model free approaches to investigating the neural mechanisms of working memory

68
10075476 Tom Minor Narender Ramnani MODEL FREE APPROACHES TO INVESTIGATING

Upload: tom-minor

Post on 04-Mar-2016

220 views

Category:

Documents


0 download

DESCRIPTION

Undergraduate Thesis BSc Psychology Royal Holloway University of London

TRANSCRIPT

Page 1: Model free approaches to investigating the neural mechanisms of working memory

10075476Tom Minor

Narender Ramnani

MODEL FREE APPROACHES TO INVESTIGATING THE NEURAL MECHANISMS OF WORKING MEMORY

Page 2: Model free approaches to investigating the neural mechanisms of working memory

ABSTRACT

A re-analysis of data from a study into the cerebellar contributions to working

memory (Hayter et al., 2007), was conducted, using the same General Linear Model

design matrix and an extension using Independent Component Analysis allowed for

a comparison of model-driven and data-driven statistical approaches and their

contributions to understanding the neural dynamics recruited during a verbal

working memory (VWM) task (PASAT). The increase in cognitive load within the

ADD condition, and the resulting changes in BOLD response was operationalised as

the variable most likely to demonstrate the execution of integrated cognitive

control. Focussing on brain areas that were found to be significantly activated

during the ADD > REPEAT condition, an attempt to evaluate model-free approaches

to neuroimaging is made and the suggestion that the use of both data-led and

hypothesis-driven models, may best suit an area of science that is renowned for the

difficulty in reliably interpreting incredibly noisy and complex data-sets, as in

functional Magnetic resonance imaging (fMRI)

Page 3: Model free approaches to investigating the neural mechanisms of working memory

INTRODUCTION

D’Esposito (2007) defines Working Memory (WM) as the temporary retention of

information recently experienced or retrieved from long-term memory that no

longer exists in the external environment. A good test of WM is the Paced Auditory

Serial Addition Task (PASAT) because it enables researchers to manipulate

cognitive and sensory demands within a task that can be learned and executed

within neuroimaging studies and allows for functional and anatomical introspection

into the inner workings of WM at the level of the brain. Evidence related to the

neural substrates of WM will be considered before introducing 2 different statistical

approaches, as the case for either model-based or data-driven methodologies to

interpreting fMRI, data from a WM task is assessed. Neuroscientific data, from

primate studies will be discussed to generalise and further our appreciation of the

range of brain areas that appear to be related to WM in humans .

Smith et al., (1996), confirms that the posterior parietal cortex is implemented left-

hemispherically, as a neural storage component to VWM, including a sub vocal

rehearsal component in Broca’s area and premotor and supplementary motor

(frontal) areas. Smith, (1998) also concludes that there is scientific convergence on

the idea that there are separate systems for verbal and spatial memory, in support

Page 4: Model free approaches to investigating the neural mechanisms of working memory

for Baddeley’s decomposition of VWM into a phonological buffer and rehearsal

process within the left posterior parietal cortex (Brodmanns areas/BA40), Broca’s

area (BA44), left premotor area (BA6) and left supplementary motor areas (BA6),

which are found to be activated in the maintenance and manipulation of information

in VWM. Broca’s aphasics show a steeper forgetting curve than controls

(Korsakoff’s patients), suggesting that this poorer performance is likely to be caused

by a lack of subvocal rehearsal, also left posterior parietal cortex lesions seem to

interrupt verbal repetition, implying a deficit in the storage component of VWM and

offering support to the idea that the posterior parietal area is generally involved in

phonological processing and not restricted to memorisation (Smith et al., 1996).

Majerus et al., (2006) note that the intraparietal sulcus (IPS) is one of the most

consistently activated regions during verbal short-term memory (VSTM) tasks,

although the precise role it plays is a matter of controversy. IPS activity was found

in Hayter et al., (2007), and the current reanalysis, providing a motivation to model

and try to theoretically understand why this region (and many others circumventing

the IPS) show significant activation in this experiment and what that might tell us

about the functional connectivity of potential networks of structurally independent

but functionally integrated brain regions involved in VWM. The IPS is functionally

connected to serial/temporal order processing areas in the right IPS, premotor and

cerebellar cortices (Majerus et al., 2006), suggesting some functional inter-

hemispheric interaction and strengthening the case that bilateral IPS activity may be

related to the cognitive demands of the task. During STM tasks, the left IPS shows

Page 5: Model free approaches to investigating the neural mechanisms of working memory

connectivity to the phonological, and orthographic processing areas in the superior

temporal and fusiform gyri (Marjerus et al., 2006), suggesting that the IPS acts as an

attentional modulation area for distant neural networks specialised in language or

order representation and strengthening attention-based accounts of VSTM.

Chen and Desmond, (2005) have implicated inferior and premotor frontal cortex,

the insula and the cerebellum in VSTM tasks, demonstrating that areas outside of

the parietal lobe are also involved in processes related to the phonological store and

providing evidence for a fronto-parietal-cerebellar network for VSTM. These

researchers also found that the IPS is connected to the medial superior frontal areas,

the left cuneus, and the posterior cingulate extending to the precuneus. Again, these

results reflect the original findings from the Hayter et al., (2007) study and seem to

be best represented in the current results from the extension analysis with a model-

free approach as opposed to the model-driven approach. Due to a significant part of

the operated experiment being verbal instruction and potential subvocal rehearsal,

it seems poignant to include the IPS and its connectivity, focussing on its

contribution to an integrated network of neural dynamics, evidenced to underlie

VSTM.

