Post-task responses following working memory and movement are driven by transient spectral bursts with similar characteristics.

The post-movement beta rebound has been studied extensively using magnetoencephalography (MEG) and is reliably modulated by various task parameters as well as illness. Our recent study showed that rebounds, which we generalise as "post-task responses" (PTRs), are a ubiquitous phenomenon in the brain, occurring across the cortex in theta, alpha, and beta bands. Currently, it is unknown whether PTRs following working memory are driven by transient bursts, which are moments of short-lived high amplitude activity, similar to those that drive the post-movement beta rebound. Here, we use three-state univariate hidden Markov models (HMMs), which can identify bursts without a priori knowledge of frequency content or response timings, to compare bursts that drive PTRs in working memory and visuomotor MEG datasets. Our results show that PTRs across working memory and visuomotor tasks are driven by pan-spectral transient bursts. These bursts have very similar spectral content variation over the cortex, correlating strongly between the two tasks in the alpha (R2 = .89) and beta (R2 = .53) bands. Bursts also have similar variation in duration over the cortex (e.g., long duration bursts occur in the motor cortex for both tasks), strongly correlating over cortical regions between tasks (R2 = .56), with a mean over all regions of around 300 ms in both datasets. Finally, we demonstrate the ability of HMMs to isolate signals of interest in MEG data, such that the HMM probability timecourse correlates more strongly with reaction times than frequency filtered power envelopes from the same brain regions. Overall, we show that induced PTRs across different tasks are driven by bursts with similar characteristics, which can be identified using HMMs. Given the similarity between bursts across tasks, we suggest that PTRs across the cortex may be driven by a common underlying neural phenomenon.

Dissecting unsupervised learning through hidden Markov modeling in electrophysiological data.

Unsupervised, data-driven methods are commonly used in neuroscience to automatically decompose data into interpretable patterns. These patterns differ from one another depending on the assumptions of the models. How these assumptions affect specific data decompositions in practice, however, is often unclear, which hinders model applicability and interpretability. For instance, the hidden Markov model (HMM) automatically detects characteristic, recurring activity patterns (so-called states) from time series data. States are defined by a certain probability distribution, whose state-specific parameters are estimated from the data. But what specific features, from all of those that the data contain, do the states capture? That depends on the choice of probability distribution and on other model hyperparameters. Using both synthetic and real data, we aim to better characterize the behavior of two HMM types that can be applied to electrophysiological data. Specifically, we study which differences in data features (such as frequency, amplitude, or signal-to-noise ratio) are more salient to the models and therefore more likely to drive the state decomposition. Overall, we aim at providing guidance for the appropriate use of this type of analysis on one- or two-channel neural electrophysiological data and an informed interpretation of its results given the characteristics of the data and the purpose of the analysis.NEW & NOTEWORTHY Compared with classical supervised methods, unsupervised methods of analysis have the advantage to be freer of subjective biases. However, it is not always clear what aspects of the data these methods are most sensitive to, which complicates interpretation. Focusing on the hidden Markov model, commonly used to describe electrophysiological data, we explore in detail the nature of its estimates through simulations and real data examples, providing important insights about what to expect from these models.

Mapping Interictal activity in epilepsy using a hidden Markov model: A magnetoencephalography study.

Epilepsy is a highly heterogeneous neurological disorder with variable etiology, manifestation, and response to treatment. It is imperative that new models of epileptiform brain activity account for this variability, to identify individual needs and allow clinicians to curate personalized care. Here, we use a hidden Markov model (HMM) to create a unique statistical model of interictal brain activity for 10 pediatric patients. We use magnetoencephalography (MEG) data acquired as part of standard clinical care for patients at the Children's Hospital of Philadelphia. These data are routinely analyzed using excess kurtosis mapping (EKM); however, as cases become more complex (extreme multifocal and/or polymorphic activity), they become harder to interpret with EKM. We assessed the performance of the HMM against EKM for three patient groups, with increasingly complicated presentation. The difference in localization of epileptogenic foci for the two methods was 7 ± 2 mm (mean ± SD over all 10 patients); and 94% ± 13% of EKM temporal markers were matched by an HMM state visit. The HMM localizes epileptogenic areas (in agreement with EKM) and provides additional information about the relationship between those areas. A key advantage over current methods is that the HMM is a data-driven model, so the output is tuned to each individual. Finally, the model output is intuitive, allowing a user (clinician) to review the result and manually select the HMM epileptiform state, offering multiple advantages over previous methods and allowing for broader implementation of MEG epileptiform analysis in surgical decision-making for patients with intractable epilepsy.

