Part of the process of EEG microstate estimation involves clustering EEG channel data at the global field power (GFP) maxima, very commonly using a modified K-means approach. Clustering has also been done deterministically, despite there being uncertainties in multiple stages of the microstate analysis, including the GFP peak definition, the clustering itself and in the post-clustering assignment of microstates back onto the EEG timecourse of interest. We perform a fully probabilistic microstate clustering and labeling, to account for these sources of uncertainty using the closest probabilistic analogue to KM called Fuzzy C-means (FCM). We train softmax multi-layer perceptrons (MLPs) using the KM and FCM-inferred cluster assignments as target labels, to then allow for probabilistic labeling of the full EEG data instead of the usual correlation-based deterministic microstate label assignment typically used. We assess the merits of the probabilistic analysis vs the deterministic approaches in EEG data recorded while participants perform real or imagined motor movements from a publicly available data set of 109 subjects. Though FCM group template maps that are almost topographically identical to KM were found, there is considerable uncertainty in the subsequent assignment of microstate labels. In general, imagined motor movements are less predictable on a time point-by-time point basis, possibly reflecting the more exploratory nature of the brain state during imagined, compared to during real motor movements. We find that some relationships may be more evident using FCM than using KM and propose that future microstate analysis should preferably be performed probabilistically rather than deterministically, especially in situations such as with brain computer interfaces, where both training and applying models of microstates need to account for uncertainty. Probabilistic neural network-driven microstate assignment has a number of advantages that we have discussed, which are likely to be further developed and exploited in future studies. In conclusion, probabilistic clustering and a probabilistic neural network-driven approach to microstate analysis is likely to better model and reveal details and the variability hidden in current deterministic and binarized microstate assignment and analyses.
12 May 2016 in Frontiers in Computational Neuroscience
Dynamic time warping, or DTW, is a powerful and domain-general sequence alignment method for computing a similarity measure. Such dynamic programming-based techniques like DTW are now the backbone and driver of most bioinformatics methods and discoveries. In neuroscience it has had far less use, though this has begun to change. We wanted to explore new ways of applying DTW, not simply as a measure with which to cluster or compare similarity between features but in a conceptually different way. We have used DTW to provide a more interpretable spectral description of the data, compared to standard approaches such as the Fourier and related transforms. The DTW approach and standard discrete Fourier transform (DFT) are assessed against benchmark measures of neural dynamics. These include EEG microstates, EEG avalanches and the sum squared error (SSE) from a multilayer perceptron (MLP) prediction of the EEG timeseries, and simultaneously acquired FMRI BOLD signal. We explored the relationships between these variables of interest in an EEG-FMRI dataset acquired during a standard cognitive task, which allowed us to explore how DTW differentially performs in different task settings. We found that despite strong correlations between DTW and DFT-spectra, DTW was a better predictor for almost every measure of brain dynamics. Using these DTW measures, we show that predictability is almost always higher in task than in rest states, which is consistent to other theoretical and empirical findings, providing additional evidence for the utility of the DTW approach.
Local perturbations within complex dynamical systems can trigger cascade-like events that spread across significant portions of the system. Cascades of this type have been observed across a broad range of scales in the brain. Studies of these cascades, known as neuronal avalanches, usually report the statistics of large numbers of avalanches, without probing the characteristic patterns produced by the avalanches themselves. This is partly due to limitations in the extent or spatiotemporal resolution of commonly used neuroimaging techniques. In this study, we overcome these limitations by using optical voltage (genetically encoded voltage indicators) imaging. This allows us to record cortical activity in vivo across an entire cortical hemisphere, at both high spatial (~30um) and temporal (~20ms) resolution in mice that are either in an anesthetized or awake state. We then use artificial neural networks to identify the characteristic patterns created by neuronal avalanches in our data. The avalanches in the anesthetized cortex are most accurately classified by an artificial neural network architecture that simultaneously connects spatial and temporal information. This is in contrast with the awake cortex, in which avalanches are most accurately classified by an architecture that treats spatial and temporal information separately, due to the increased levels of spatiotemporal complexity. This is in keeping with reports of higher levels of spatiotemporal complexity in the awake brain coinciding with features of a dynamical system operating close to criticality.
Reinforcement learning (RL) is a general-purpose powerful machine learning framework within which we can model various deterministic, non-deterministic and complex environments. We applied RL to the problem of tracking and improving human sustained attention during a simple sustained attention to response task (SART) in a proof of concept study with two subjects, using state-of-the-art deep neural network-based RL in the form of Deep Q Networks (DQNs). While others have used RL in EEG settings previously, none have applied it in a neurofeedback (NFB) setting, which seems a natural problem within Brain Computer Interfaces (BCIs) to tackle using end-to-end RL in the form of DQNs, due to both the problem’s non-stationarity and the ability of RL to learn in a continuous setting. Furthermore, while many have explored phasic alerting previously, learning optimal alerting in a personalized way in real time is a less explored field, which we believe RL to be an most suitable solution for. First, we used empirically-derived simulated data of EEG and reaction times and subsequent parameter/algorithmic exploration within this simulated model to pick parameters for the DQN that are more likely to be optimal for the experimental setup and to explore the behavior of DQNs in this task setting. We then applied the method on two subjects and show that we get different but plausible results for both subjects, suggesting something about the behavior of DQNs in this setting. For this experimental part, we used parameters suggested to us by the simulation results. This RL-based behavioral- and neuro-feedback BCI method we have developed here is input feature agnostic and allows for complex continuous actions to be learned in other more complex closed-loop behavioral or neuro-feedback approaches.