The first statistical approach used was a General Linear Model (GLM), which

assumes Gaussian distribution of data. GLM makes this assumption and sometimes

this rigid hypothesis is wrong, leading to inaccurate data that misses something

with import for the experiment, because it was not modelled (expected) in the

design matrix and therefore considered as noise. Hemodynamic activity varies

Page 6: Model free approaches to investigating the neural mechanisms of working memory

across many factors, including intra-individually, GLM assumes that all data

represents a Gaussian distribution, and ignores any data that isn’t distributed in a

Gaussian manner, as noise. Not all hemodynamic activity related to the experiment

is Gaussian in nature; there may be unexpected, non-Gaussian distributed data that

still has something to tell us about the processes of WM and this is why the use of

data-driven approaches is required. Using a General Linear Model (GLM) statistical

approach to their data analysis, Hayter et al., (2007), defined the parameters of their

design matrix, according to the variables of interest (regressors) and considered all

activity related to the temporal dynamics of their experiment, thus interpreting

results based on effects due to voxel-wise interactions that could be accounted for

by their preconceived model, driving the analysis in a hypothesis led fashion. In

order to test the usefulness of GLM, a second statistical approach was introduced,

this was Independent Component Analysis (ICA).

ICA has its own problems in that, without a defining model to refer to, there are

interpretive limitations. Fortunately a new form of ICA, used in this study, allows

for the analysis of regressors, so that the data set represents the best of model–free

and the best of model-driven statistical methods in order to more closely scrutinize

results. ICA allows for the blind source separation of a mixed data-set into its

component parts, dealing with ‘functional segregation’ by detecting focal differences

in neuropsychological effects in terms of a number of regionally specific changes

whose significance can be assessed independently of changes going on elsewhere in

the brain (Friston, 1998). ICA is a data-led approach to the analysis of fMRI data,

allowing for exploratory research. fMRI data sets, represent a mixture of signals

Page 7: Model free approaches to investigating the neural mechanisms of working memory

resulting from hemodynamic changes, reflecting neural activity, motion and

machine artefacts and signals from physiological cardiac and respiratory pulsations:

the relative contribution of each of these components is undetermined, suggesting a

role for blind source separation (Jung et al., 2000). In dealing with the ‘neural

cocktail-party problem,’ ICA allows for the delineation of a noisy signal into its

important component parts, attempting ‘true redundancy reduction’ by

‘minimalising mutual information,’ though it is important to note that true

‘independence’ does not really exist because neurons can be coupled by synapses

and share efferent connections (Brown et al., 2001). ICA is particularly effective at

detecting task-related activations, including unexpected activity that would not be

detected under the premise of a model-based analysis, which considers unexpected

activity as error. It is possible to input a regression framework into the ICA,

allowing for a combination of hypothesis testing and data-driven research (Brown

et al., 2001).

McKeown et al., (1997) state that the principle of localisation implies that each

psychomotor function is performed in a small set of brain areas, which are

anatomically segregated for each function. They point out that the goal for

decomposition of fMRI data into physiological and cognitive components is to

determine topographically distinct brain areas that are co-activated during time-

series acquisition throughout the experiment. Finally, under McKeown’s conception

of ICA, it is equivalent to saying that voxel values in one component, convey nothing

of any value about the voxels in any other component, which is a stronger criterion

Page 8: Model free approaches to investigating the neural mechanisms of working memory

than the assumption of GLM (that maps of voxel-wise interactions from different

components are uncorrelated and Gaussian in nature).

In attempting to assess the contribution of cortico-cerebellar loops in the execution

of skilled cognitive operations, Hayter et al., (2007) reported activity in the medial

portions of cerebellar cortical lobule VII, linking these activations to the motor

demands of their experimental condition (ADD) within the PASAT. Noting the

substantial evidence of prefrontal-cerebellar connectivity (Hoover & Strick, 1999,

Kelly & Strick, 2003, Middleton & Strick, 2001; Walker, 1940), Hayter et al., (2007)

tested the hypothesis that highly practiced execution of action, engages cerebellar

circuits (Wolpert et al., 1998) by manipulating the sensory demands within a task

that required speech motor control. Activity was predicted in the auditory areas of

the superior temporal gyrus and ventral areas of the Primary Motor Cortex

(PMC/M1) -- containing orofacial musculature representations, along with

cerebellar lobules, known to be components of the motor loop (IV, V and VI). In

addition to confirming their hypothesis, Hayter et al., (2007), found co-activations in

mid portions of the middle frontal gyrus (9/46 putatively), anterior portions of the

cingulate cortex and areas within the pars triangularis (inferior and middle frontal

gyri/Broca’s area), supporting literature documenting the activation of these areas

during WM in humans (Curtis & D‘Esposito, 2003; Passingham & Sakai, 2004; Smith

& Jonides. 1996). The Hayter et al., (2007) study provides support for the

‘automaticity function’ of the cerebellum by demonstrating that the cerebellum is

involved in the execution of learned actions thus freeing up cortical structures to

deal with novel information processing. It is proposed that due to the uniform

Page 9: Model free approaches to investigating the neural mechanisms of working memory

cytoarchitecture of the cerebellum, considered alongside the consistent findings of

activity related to the control of motor speech production and prefrontal cognition,

that the mechanisms supporting automaticity of action and automatic information

processing may be similar (Ramnani 2006). Hayter et al., (2007) add to the

consistently robust finding of activations in the lateral convexity of the PFC in

relation to maintenance and manipulation-related demands of WM (Curits &

D’esposito, 2003; Passingham & Sakai, 2004), highlighting the role of the cerebellum

in processes linked to verbal WM and emphasising the use of these processes in

acquiring a skill through the process of automaticity.

Cerebellar contributions to WM form the starting point of this investigation. Using

the rationale and data from the Hayter et al., (2007) paper, this investigation is

based on discovering what impact model based, hypothesis-driven techniques as

compared to minimal assumptions, data-led approaches to analysing fMRI data,

have on the actual statistical processing and interpretation of results. As such, part

of this investigation is a re-analysis of the Hayter et al., (2007) data set, using the

same design matrix within a GLM analysis (though on a later SPM5 program) and a

comparison to an extended analysis, using ICA, including a GLM model-fit,

displaying the variance of results based on the 2 techniques. On the whole, results

from the re-run, corresponded to the detail in the Hayter et al., (2007) study, some

differences were found and these will be discussed in terms of the operating

systems and potential human error, within a broader discussion of the techniques

used and potential fronto-parietal-cerebellar network for VSTM. The further

Page 10: Model free approaches to investigating the neural mechanisms of working memory

analysis (ICA) was able to allow for potentially more detailed, and more selective

choice of peak-voxel activation and as such, brings the possibility of a more complex

view of the structure and function of the hemodynamic activity relative to the

original experiment.

