Interpretable many-class decoding for MEG.

Multivariate pattern analysis (MVPA) of Magnetoencephalography (MEG) and Electroencephalography (EEG) data is a valuable tool for understanding how the brain represents and discriminates between different stimuli. Identifying the spatial and temporal signatures of stimuli is typically a crucial output of these analyses. Such analyses are mainly performed using linear, pairwise, sliding window decoding models. These allow for relative ease of interpretation, e.g. by estimating a time-course of decoding accuracy, but have limited decoding performance. On the other hand, full epoch multiclass decoding models, commonly used for brain-computer interface (BCI) applications, can provide better decoding performance. However interpretation methods for such models have been designed with a low number of classes in mind. In this paper, we propose an approach that combines a multiclass, full epoch decoding model with supervised dimensionality reduction, while still being able to reveal the contributions of spatiotemporal and spectral features using permutation feature importance. Crucially, we introduce a way of doing supervised dimensionality reduction of input features within a neural network optimised for the classification task, improving performance substantially. We demonstrate the approach on 3 different many-class task-MEG datasets using image presentations. Our results demonstrate that this approach consistently achieves higher accuracy than the peak accuracy of a sliding window decoder while estimating the relevant spatiotemporal features in the MEG signal.

Group-level brain decoding with deep learning.

Decoding brain imaging data are gaining popularity, with applications in brain-computer interfaces and the study of neural representations. Decoding is typically subject-specific and does not generalise well over subjects, due to high amounts of between subject variability. Techniques that overcome this will not only provide richer neuroscientific insights but also make it possible for group-level models to outperform subject-specific models. Here, we propose a method that uses subject embedding, analogous to word embedding in natural language processing, to learn and exploit the structure in between-subject variability as part of a decoding model, our adaptation of the WaveNet architecture for classification. We apply this to magnetoencephalography data, where 15 subjects viewed 118 different images, with 30 examples per image; to classify images using the entire 1 s window following image presentation. We show that the combination of deep learning and subject embedding is crucial to closing the performance gap between subject- and group-level decoding models. Importantly, group models outperform subject models on low-accuracy subjects (although slightly impair high-accuracy subjects) and can be helpful for initialising subject models. While we have not generally found group-level models to perform better than subject-level models, the performance of group modelling is expected to be even higher with bigger datasets. In order to provide physiological interpretation at the group level, we make use of permutation feature importance. This provides insights into the spatiotemporal and spectral information encoded in the models. All code is available on GitHub (https://github.com/ricsinaruto/MEG-group-decode).

The relationship between frequency content and representational dynamics in the decoding of neurophysiological data.

Decoding of high temporal resolution, stimulus-evoked neurophysiological data is increasingly used to test theories about how the brain processes information. However, a fundamental relationship between the frequency spectra of the neural signal and the subsequent decoding accuracy timecourse is not widely recognised. We show that, in commonly used instantaneous signal decoding paradigms, each sinusoidal component of the evoked response is translated to double its original frequency in the subsequent decoding accuracy timecourses. We therefore recommend, where researchers use instantaneous signal decoding paradigms, that more aggressive low pass filtering is applied with a cut-off at one quarter of the sampling rate, to eliminate representational alias artefacts. However, this does not negate the accompanying interpretational challenges. We show that these can be resolved by decoding paradigms that utilise both a signal's instantaneous magnitude and its local gradient information as features for decoding. On a publicly available MEG dataset, this results in decoding accuracy metrics that are higher, more stable over time, and free of the technical and interpretational challenges previously characterised. We anticipate that a broader awareness of these fundamental relationships will enable stronger interpretations of decoding results by linking them more clearly to the underlying signal characteristics that drive them.

Phasic modulation of visual representations during sustained attention.

Sustained attention has long been thought to benefit perception in a continuous fashion, but recent evidence suggests that it affects perception in a discrete, rhythmic way. Periodic fluctuations in behavioral performance over time, and modulations of behavioral performance by the phase of spontaneous oscillatory brain activity point to an attentional sampling rate in the theta or alpha frequency range. We investigated whether such discrete sampling by attention is reflected in periodic fluctuations in the decodability of visual stimulus orientation from magnetoencephalographic (MEG) brain signals. In this exploratory study, human subjects attended one of the two grating stimuli, while MEG was being recorded. We assessed the strength of the visual representation of the attended stimulus using a support vector machine (SVM) to decode the orientation of the grating (clockwise vs. counterclockwise) from the MEG signal. We tested whether decoder performance depended on the theta/alpha phase of local brain activity. While the phase of ongoing activity in the visual cortex did not modulate decoding performance, theta/alpha phase of activity in the frontal eye fields and parietal cortex, contralateral to the attended stimulus did modulate decoding performance. These findings suggest that phasic modulations of visual stimulus representations in the brain are caused by frequency-specific top-down activity in the frontoparietal attention network, though the behavioral relevance of these effects could not be established.