Dynamic analysis on simultaneous iEEG-MEG data via hidden Markov model.

BACKGROUND: Intracranial electroencephalography (iEEG) recordings are used for clinical evaluation prior to surgical resection of the focus of epileptic seizures and also provide a window into normal brain function. A major difficulty with interpreting iEEG results at the group level is inconsistent placement of electrodes between subjects making it difficult to select contacts that correspond to the same functional areas. Recent work using time delay embedded hidden Markov model (HMM) applied to magnetoencephalography (MEG) resting data revealed a distinct set of brain states with each state engaging a specific set of cortical regions. Here we use a rare group dataset with simultaneously acquired resting iEEG and MEG to test whether there is correspondence between HMM states and iEEG power changes that would allow classifying iEEG contacts into functional clusters. METHODS: Simultaneous MEG-iEEG recordings were performed at rest on 11 patients with epilepsy whose intracranial electrodes were implanted for pre-surgical evaluation. Pre-processed MEG sensor data was projected to source space. Time delay embedded HMM was then applied to MEG time series. At the same time, iEEG time series were analyzed with time-frequency decomposition to obtain spectral power changes with time. To relate MEG and iEEG results, correlations were computed between HMM probability time courses of state activation and iEEG power time course from the mid contact pair for each electrode in equally spaced frequency bins and presented as correlation spectra for the respective states and iEEG channels. Association of iEEG electrodes with HMM states based on significant correlations was compared to that based on the distance to peaks in subject-specific state topographies. RESULTS: Five HMM states were inferred from MEG. Two of them corresponded to the left and the right temporal activations and had a spectral signature primarily in the theta/alpha frequency band. All the electrodes had significant correlations with at least one of the states (p < 0.05 uncorrected) and for 27/50 electrodes these survived within-subject FDR correction (q < 0.05). These correlations peaked in the theta/alpha band. There was a highly significant dependence between the association of states and electrodes based on functional correlations and that based on spatial proximity (p = 5.6e-6,χ2 test for independence). Despite the potentially atypical functional anatomy and physiological abnormalities related to epilepsy, HMM model estimated from the patient group was very similar to that estimated from healthy subjects. CONCLUSION: Epilepsy does not preclude HMM analysis of interictal data. The resulting group functional states are highly similar to those reported for healthy controls. Power changes recorded with iEEG correlate with HMM state time courses in the alpha-theta band and the presence of this correlation can be related to the spatial location of electrode contacts close to the individual peaks of the corresponding state topographies. Thus, the hypothesized relation between iEEG contacts and HMM states exists and HMM could be further explored as a method for identifying comparable iEEG channels across subjects for the purposes of group analysis.

The psychological correlates of distinct neural states occurring during wakeful rest.