The reliability of neuroscientific interpretations of physiological data sets is

intimately linked to the quality of the scientific tools and methods used in detecting

and localising brain activity, which relies on a less than uniform nomenclature and

differing sets of anatomical atlases. The search for detailed knowledge of how the

brain works, and how interconnected but distinct brain areas communicate with

one another; through which pathways and in what serial order, has been greatly

advanced using non-human primate research which allows for deeper introspection

into the mechanics and architectures of brain areas that are typically out of reach in

human subjects, due to their critical, sub-cortical or frontopolar location.

Petrides & Pandya., (2001) have done much to assist in establishing meaningful

comparisons between the human and macaque brain by developing a common

architectonic map, allowing for delineations based on functionally and structurally

evolved parts of the human PFC and also contributing to our understanding of brain

function. Comparing the cytoarchitectonic nature of the human and macaque

ventrolateral PFC (vlPFC) and corticocortical connection patterns in the monkey, via

injecting tracers into Walkers area 12, herewith referred to as area 47/12. The

multimodal input into this area can be assessed and fed into a model of how this

Page 11: Model free approaches to investigating the neural mechanisms of working memory

area differs in its function to the corresponding brain area in humans, and what it is

about its function in monkeys that can add to our ‘as-close-to’ complete view.

Tracers were found in the rostral inferotemporal visual association cortex and in

temporal limbic areas such as the parahippocampal cortex. Dorsally adjacent to

47/12, lies area 45 which represents the dorsal part of the vlPF convexity, just

below area 9/46 and caudodorsal to Brodmanns area 47. Injections to area 45,

labelled the Superior Temporal Gyrus (STG) (auditory association cortex) and the

multimodal cortex in the upper bank of the Superior Temporal Sulcus (STS). Area

45 occupies a large part of the pars traingularis. Tracers beginning their journey

here were also found in dorsolateral frontal areas 6, 8Ad, 8B and 10, as well as area

24 of the cingulate gyrus. Petrides & Pandya (2001), suggest that area 45 subserves

higher-order aspects of organisation of linguistic processing, but is not limited to

that; due to the existence of these areas in monkey, the implication of a more

general and fundamental role in cognition which may have adapted in humans to

serve linguistic processing in the left hemisphere, is posited. Petrides, (2005)

believes that ventrolateral prefrontal cognition controls information processing in

posterior cortical areas necessary for active retrieval of information from memory.

Area 47/12, with its strong links to rostral inferotemporal visual association cortex

and ventral limbic areas (rostral parahippocampal gyrus) may be a part of a frontal

mid-ventrolateral exectutive system that is involved in active judgements on stimuli

that are coded and held in posterior association cortex (Petrides, 2005). This body

of work, does much to illuminate some of the findings from the current re-analysis

of the Hayter et al., (2007) data, using an ICA, as all the areas mentioned, show

Page 12: Model free approaches to investigating the neural mechanisms of working memory

activation and show tentative support for these hypotheses (frontal-

midventrolateral system).

Fuster (2004), refers to Koechlin et als., (2007) hierarchical cascading model of

executive processes in the lateral cortex of the frontal lobe. Discussing the upper

processing stages of the perception-action cycle, he concludes that at all levels of the

CNS, sensory processing happens sequentially and along a posteroanterior axis,

with feedback at every level. Cortically, information is believed to flow circularly

through a series of hierarchically organised areas and connections constituting the

perception-action cycle. Within this conception, automaticity (cerebellar learning in

order to free up frontal areas for novel information processing) is implied. Koechlin

et al., (2007), view the lateral PFC (lPFC) as a cascade of executive processes,

organised from premotor to anterior PFC regions, that control behaviour according

to the stimulus, perceptual context and temporal dynamics of the stimulus,

providing a unified and modular account of cognitive control. This top-down

interaction between the lPFC regions and premotor or posterior association cortices

predicts that episodic, contextual and sensory contributions to cognition, gradually

emerge from rostral to caudal lPFC and premotor regions, and that the increasing

demands of these factors have additive effects of behavioural reaction times, this is

characterised by Weber’s Law, (Blake & Sekuler, 2006). On the basis of this

research, it is believed that rostral activations of the lPFC (inferior/middle frontal

gyrus; pars triangularis) will result from variations in WM load, (maintaining

instructions relative to cues over the course of experimental episodes) and that

these areas are engaged in selection of appropriate representation-for-action

Page 13: Model free approaches to investigating the neural mechanisms of working memory

modules, processed across episodes and exert control on the caudal lPFC, but not

directly on the premotor representations – it is the caudal regions that select

premotor responses to task sets according to the context, and rostral regions select

caudal lPFC representations, modelling the classical theory of executive control

based on central control of multiple slave systems (Koechlin et al., 2007)

The caudate nucleus seems to function similarly to the PFC, in that damage to both

areas result in similar deficits due to the sharing of inputs from frontal and

temporo-parietal cortical areas and outputs to the basal ganglia. Part of the fronto-

caudate pathway projects via the thalamus back to BA 46 in PFC, suggesting that the

cognitive functions of the dorsolateral prefrontal (dlPF) and anterior cingulate (AC)

pathways implement the caudate (Abdullaev et al., (1997). The caudate (tail &

head) is activated in Hayter et al., (2007) and in the current re-analysis, including

the GLM and ICA, these results fits with cascade model along the posteroanterior

axis for corticosubcortical WM operations.