Investigating the effects of pre-stimulus cortical oscillatory activity on behavior.

Rhythmic brain activity may reflect a functional mechanism that facilitates cortical processing and dynamic interareal interactions and thereby give rise to complex behavior. Using magnetoencephalography (MEG), we investigated rhythmic brain activity in a brain-wide network and their relation to behavior, while human subjects executed a variant of the Simon task, a simple stimulus-response task with well-studied behavioral effects. We hypothesized that the faster reaction times (RT) on stimulus-response congruent versus incongruent trials are associated with oscillatory power changes, reflecting a change in local cortical activation. Additionally, we hypothesized that the faster reaction times for trials following instances with the same stimulus-response contingency (the so-called Gratton effect) is related to contingency-induced changes in the state of the network, as measured by differences in local spectral power and interareal phase coherence. This would be achieved by temporarily upregulating the connectivity strength between behaviorally relevant network nodes. We identified regions-of-interest that differed in local synchrony during the response phase of the Simon task. Within this network, spectral power in none of the nodes in either of the studied frequencies was significantly different in the pre-cue window of the subsequent trial. Nor was there a significant difference in coherence between the task-relevant nodes that could explain the superior behavioral performance after compatible consecutive trials.

Stimulus-induced gamma power predicts the amplitude of the subsequent visual evoked response.

The efficiency of neuronal information transfer in activated brain networks may affect behavioral performance. Gamma-band synchronization has been proposed to be a mechanism that facilitates neuronal processing of behaviorally relevant stimuli. In line with this, it has been shown that strong gamma-band activity in visual cortical areas leads to faster responses to a visual go cue. We investigated whether there are directly observable consequences of trial-by-trial fluctuations in non-invasively observed gamma-band activity on the neuronal response. Specifically, we hypothesized that the amplitude of the visual evoked response to a go cue can be predicted by gamma power in the visual system, in the window preceding the evoked response. Thirty-three human subjects (22 female) performed a visual speeded response task while their magnetoencephalogram (MEG) was recorded. The participants had to respond to a pattern reversal of a concentric moving grating. We estimated single trial stimulus-induced visual cortical gamma power, and correlated this with the estimated single trial amplitude of the most prominent event-related field (ERF) peak within the first 100 ms after the pattern reversal. In parieto-occipital cortical areas, the amplitude of the ERF correlated positively with gamma power, and correlated negatively with reaction times. No effects were observed for the alpha and beta frequency bands, despite clear stimulus onset induced modulation at those frequencies. These results support a mechanistic model, in which gamma-band synchronization enhances the neuronal gain to relevant visual input, thus leading to more efficient downstream processing and to faster responses.

osl-dynamics, a toolbox for modeling fast dynamic brain activity.

Neural activity contains rich spatiotemporal structure that corresponds to cognition. This includes oscillatory bursting and dynamic activity that span across networks of brain regions, all of which can occur on timescales of tens of milliseconds. While these processes can be accessed through brain recordings and imaging, modeling them presents methodological challenges due to their fast and transient nature. Furthermore, the exact timing and duration of interesting cognitive events are often a priori unknown. Here, we present the OHBA Software Library Dynamics Toolbox (osl-dynamics), a Python-based package that can identify and describe recurrent dynamics in functional neuroimaging data on timescales as fast as tens of milliseconds. At its core are machine learning generative models that are able to adapt to the data and learn the timing, as well as the spatial and spectral characteristics, of brain activity with few assumptions. osl-dynamics incorporates state-of-the-art approaches that can be, and have been, used to elucidate brain dynamics in a wide range of data types, including magneto/electroencephalography, functional magnetic resonance imaging, invasive local field potential recordings, and electrocorticography. It also provides novel summary measures of brain dynamics that can be used to inform our understanding of cognition, behavior, and disease. We hope osl-dynamics will further our understanding of brain function, through its ability to enhance the modeling of fast dynamic processes.