When unoccupied by an explicit external task, humans engage in a wide range of different types of self-generated thinking. These are often unrelated to the immediate environment and have unique psychological features. Although contemporary perspectives on ongoing thought recognise the heterogeneity of these self-generated states, we lack both a clear understanding of how to classify the specific states, and how they can be mapped empirically. In the current study, we capitalise on advances in machine learning that allow continuous neural data to be divided into a set of distinct temporally re-occurring patterns, or states. We applied this technique to a large set of resting state data in which we also acquired retrospective descriptions of the participants' experiences during the scan. We found that two of the identified states were predictive of patterns of thinking at rest. One state highlighted a pattern of neural activity commonly seen during demanding tasks, and the time individuals spent in this state was associated with descriptions of experience focused on problem solving in the future. A second state was associated with patterns of activity that are commonly seen under less demanding conditions, and the time spent in it was linked to reports of intrusive thoughts about the past. Finally, we found that these two neural states tended to fall at either end of a neural hierarchy that is thought to reflect the brain's response to cognitive demands. Together, these results demonstrate that approaches which take advantage of time-varying changes in neural function can play an important role in understanding the repertoire of self-generated states. Moreover, they establish that important features of self-generated ongoing experience are related to variation along a similar vein to those seen when the brain responds to cognitive task demands.

Functional network dynamics in a neurodevelopmental disorder of known genetic origin.

Dynamic connectivity in functional brain networks is a fundamental aspect of cognitive development, but we have little understanding of the mechanisms driving variability in these networks. Genes are likely to influence the emergence of fast network connectivity via their regulation of neuronal processes, but novel methods to capture these rapid dynamics have rarely been used in genetic populations. The current study redressed this by investigating brain network dynamics in a neurodevelopmental disorder of known genetic origin, by comparing individuals with a ZDHHC9-associated intellectual disability to individuals with no known impairment. We characterised transient network dynamics using a Hidden Markov Model (HMM) on magnetoencephalography (MEG) data, at rest and during auditory oddball stimulation. The HMM is a data-driven method that captures rapid patterns of coordinated brain activity recurring over time. Resting-state network dynamics distinguished the groups, with ZDHHC9 participants showing longer state activation and, crucially, ZDHHC9 gene expression levels predicted the group differences in dynamic connectivity across networks. In contrast, network dynamics during auditory oddball stimulation did not show this association. We demonstrate a link between regional gene expression and brain network dynamics, and present the new application of a powerful method for understanding the neural mechanisms linking genetic variation to cognitive difficulties.

Unpacking Transient Event Dynamics in Electrophysiological Power Spectra.

Electrophysiological recordings of neuronal activity show spontaneous and task-dependent changes in their frequency-domain power spectra. These changes are conventionally interpreted as modulations in the amplitude of underlying oscillations. However, this overlooks the possibility of underlying transient spectral 'bursts' or events whose dynamics can map to changes in trial-average spectral power in numerous ways. Under this emerging perspective, a key challenge is to perform burst detection, i.e. to characterise single-trial transient spectral events, in a principled manner. Here, we describe how transient spectral events can be operationalised and estimated using Hidden Markov Models (HMMs). The HMM overcomes a number of the limitations of the standard amplitude-thresholding approach to burst detection; in that it is able to concurrently detect different types of bursts, each with distinct spectral content, without the need to predefine frequency bands of interest, and does so with less dependence on a priori threshold specification. We describe how the HMM can be used for burst detection and illustrate its benefits on simulated data. Finally, we apply this method to empirical data to detect multiple burst types in a task-MEG dataset, and illustrate how we can compute burst metrics, such as the task-evoked timecourse of burst duration.

Altered transient brain dynamics in multiple sclerosis: Treatment or pathology?

Multiple sclerosis (MS) is a demyelinating, neuroinflammatory, and -degenerative disease that affects the brain's neurophysiological functioning through brain atrophy, a reduced conduction velocity and decreased connectivity. Currently, little is known on how MS affects the fast temporal dynamics of activation and deactivation of the different large-scale, ongoing brain networks. In this study, we investigated whether these temporal dynamics are affected in MS patients and whether these changes are induced by the pathology or by the use of benzodiazepines (BZDs), an important symptomatic treatment that aims at reducing insomnia, spasticity and anxiety and reinforces the inhibitory effect of GABA. To this aim, we employed a novel method capable of detecting these fast dynamics in 90 MS patients and 46 healthy controls. We demonstrated a less dynamic frontal default mode network in male MS patients and a reduced activation of the same network in female MS patients, regardless of BZD usage. Additionally, BZDs strongly altered the brain's dynamics by increasing the time spent in the deactivating sensorimotor network and the activating occipital network. Furthermore, BZDs induced a decreased power in the theta band and an increased power in the beta band. The latter was strongly expressed in those states without activation of the sensorimotor network. In summary, we demonstrate gender-dependent changes to the brain dynamics in the frontal DMN and strong effects from BZDs. This study is the first to characterise the effect of multiple sclerosis and BZDs in vivo in a spatially, temporally and spectrally defined way.