Desikan et al., (2006) warn that there is limited ability to qualify critical dimensions

of interest in localisation of fMRI data, because of the considerable inter-individual

variability of topographic cortical features. Stating that the banks of the superior

temporal gyrus (STG) and pericalcarine cortex have larger interindividual

variability due to their pure sulcal nature. Bearing this mind, a comparison of

results from the original data set and the current reanalysis with the same GLM on a

newer SPM5 programme and a new version of ICA that allows for GLM regression

Page 14: Model free approaches to investigating the neural mechanisms of working memory

analysis and illustrates the variations in results between the two approaches,

allowing for some good comparisons and elucidating the neural mechanics of VWM

in the PASAT.

METHODSThis is a report from an already published data set, we report the salient issues.

PARTICIPANTS:

The original study involved 15 Right-handed volunteers between the ages of 18 and

29 (6 males) who gave written and informed consent confirming that they had no

neurological or psychiatric history. The study ran in line with the Royal Holloway

University of London (RHUL) MRI Rules of Operation and the Medical Devices

Agency. Ethical approval from the RHUL Psychology Department Ethics Committee

was received for a reanalysis of the original results and no experiment took place.

EXPERIMENTAL DESIGN:

The aim of this study was to investigate what differences were found in the results

of a reanalysis of the original data from the Hayter et al., (2007) paper, using the

same General Linear Model used in the original study and an Independent

Components Analysis.

Page 15: Model free approaches to investigating the neural mechanisms of working memory

Paced Auditory Serial Addition Test (PASAT):

Tombaugh (2005) recognises that the PASAT measures the performance of subjects

across multiple sensory modalities due to the requirement of successful completion

of a variety for functions related to attention, including audition, visual perception,

numerical cognition and speech. In the original study, the adaptation of the PASAT

allowed for the investigation of BOLD activity related to the specific cognitive

demands of the task, whilst enforcing strict experimental controls. The preparation

phase, outside of the scanner allowed for familiarisation with requirements, then

subjects practiced the test inside the scanner, to familiarise themselves with the test

environment. The experimental phase consisted of 35 blocks of experimental

(ADD), 35 block of control (REPEAT), and 10 null blocks (with no stimuli), lasting 32

minutes altogether. Blocks were pseudo-randomly intermixed. For more detail

about the PASAT test, refer to Hayter et al., (2007)

APPARATUS:

Participants lay supine within the MRI Scanner with padded restraints immobilising

their head. Instructions were received audibly through headphones that were

compatible with MRI apparatus. Verbal responses were measured using an MRI-

compatible microphone. An overhead mirror mounted on a head coil made the

visually presented instructions viewable as they were back projected onto a screen.

Page 16: Model free approaches to investigating the neural mechanisms of working memory

Experimental and control stimuli were controlled by a computer. For detail and

timings of experimental episodes refer to Hayter et al., (2007).

IMAGE ACQUISITION:

Participants were scanned with the 3-T Siemans Trio MRI scanner at RHUL. First

structural images were acquired. The functional sequence during the experimental

phase was made up of 644 EPI images and 4 volumes were collected before the

experiment began to minimise longitudinal relaxation artefacts (Hayter et al, 2007).

SPARSE SAMPLING:

Participants produced overt verbal responses during periods of deliberate scanner

silence (sparse-sampling) in order to allow for and cope with the head motion

artefact expected during the spoken response.

MRI Image Analysis:

Image processing and analysis for the original replication were carried out in SPM5

(www.fil.ion.ucl.uk/spm) on a dual core AMD Athlon 64 MHz PC with 2 GB of RAM

Linux server 2003 and Matlab 6.5 (MathWorks.Inc., MA, USA)

Page 17: Model free approaches to investigating the neural mechanisms of working memory

For the ICA, the GLM from the previous analysis was fed into the ICA via the FSL

programme on the same PC and server using Linux.

Preprocessing:

REALIGNMENT

Realignment was computed with reference to the first EPI time-series and the

resulting head motion parameters (3 rotations and 3 translations) were saved and

included as regressors within the GLM. This step attempts to model the acquired

head motion regressors, statistically, to allow for accuracy in later steps based on

the variance in results that may be caused by the subjects translating or rotating

their head during the verbal response phase of the experiment and thus influencing

the maps of activity.

NORMALISATION

Brains aren’t all the same, so in order to meaningfully interpret the results, subjects’

individual brains need to be adjusted statistically to reduce individual differences

and provide a common stereotaxic space from within which to compare localised

activations across the board. Normalising the volumes to the ICBM template

(Montreal Neurological Institute series; MNI) of standard stereotaxic space was

achieved using rigid, linear scaling that brings the individual brain into the realm of

stereotaxic space and then non-linear warping which selects particular bits of the

Page 18: Model free approaches to investigating the neural mechanisms of working memory

brain that seem to have evaded normalisation through linear transformations and

require particular attention in order to match the template meaningfully.

SMOOTHING

An 8mm smoothing kernel was applied at the final stage of preprocessing. Within

the GLM analysis, these steps had to be processed manually, with the aid of the

SPM5 programme and within FSL, for the ICA, these steps were computed

automatically. Smoothing data amplifies significant signals and reduces signals of no

interest (noise). The size of the kernel is essentially an arbitrary value, however

using a smaller kernel may have magnified many more subcortical activations and it

would be extremely unfruitful to use higher than 10mm kernel, because at that

range, segregating activations loses specificity, especially between anatomically

proximal areas. Smoothing kernels need to be at least twice the size of the voxels to

be able to tell us anything meaningful, otherwise they may miss minute details in

signal significance, reducing something interesting or elaborating something that is

in fact noise and of no interest.

