Machine learning of brain-specific biomarkers from EEG.

BACKGROUND: 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 artefacts, which may cause problems for deriving biomarkers specific to the central nervous system (CNS). METHODS: We present a framework for conceptualising machine learning from CNS versus peripheral signals measured with EEG. A signal representation 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 based on covariance matrices. Using more than 2600 EEG recordings from large public databases (TUAB, TDBRAIN), we studied the impact of peripheral signals and artefact removal techniques on ML models in age and sex prediction analyses. FINDINGS: Across benchmarks, basic artefact 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. INTERPRETATION: We show that brain signals and body signals, both present in the EEG, allow for prediction of personal characteristics. While these results may depend on specific applications, our work suggests that great care is needed to separate these signals when the goal is to develop CNS-specific biomarkers using ML. FUNDING: All authors have been working for F. Hoffmann-La Roche Ltd.

Effects of Ketamine and Midazolam on Simultaneous EEG/fMRI Data During Working Memory Processes.

Reliable measures of cognitive brain activity from functional neuroimaging techniques may provide early indications of efficacy in clinical trials. Functional magnetic resonance imaging and electroencephalography provide complementary spatiotemporal information and simultaneous recording of these two modalities can remove inter-session drug response and environment variability. We sought to assess the effects of ketamine and midazolam on simultaneous electrophysiological and hemodynamic recordings during working memory (WM) processes. Thirty participants were included in a placebo-controlled, three-way crossover design with ketamine and midazolam. Compared to placebo, ketamine administration attenuated theta power increases and alpha power decreases and midazolam attenuated low beta band decreases to increasing WM load. Additionally, ketamine caused larger blood-oxygen-dependent (BOLD) signal increases in the supplementary motor area and angular gyrus, and weaker deactivations of the default mode network (DMN), whereas no difference was found between midazolam and placebo. Ketamine administration caused positive temporal correlations between frontal-midline theta (fm-theta) power and the BOLD signal to disappear and attenuated negative correlations. However, the relationship between fm-theta and the BOLD signal from DMN areas was maintained in some participants during ketamine administration, as increasing theta strength was associated with stronger BOLD signal reductions in these areas. The presence of, and ability to manipulate, both positive and negative associations between the BOLD signal and fm-theta suggest the presence of multiple fm-theta components involved in WM processes, with ketamine administration disrupting one or more of these theta-linked WM strategies.

Modulation of simultaneously collected hemodynamic and electrophysiological functional connectivity by ketamine and midazolam.

The pharmacological modulation of functional connectivity in the brain may underlie therapeutic efficacy for several neurological and psychiatric disorders. Functional magnetic resonance imaging (fMRI) provides a noninvasive method of assessing this modulation, however, the indirect nature of the blood-oxygen level dependent signal restricts the discrimination of neural from physiological contributions. Here we followed two approaches to assess the validity of fMRI functional connectivity in developing drug biomarkers, using simultaneous electroencephalography (EEG)/fMRI in a placebo-controlled, three-way crossover design with ketamine and midazolam. First, we compared seven different preprocessing pipelines to determine their impact on the connectivity of common resting-state networks. Independent components analysis (ICA)-denoising resulted in stronger reductions in connectivity after ketamine, and weaker increases after midazolam, than pipelines employing physiological noise modelling or averaged signals from cerebrospinal fluid or white matter. This suggests that pipeline decisions should reflect a drug's unique noise structure, and if this is unknown then accepting possible signal loss when choosing extensive ICA denoising pipelines could engender more confidence in the remaining results. We then compared the temporal correlation structure of fMRI to that derived from two connectivity metrics of EEG, which provides a direct measure of neural activity. While electrophysiological estimates based on the power envelope were more closely aligned to BOLD signal connectivity than those based on phase consistency, no significant relationship between the change in electrophysiological and hemodynamic correlation structures was found, implying caution should be used when making cross-modal comparisons of pharmacologically-modulated functional connectivity.

Spindle activity phase-locked to sleep slow oscillations.

The <1Hz slow oscillation (SO) and spindles are hallmarks of mammalian non-rapid eye movement and slow wave sleep. Spindle activity occurring phase-locked to the SO is considered a candidate mediator of memory consolidation during sleep. We used source localization of magnetoencephalographic (MEG) and electroencephalographic (EEG) recordings from 11 sleeping human subjects for an in-depth analysis of the temporal and spatial properties of sleep spindles co-occurring with SOs. Slow oscillations and spindles were identified in the EEG and related to the MEG signal, providing enhanced spatial resolution. In the temporal domain, we confirmed a phase-locking of classical 12-15Hz fast spindle activity to the depolarizing SO up-state and of 9-12Hz slow spindle activity to the up-to-down-state transition of the SO. In the spatial domain, we show a broad spread of spindle activity, with less distinct anterior-posterior separation of fast and slow spindles than commonly seen in the EEG. We further tested a prediction of current memory consolidation models, namely the existence of a spatial bias of SOs over sleep spindles as a mechanism to promote localized neuronal synchronization and plasticity. In contrast to that prediction, a comparison of SOs dominating over the left vs. right hemisphere did not reveal any signs of a concurrent lateralization of spindle activity co-occurring with these SOs. Our data are consistent with the concept of the neocortical SO exerting top-down control over thalamic spindle generation. However, they call into question the notion that SOs locally coordinate spindles and thereby inform spindle-related memory processing.

Learning viewpoint invariant object representations using a temporal coherence principle.

Invariant object recognition is arguably one of the major challenges for contemporary machine vision systems. In contrast, the mammalian visual system performs this task virtually effortlessly. How can we exploit our knowledge on the biological system to improve artificial systems? Our understanding of the mammalian early visual system has been augmented by the discovery that general coding principles could explain many aspects of neuronal response properties. How can such schemes be transferred to system level performance? In the present study we train cells on a particular variant of the general principle of temporal coherence, the "stability" objective. These cells are trained on unlabeled real-world images without a teaching signal. We show that after training, the cells form a representation that is largely independent of the viewpoint from which the stimulus is looked at. This finding includes generalization to previously unseen viewpoints. The achieved representation is better suited for view-point invariant object classification than the cells' input patterns. This property to facilitate view-point invariant classification is maintained even if training and classification take place in the presence of an--also unlabeled--distractor object. In summary, here we show that unsupervised learning using a general coding principle facilitates the classification of real-world objects, that are not segmented from the background and undergo complex, non-isomorphic, transformations.