2019).…”