Statistics:

GLM is conducted using a Linux system, Statistical Parametric Mapping (SPM)

programme within MATLAB. This programme allows you to create the design

matrix, defining the parameters of the variables of interest. In this case, head

Page 19: Model free approaches to investigating the neural mechanisms of working memory

motion regressors (3 x rotations/ 3 x translations), conditions (instruction, add,

repeat) and error were modelled within the design matrix. Preprocessing the data

for SPM5 involved realigning the individual subject scans on the basis of the head

motion artefact to measure the extent to which a particular scan deviates from the

first time-series. By modelling approximate estimations of head motion from fMRI

data, using sparse sampling as a way to measure minute differences in head position

before and after overt verbal responses are made during scanner silence, SPM5 re-

slices the brain scan, adjusting the data according to the variance in signal due to

motion, as a first step to creating statistical validity. Normalising individual brain

images (T1s) into a common stereotaxic space (MNI template) requires telling SPM

which brain needs normalising and selecting the correct template within which to fit

it. The linear affine scaling and non-linear warping of the images that fits the

individual subjects brain into the template brain requires selecting the mean image

from the newly realigned files as the source image, and then selecting the 644

realigned files to be normalised to an EPI image. The last step before setting up the

GLM design matrix, is smoothing, which requires inputting the voxel dimensions of

the isotropic smoothing kernel desired for use (8mm x 8mm x 8mm).

Event definition and modelling:

The four events modelled from the original experiment were: Instruction (1s); ADD

blocks (15s); REPEAT blocks (15s); and Error blocks (15s).

Page 20: Model free approaches to investigating the neural mechanisms of working memory

GLM explains how much each regressor explains the variance in the data, in order to

do this GLM needs to know the temporal dynamics of the experiment so that it can

identify different time points of each block and convulve them with the

hemodynamic response recorded.

First-level single-subject analysis (Parameter estimation of GLM):

Within SPM5, after all preprocessing has been completed, specifying the 1st level

statistics begins by selecting the subject’s smoothed files, specifying the interscan

interval as 3 seconds and manually inputting the details of the 4 conditions of the

experiment (Instruction, Add, Repeat and Error), which later allow for t-contrasts.

For each condition, the Onset files have to be imported from the original data and

the duration of the blocks have to be specified as 15 seconds, except for Instruction

which was 1 second. A High Pass filter was set to 60 and a canonical hemodynamic

response function was selected as the basis function (expected BOLD response)

before the model was saved. Model setup completed the estimation stage of the

GLM, at which point some checks were necessary. By selecting review in SPM5, the

orthogonality of the design matrix is displayed, this allows for a quick check that

there are the right number of regressors within the design matrix and that they have

been estimated and modelled correctly. Once satisfied that the model is correct, the

GLM analysis for 1st level were run (See Table 3 in results for 4 individual GLM 1st

level analyses). T-contrast comparisons were run on: (1) INSTRUCTION only; (2)

ADD only; (3) REPEAT only; and (4) ADD > REPEAT.

Page 21: Model free approaches to investigating the neural mechanisms of working memory

T-contrasts and statistical parametric mapping (SPM{t}) maps highlighted voxels

within MNI space where Blood Oxygenation Level Dependent (BOLD) responses

were significantly different between the conditions ADD and REPEAT. The SPM.mat

file that is produced from the 1st level GLM is specified within SPM5, this requires

manual input of the values of different conditions, allowing SPM to bring up the

brain activations that relate to the specific demands of the contrast. For example,

Add greater than Repeat would be defined as Add carrying a weight of 1 and Repeat

carrying a weight of zero, for the statistics to produce maps that reflect brain

activity that was specific to the Add condition and greater than the activity found in

the Repeat condition. At this point, whole brain threshold tables are produced,

highlighting peak-voxel clusters, and SPM5 allows overlays onto canonical

templates which are a useful tool for localisation, as they share similar landmarks to

the anatomy in the atlas used for localisation (Duvernoy & Bourgouin, 1999), and

match the template in the MRICRO programme which allows for manual input of

MNI co-ordinates in order to specify where activity is in MNI space. SPM5 also

allows for overlaying back to individual T1 images, providing closer, adherence to

individual variations in brain size and shape.

Second-Level between-subjects analysis:

This involves a one-sample t-test applied to the contrasts from the first level

statistics (all results thresholded p<0.05, FWE corrected). Specifying 2nd level

Page 22: Model free approaches to investigating the neural mechanisms of working memory

statistics for GLM on SPM5 simply plugs the output from the 1st level statistics into

the group analysis. By selecting the contrast image files, saved from the 1st level,

SPM5 computes group level analysis on the contrasts you ask it to, finding

overwhelming common activations between subjects. Classical as opposed to

Bayesian inference methods were chosen and results were corrected for Family

Wise Error, which is a more commonly used, stringent correction for multiple

comparisons, applied to each and every voxel, meaning that GLM computes 100,000

t-tests at this level. The maps produced at 2nd level can be treated as in the 1st level,

with interactive thresholds and overlays within the SPM5 programme. Since group

level statistics, on the whole, tell us more about strong findings, a table of the results

of the contrast of Add > Repeat is presented in Table 1 of the results. Areas of

significant activation were localised using a overlay function within the SPM5 setup;

overlaying the T1 functional results (SPM{t} maps) onto a canonical anatomical

representative T1 image (MNI Series). Due to the variation of landmarks within

MNI space, a cross check involving overlaying individual SPM{t} maps back onto the

normalised T1 scans for each subjects’ results from the original study was done.

Independent Component Analysis:

ICA does not require predetermined regressors in order to separate the mixed

signal into its significant component parts. It is possible however to enter the GLM

model into the ICA on the FSL programme, allowing for an inspection of how well

ICA separates components of head motion, machine and physiological artefact from

Page 23: Model free approaches to investigating the neural mechanisms of working memory

interesting neuronal activity and the relevant, task-related hemodynamic

components. Within FSL, it is possible to import head motion vectors into the

analysis and build them into the weak model. 4 event variables were built into the

model (Instruction, Add, Repeat and Error), and prestats were computed using

MCFLIRT. The smoothing kernel was set to 8, as in the GLM. Registration to MNI

space, setting the high pass filter to 60 and the number of components set to 50

completes the model setup and then ICA is run on individual subjects data in

MELODIC within FSL. Once the ICA has run, within FEAT-FMRI in FSL, it is possible

to load the 4-dimensional EPI files for each subject and to inspect each component

that ICA found that could be separated from the source signal. Checks were made as

to the periodicity of activation within the component and whether it related to the

periodicity of the experiment, using temporal timecourses (see results). Tables with

p-values highlighted the most significant contrast within each component, allowed

for relation to specific contrasts such as Add > Repeat. Gamma distribution graphs

(see results) also helped to demonstrate that ICA was indeed delineating Gaussian

distributed data (components) from a mixed non-Gaussian source. ICA produces

many images for each component, mapping brain activity in different slices of the

brain, from the cortex down to the cerebellum. It also produces maps of relative

deactivation and colours these differently to activations, providing interactive

thresholds at every level. It is not possible, however to overlay the images produced

from the results of ICA within FSL, onto the individual TI brains, or into a canonical

template MNI space, this makes localising activations, more vulnerable to human

Page 24: Model free approaches to investigating the neural mechanisms of working memory

error. Duvernoy & Bourgouin (1999) was the anatomical atlas used in both the GLM

and ICA investigation.

