Revealing Distinct Neural Signatures in Magnetoencephalography with Hidden Markov Models

Seedat, Zelekha Abid (2022) Revealing Distinct Neural Signatures in Magnetoencephalography with Hidden Markov Models. PhD thesis, University of Nottingham.

PDF (Thesis - as examined) - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Available under Licence Creative Commons Attribution.
Download (8MB) | Preview


Magnetoencephalography (MEG) is a functional neuroimaging method which measures the magnetic fields produced by neural communication in the brain. Specifically, the fields induced by dendritic current flow in assemblies of pyramidal neurons. Because these magnetic fields are generated directly by brain electrophysiology, and are mostly unperturbed by the skull, MEG data are rich in spatial and temporal information. This thesis is chiefly concerned with interpreting these data in a way that produces useful results whilst minimising bias.

Hidden Markov modelling (HMM) is a robust statistical method which has been applied to fields as diverse as speech recognition and financial market prediction. It parses data into a number of ‘hidden states’, each with their own unique characteristics, in an unsupervised way. Because it is data-driven, it can create a model unique to each participant’s brain activity and specific to each task. In addition, the HMM framework itself is flexible so it can be applied to both sensor and source-space data and can be applied to multiple channels (multivariate) or to a single time course (univariate). Choice of an observation model allows states to be characterised by amplitude, spatial, or spectral content depending on the research question.

The aim of this thesis is to apply hidden Markov modelling (HMM) to whole head MEG data to identify repeated patterns of transient neural activity occurring throughout the brain. Once these patterns were identified, the interaction between these short ‘bursts’ of activity across the cortex was established which provided a unique measure of functional connectivity.

Three studies were undertaken:

The role of transient spectral bursts in MEG functional connectivity: In recent years, the smoothly varying neural oscillations often studied in MEG (such as those trial-averaged responses in the traditional neurophysiological (such as alpha/beta) frequency bands) have been shown to be made up of single-trial high-amplitude ‘bursts’ of activity. These bursts can be observed in the beta frequency band and are therefore often referred to as beta bursts. In this study, a novel time-delay embedded HMM was used to identif bursts in broadband data based on their spectral content for MEG data from 66 healthy adult participants. The burst amplitude, duration and frequency of occurrence were characterised across the cortex in resting state data, and in a motor task the classic movement-related beta desynchronisation and post movement beta rebound were shown to be made up of changes in burst occurrence. A novel functional connectivity metric was then introduced based on the coincidence of bursts from distal brain regions, allowing the known beta band functional connectome to be reproduced. Bursts coincident across spatially separate brain regions were also shown to correspond to periods of heightened coherence, lending evidence to the communication by coherence (Fries 2005, 2015) hypothesis.

Post-stimulus responses across the cortex: During a motor task, both primary (during stimulation) and post stimulus responses (PSR) can be observed. These are well characterised in the literature, but little is known about their functional significance. The PSR in particular is modified in a range of seemingly unrelated neurological conditions with variable symptoms, such as schizophrenia (Robson et al. 2016), autism spectrum disorder (Gaetz et al. 2020) and multiple sclerosis (Barratt et al. 2017), indicating that the PSR is a fundamental neurophysiological process, the disturbance of which has implications on both healthy and pathological brain function. This work therefore tested the hypothesis that the PSR is present across the cortex. MEG data were acquired and analysed from two experiments with 15 healthy adult volunteers each – the first was a right-hand grip task with visual feedback, the second involved passive left visual field stimulation. Both experiments varied stimulus duration (2s, 5s and 10s) with a 30s rest-period between trials to allow characterisation of the full PSR. A univariate 3-state time-delay-embedded hidden Markov model (HMM) was used to characterise the spatial distributions of the primary and PSR across the cortex for both tasks. Results showed that for both tasks, the primary response state was more bilateral over the sensorimotor or visual areas (depending on task) where the PSR state was more unilateral and confined to the contralateral sensorimotor or visual areas (again, dependant on task). A state coincidence metric was then used to investigate the integration of the primary and PSR states across brain regions as a measure of task-related functional connectivity.

Hidden Markov modelling of the interictal brain: Epilepsy is a highly heterogeneous disease with variations in the temporal morphology and localisation of epileptiform activity across patients. Unsupervised machine learning techniques like the HMM allow us to take into account this variability and ensure that every model is tailored to each individual. In this work, a multivariate time-delay embedded HMM was used to identify brain states based on their spatial and spectral properties in sensor-level MEG data acquired as part of standard clinical care for patients at the Children’s Hospital of Philadelphia. State allocations were used together with a linearly constrained minimum variance (LCMV) beamformer to produce a 3D map of state variance, hence localising probable epileptogenic foci. Clinical MEG epilepsy data are routinely analysed by excess kurtosis mapping (EKM) and so the performance of the HMM was assessed against this for three patient groups, each with increasingly complex epilepsy manifestation (10 patients in total). The difference in localization of epileptogenic foci for the two methods was 7 ± 2mm (mean ± SD over all 10 patients); and 94 ± 13% of EKM temporal markers were matched by an HMM state visit. It is therefore clear that this method localizes epileptogenic areas in agreement with EKM and in patients with more than one focus the HMM provides additional information about the relationship between them.

Item Type: Thesis (University of Nottingham only) (PhD)
Supervisors: Brookes, Matthew J.
Mullinger, Karen J.
Keywords: Magnetoencephalography; MEG; Hidden Markov Model; HMM; Epilepsy; Neuroscience; Functional Connectivity
Subjects: Q Science > QC Physics > QC501 Electricity and magnetism
R Medicine > RC Internal medicine > RC 321 Neuroscience. Biological psychiatry. Neuropsychiatry
Faculties/Schools: UK Campuses > Faculty of Science > School of Physics and Astronomy
Item ID: 69462
Depositing User: Seedat, Zelekha
Date Deposited: 23 Aug 2023 12:34
Last Modified: 23 Aug 2023 12:34

Actions (Archive Staff Only)

Edit View Edit View