Task-Evoked Dynamic Network Analysis Through Hidden Markov Modeling.

Complex thought and behavior arise through dynamic recruitment of large-scale brain networks. The signatures of this process may be observable in electrophysiological data; yet robust modeling of rapidly changing functional network structure on rapid cognitive timescales remains a considerable challenge. Here, we present one potential solution using Hidden Markov Models (HMMs), which are able to identify brain states characterized by engaging distinct functional networks that reoccur over time. We show how the HMM can be inferred on continuous, parcellated source-space Magnetoencephalography (MEG) task data in an unsupervised manner, without any knowledge of the task timings. We apply this to a freely available MEG dataset in which participants completed a face perception task, and reveal task-dependent HMM states that represent whole-brain dynamic networks transiently bursting at millisecond time scales as cognition unfolds. The analysis pipeline demonstrates a general way in which the HMM can be used to do a statistically valid whole-brain, group-level task analysis on MEG task data, which could be readily adapted to a wide range of task-based studies.

Spontaneous cortical activity transiently organises into frequency specific phase-coupling networks.

Frequency-specific oscillations and phase-coupling of neuronal populations are essential mechanisms for the coordination of activity between brain areas during cognitive tasks. Therefore, the ongoing activity ascribed to the different functional brain networks should also be able to reorganise and coordinate via similar mechanisms. We develop a novel method for identifying large-scale phase-coupled network dynamics and show that resting networks in magnetoencephalography are well characterised by visits to short-lived transient brain states, with spatially distinct patterns of oscillatory power and coherence in specific frequency bands. Brain states are identified for sensory, motor networks and higher-order cognitive networks. The cognitive networks include a posterior alpha (8-12 Hz) and an anterior delta/theta range (1-7 Hz) network, both exhibiting high power and coherence in areas that correspond to posterior and anterior subdivisions of the default mode network. Our results show that large-scale cortical phase-coupling networks have characteristic signatures in very specific frequency bands, possibly reflecting functional specialisation at different intrinsic timescales.

Discovering dynamic brain networks from big data in rest and task.

Brain activity is a dynamic combination of the responses to sensory inputs and its own spontaneous processing. Consequently, such brain activity is continuously changing whether or not one is focusing on an externally imposed task. Previously, we have introduced an analysis method that allows us, using Hidden Markov Models (HMM), to model task or rest brain activity as a dynamic sequence of distinct brain networks, overcoming many of the limitations posed by sliding window approaches. Here, we present an advance that enables the HMM to handle very large amounts of data, making possible the inference of very reproducible and interpretable dynamic brain networks in a range of different datasets, including task, rest, MEG and fMRI, with potentially thousands of subjects. We anticipate that the generation of large and publicly available datasets from initiatives such as the Human Connectome Project and UK Biobank, in combination with computational methods that can work at this scale, will bring a breakthrough in our understanding of brain function in both health and disease.

Fusion of magnetometer and gradiometer sensors of MEG in the presence of multiplicative error.