Localising activation for both analyses, involved choosing sagittal and axial planes of

view of the brain within the atlas and finding landmarks that reflected the

dorsoventral and medial-lateral position of the images within the program. It was

important to use at least 2 planes of view, in order to make sure, something than

looked cortical, was not in fact subcortical and vice-a-versa. Unfortunately, the atlas

carries a tremendous amount more detail than either SPM5 of FSL seem capable of

producing at this stage, and even with the assistance of the MRICRO programme

within SPM5, considerable attacks on confidence levels for the accuracy of

localisation could be made. As far as this study goes, it is believed strongly that the

author’s localising skill is one of the more reliable aptitudes brought to this project,

and much painstaking time was spent, to ensure that the brain area selected as

representative of the result, was indeed, the closest possible result that could be

reported. Clearly, there were times when the activity could not be definitely located

to a marked area within the atlas, but in these cases, sulcul and gyral anatomy was

used as a gross morphological attempt to localise as accurately as possible.

Page 25: Model free approaches to investigating the neural mechanisms of working memory

RESULTSTable 1: (2nd level analysis of 1st level GLMs )ADD>REPEAT (FWE corrected for multiple comparisons, p<0.05, random effects analysis).

LOBE Gyri/Sulci MNI Co-ords Stats StatsT Z

Frontal (PF/PMC) Superior Frontal Gyrus (L) 0 12 48 13.25 5.81PFC Middle Frontal Gyrus (L) -42 0 52 13.08 5.78

Cingulate Gyrus (L) 8 26 32 11.59 5.57Cingulate Sulcus [R] -6 -68 52 12.34 5.66Central Gyrus/Sulcus [R] -46 -36 44 11.38 5.49

PMC Precentral Gyrus [R] -40 -48 55 11.38 5.49

Superior Precentral sulcus[R] -40 -48 55 11.38 5.49

Temporal Circular Insular sulcus [R] -22 24 8 11.47 5.51Circular Insular Sulcus (L) 38 20 -4 10.56 5.33Anterior/Intermediate transverse Temporal Gyrus (L) 38 20 -4 10.56 5.33

SUBCORTICALBasal Ganglia Caudate Nucleus [R] -18 -8 24 12.46 5.51

Subthalamic nucleus 14 -4 -10 10.27 5.28/Substantia Nigra (L)Cerebellum [R] -6 -50 -38 12.11 5.62

Results were corrected for Family Wise Error, a standard conservative control, which finds activations that are above 0.05 (p-value). If activations survive FWE correction, we know that there is absolutely something happening in the voxels highlighted. FDR correction finds more reasonable activations, but it was felt these gross voxel clusters would inhibit fine-tuned localisation with certainty.

Below is an example of the output in SPM5 for the group level analyses of 1st level GLM. Whole brain interactive thresholds, to select peak activity within voxel clusters

Page 26: Model free approaches to investigating the neural mechanisms of working memory

Table 2 (ICA results of individual 1st level analysis)LOBE Gyral/Sulcal anatomy STATISTICS Lateralisation

z pFrontal Superior frontal gyrus z = 5.64 p < 0.0000 Bilateral

Middle frontal gyrus z = 5.64 p < 0.0000 BilateralCingulate gyrus z = 0.68 p < 0.03501 Bilateral

Broca's area Inferior frontal gyrus z = 5.64 p < 0.0000 BilateralSuperior frontal sulcus z = 5.64 p < 0.0000 BilateralMedial orbital gyrus z – 1.88 P < 0.02979 Bilateral

Central sulcus z = 4.48 p < 0.0000 Bilateral

Inferior precentral sulcus z = 5.02 p < 0.00000 BilateralSuperior precentral sulcus z = 4.48 p < 0.0000 BilateralSuperior postcentral sulcus z = 6.18 p < 0.0000 BilateralParacentral sulcus z = 6.18 p < 0.0000 BilateralSuperior lingual gyrus z = 0.68 p < 0.03501 Bilateral

Parietal Supramarginal gyrus z = 5.02 p < 0.00000 BilateralSubparietal sulcus z = 0.68 p < 0.03501 BilateralInterparietal sulcusv z = 0.68 p < 0.03501 BilateralPost central gyrus z = 4.48 p < 0.0000 Bilateral

Superior parietal gyrus z = 5.64 p < 0.00000 BilateralPrecuneus z = 5.02 p < 0.00000 BilateralAngular gyrus z = 5.02 p < 0.00000

Occipital Lingual sulcus z = 0.68 p < 0.03501 BilateralInferior lingual gyrus z = 4.66 p < 0.0000 BilateralCalcarine sulcus z = 4.66 p < 0.0000 BilateralPosterior transverse Collateral BilateralSulcus z = 6.18 p < 0.0000

Temporal Inferior temporal gryus z = 6.18 p < 0.0000 BilateralSuperior temporal gyrus z = 6.18 p < 0.0000 BilateralMiddle temporal gyrus z = 6.18 p < 0.0000 BilateralSuperior temporal sulcus z = 6.18 p < 0.0000 BilateralFusiform gyrus z = 3.06 p < 0.00111 BilateralSubcollosal Gyrus z = 2.53 p < 0.00570 BilateralParahippocampal gyrus z = 5.59 p < 0.0000 Bilateral

Page 27: Model free approaches to investigating the neural mechanisms of working memory

Subcortical Hippocampus z = 6.18 p < 0.0000 BilateralBasal ganlia Head of Caudate nucleus z = 2.53 p < 0.00570 Bilateral

Basal nucleus of amygdala z = 5.59 p < 0.0000 BilateralCrus Cerebri z = 6.18 p < 0.0000 Bilateral

Some examples of the bilateral activation.Table 3. Common activations, individual comparisons of 1st level GLM and individual ICA analyses.Subject

Common Activation (ICA & GLM

GLM only ICA only

AM lobule VII superior lingual gyruscrus ii inferior lingual gyrusprecuneus inferior temporal gyrustail of caudate superior temporal gyrusSuperior parietal gyrus middle temporal gyruscentral sulcus superior temporal sulcusSuperior frontal sulcus Paracentral sulcus

Superior postcentral sulcus

FK Inferior frontal sulcus Cingulate sulcus Inferior lingual gyrusSuperior frontal gyrus Crus I head of the caudateSuperior temporal gyrus Inferior frontal sulcus inferior frontal gyrus

Superior frontal gyrus superior frontal sulcusCingulate gyrus Middle frontal gyrus

central sulcusprecentral gyruspostcentral gyrusPrecuneus

RO Superior frontal sulcus Lobule VII Head of the hippocampusSuperior frontal gyrus Crus II PrecuneusInferior frontal sulcus Superior precentral Superior temporal sulcusInferior frontal gyrus sulcus Inferior precentral sulcusMiddle frontal gyrus Supramarginal sulcus Supramarginal gyrusPost central gyrus Putamen Central sulcusLateral Fissure Superior precentral sulcus Precentral gyrusAngular gyrus Superior Parietal gyrus

SD Caudate nucleus tail of caudate head of caudateLateral fissure Superior frontal sulcus Inferior frontal gyrus

Superior frontal gyrus Inferior temporal gyrusPutamen Superior postcentral sulcusSuperior temporal sulcus Lingual sulcus

cingulate gyrusprecuneusInferior precental sulcusCentral sulcusSubparital sulcus

Page 28: Model free approaches to investigating the neural mechanisms of working memory

Superior parietal gyrusIntraparietal sulcusAngular gyrusParacentral lobuleCingulate sulcus

DISCUSSION

Seeing as the implicit goal of this thesis was to establish the methodological and

interpretational effects of using model-free approaches to biomedical signalling

experiments, this discussion will begin with focus on the statistics including the

technical issues related to the operating systems used for each and then go on to

discuss the specific anatomical discrepancies that show up in a comparison between

results from a model-based GLM and model-free ICA, relating to the research

introduced at the start of this paper.

The Hayter et al., (2007) hypothesis focussed on testing the interconnectedness of

the prefrontal cortex with the cerebellum in line with past research referred to in

the introduction and they used a GLM statistical approach. This study attempting to

discover what more could be said about the neural dynamics of the same data from

a PASAT VWM task, using a different, model-free statistical approach (ICA).

To create a GLM, manual input of the regressors of interest into a design matrix is

required, within SPM5 and MATLAB, allowing for a close inspection of the BOLD

responses relevant to the experimental episodes of ADD vs REPEAT condition.

Maps of peak-voxel cluster activity were produced and there were some definite

Page 29: Model free approaches to investigating the neural mechanisms of working memory

similarities as well as considerable lack of matching activity between the current

results of the earlier results from Hayter et al., (2007), especially in the Occipital and

Parietal lobes.

These differences, became, considerably more detailed in the ICA maps, which

produced nearly 600 less components for the analysis of temporally relevant

activations. Due to the sheer density of human programming, it can be estimated

that some of the lacking activations within the GLM re-run, were due to human

error, and disparate inter-observer reliability but more importantly, the fact that the

Hayter et al., (2007) study was solely driven by a hypothesis that expected to see

considerable activity in the cerebellum may have lead to their finding more supra

threshold voxels in that area. It is not suggested that the previous researchers,

made up these findings, only that the considerable variation could be due to the

inadequacies of the programs used to produce such maps and the substantial room

for selectivity in honing in on particular areas within whole brain activity that

support predetermined hypotheses in the minds of the researcher and within a

model-based approach in general.

There are several comparable differences and difficulties regarding the localisation

of activations from data from multiple subjects into a standardised space that allows

for direct comparisons, aside from the issues mentioned above, both FSL and SPM5

have their strengths and drawbacks with regards to this salient issue.

Page 30: Model free approaches to investigating the neural mechanisms of working memory

Whereas the GLM analysis within SPM5, lists tables with peak-voxel activations,

allowing you to explore the variously spread clusters of activation, to find the most

significantly activated voxel amongst a cluster of activations. FSL does not provide

interactive thresholds; it is possible to go into the components and establish which

activation was ‘peak’, but as for a simple method for doing that, there seems not to

be none, other than the colour intensity bars produced aside the component images,

reflecting the significance of the signal, using the colour spectrum as representative,

but it is extremely difficult to see, clearly, which areas are most significantly

activated and to get a sense of the relative activity across the brain; essentially its

like trying to look a jigsaw with lots of shades to blue and red and trying to see

which ones were more important, it is possible, but not very reliable. From a

general view, localisation from SPM5, with the ease of overlaying onto template

brains and the use of coordinates that are transferable into the MRICRO program,

which allows for even more scrutiny regarding localisation, seems to provide more

confidence in the process of localising, and somewhat helps to narrow down the

areas of interest, but then again, the GLM approach has the benefit of the model. FSL

appears to show you much more and makes the job of discerning important and

relevant localisations, not only more time consuming, but less reliable, as there is no

direct method of overlaying the activations from independent components onto a

stereotaxically homologous space. This vital tool, of being able to directly transfer

to image acquisitions onto a definite brain, with landmark features to assist in the

localisation, which by any stretch of the imagination is not without the risk of

human error, is a crucial failing of the FSL program approach to univariate analyses.