Novel neuroimaging techniques have provided unprecedented information on the structure and function of the living human brain. Multimodal fusion of data from different sensors promises to radically improve this understanding, yet optimal methods have not been developed. Here, we demonstrate a novel method for combining multichannel signals. We show how this method can be used to fuse signals from the magnetometer and gradiometer sensors used in magnetoencephalography (MEG), and through extensive experiments using simulation, head phantom and real MEG data, show that it is both robust and accurate. This new approach works by assuming that the lead fields have multiplicative error. The criterion to estimate the error is given within a spatial filter framework such that the estimated power is minimized in the worst case scenario. The method is compared to, and found better than, existing approaches. The closed-form solution and the conditions under which the multiplicative error can be optimally estimated are provided. This novel approach can also be employed for multimodal fusion of other multichannel signals such as MEG and EEG. Although the multiplicative error is estimated based on beamforming, other methods for source analysis can equally be used after the lead-field modification.

Assessment of arterial arrival times derived from multiple inversion time pulsed arterial spin labeling MRI.

The purpose of this study was to establish a normal range for the arterial arrival time (AAT) in whole-brain pulsed arterial spin labeling (PASL) cerebral perfusion MRI. Healthy volunteers (N = 36, range: 20 to 35 years) provided informed consent to participate in this study. AAT was assessed in multiple brain regions, using three-dimensional gradient and spin echo (GRASE) pulsed arterial spin labeling at 3.0 T, and found to be 641 +/- 95, 804 +/- 91, 802 +/- 126, and 935 +/- 108 ms in the temporal, parietal, frontal, and occipital lobes, respectively. Mean gray matter AAT was found to be 694 +/- 89 ms for females (N = 15), which was significantly shorter than for men, 814 +/- 192 ms (N = 21; P < 0.0003), and significant after correcting for brain volume (P < 0.001). Significant AAT sex differences were also found using voxelwise permutation testing. An atlas of AAT values across the healthy brain is presented here and may be useful for future experiments that aim to quantify cerebral blood flow from ASL data, as well as for clinical comparisons where disease pathology may lead to altered AAT. Pulsed arterial spin labeling signals were simulated using an identical sampling scheme as the empiric study and revealed AAT can be estimated robustly when simulated arrival times are well beyond the normal range.

Multilevel linear modelling for FMRI group analysis using Bayesian inference.

Functional magnetic resonance imaging studies often involve the acquisition of data from multiple sessions and/or multiple subjects. A hierarchical approach can be taken to modelling such data with a general linear model (GLM) at each level of the hierarchy introducing different random effects variance components. Inferring on these models is nontrivial with frequentist solutions being unavailable. A solution is to use a Bayesian framework. One important ingredient in this is the choice of prior on the variance components and top-level regression parameters. Due to the typically small numbers of sessions or subjects in neuroimaging, the choice of prior is critical. To alleviate this problem, we introduce to neuroimage modelling the approach of reference priors, which drives the choice of prior such that it is noninformative in an information-theoretic sense. We propose two inference techniques at the top level for multilevel hierarchies (a fast approach and a slower more accurate approach). We also demonstrate that we can infer on the top level of multilevel hierarchies by inferring on the levels of the hierarchy separately and passing summary statistics of a noncentral multivariate t distribution between them.

Constrained linear basis sets for HRF modelling using Variational Bayes.

FMRI modelling requires flexible haemodynamic response function (HRF) modelling, with the HRF being allowed to vary spatially and between subjects. To achieve this flexibility, voxelwise parameterised HRFs have been proposed; however, inference on such models is very slow. An alternative approach is to use basis functions allowing inference to proceed in the more manageable General Linear Model (GLM) framework. However, a large amount of the subspace spanned by the basis functions produces nonsensical HRF shapes. In this work we propose a technique for choosing a basis set, and then the means to constrain the subspace spanned by the basis set to only include sensible HRF shapes. Penny et al. showed how Variational Bayes can be used to infer on the GLM for FMRI. Here we extend the work of Penny et al. to give inference on the GLM with constrained HRF basis functions and with spatial Markov Random Fields on the autoregressive noise parameters. Constraining the subspace spanned by the basis set allows for far superior separation of activating voxels from nonactivating voxels in FMRI data. We use spatial mixture modelling to produce final probabilities of activation and demonstrate increased sensitivity on an FMRI dataset.