Page 31: Model free approaches to investigating the neural mechanisms of working memory

GLM attempts to extract a specific signal, that best fits the specified temporal model,

identifying regions of the brain that activate with similar time courses to the

experiment, depending on the Gaussian nature of the data. ICA extracts non-

gaussian data from a Gaussian, mixed source rather than the data that best fit the

model. ICA assumes that signal components are not only uncorrelated but also

‘statistically independent’, and attempts to find them from the mixed data-set of

fMRI timecourses and assumes that these signals derive from physically different

processes, thus maximising a measure of the joint entropy of the extracted signals

(Friston, 1998). Both ICA and GLM can be used; if there is sound reason for specific

hypothesis-testing, model-based approaches would be preferable, but if a model-

based approach wouldn’t extract all the signals of interest (as has been

demonstrated in this study), exploratory, data-driven methods can be more

appropriate.

FSL exceeds SPM5 in terms of options for viewing results, with one-click changes

from tables to graphs, displaying Gamma distributions of time courses that

represent independent components, allowing you to assess their affinity to the

experimental timings, and thus making sounder inferences regarding the relevance

of a particular component (if the time course matches that of the experiment, it is

likely to have something to do with experiment). You can plot graphs with SPM5,

but it lags in the immediacy of the FSL program, and somewhat in its usefulness.

SPM5 allows for whole brain p-values, which create tables of MNI co-ordinates

Page 32: Model free approaches to investigating the neural mechanisms of working memory

allowing for more certainty in localisation (with the use of MRICRO), whereas FSL

cannot do this. As a first time user, these comments may merely show a lack of

complete knowledge about the systems being discussed, but overall, there were

mixed pros and cons with both SPM5 and FSL. During different stages of the data

management and processing, each technique offered sometimes easier and

sometimes more difficult sets of command chains to go through in order to do the

necessary computations. Although the reanalysis did align generally to the results

from the previous study, there were some discrepancies, and it would seem almost

impossible for there not to be differences, because each time you reanalysed the

same set of data, there are clearly far to many areas for operational error.

In summary, GLM allows for partitioning of variance, in understanding the degree to

which each regressor explains the data. SPM5 computes a GLM on each voxel

serially, demanding Gaussian distribution of data. ICA takes an essentially Gaussian

mixture and tries to delineate the non-Gaussian distributions within it. It is

suggested that the use of both methods concurrently, is so far the best option for

trying to gather a holistic picture of brain activity during neuroimaging tasks, and

especially in the current climate of comparing each method to establish its relative

contribution; at this stage, neither stands out as a clear victor, so fortunately for

neuroscientists, applying multiple statistical techniques of brain imaging data,

seems to be the most sensible way forward

Page 33: Model free approaches to investigating the neural mechanisms of working memory

Understanding the neural mechanisms underlying active maintenance of task

relevant information hinges on how we resolve the nature of stored

representations, in addition to the types of operations performed on such

representations (D’Esposito, 2007). Representations equate to symbols and codes

for information that are activated transiently or continuously within neural

networks, and operations or processes (computations) are performed on these

representations. Areas of the multimodal cortex (PFC & parietal cortex) are in a

position to integrate representations through connectivity to the unimodal

association cortex and are also critically involved in the active maintenance of task

relevant information (D’esposito, 2007). Fuster, (1997), expounds that the PFC is

critically responsible for temporal integration and mediation of events that are

separated in time such as in the ADD condition of this PASAT test. To explore this

problem further specific results from both GLM and ICA will be discussed in relation

to the original Hayter et al., (2007) findings.

Frontal Lobe

At 2nd level (between groups), GLM analysis found similarities with the original

findings in the superior/middle frontal gyrus, cingulate gyrus, central sulcus,

precentral gyrus and superior precentral sulcus.

ICA finds inferior frontal gyrus, superior frontal sulcus, medial orbital gyrus, inferior

precentral sulcus, central gyrus, inferior precentral sulcus, superior postcentral

Page 34: Model free approaches to investigating the neural mechanisms of working memory

sulcus, paracentral sulcus, superior lingual gyrus, which did not show up from the

GLM.

According to Abdullaev (1997), bilateral premotor region (BA 6) – middle frontal

gyrus, activates due to stimulus effects, whereas BA 44/45 – inferior frontal gyrus,

shows effects of context but not stimulus and episode effects occur in both. The

frontal lobe activity found here, seems to support rostro-caudal accounts of lateral

PFC and premotor regions, dealing with cumulative cognitive demands, in a

cascading manner and fits within the classical theory of executive control (Koechlin

et al., 1997)

Ist level (individual differences) within GLM and ICA analyses highlight some other

discrepancies between the results. Some examples from individuals upon whose

results, both ICA and GLM analyses were computed.

Subject AM:

No common activations between GLM and ICA were found however GLM found

almost equal number of activations that ICA did not, including the central sulcus and

superior frontal sulcus. ICA found activity in the paracentral sulcus and superior

postcentral sulcus. These differences could be down to minute discrepancies in

localisation procedures, or they could reflect operational disparities between the 2

methods under discussion. Over all, these differing frontal activations, on their own,

Page 35: Model free approaches to investigating the neural mechanisms of working memory

do not negate other findings, indeed they may add to them, if we take the view that

both methods show us things that the other is not capable of, and by combining

data, our understanding is further enriched. Paracentral, superior postcentral and

central sulci move dorsoventrally from medial cortical areas towards and into the

superior frontal gyrus, it could be that these areas, play a significant role in the

particular demands of the VWM task but it would be difficult establish this from

group statistics alone, only by combining individual activations within 1st level GLM

results and the robust components from the ICA, can we begin to see, how both

methods, compliment, as oppose to exclude the other.

Parietal

2nd level GLM statistics, failed to retrieve any activations in the parietal lobe, despite

their being found with ICA and previously in Hayter et al., (2007). Considering, that

the GLM was input into the ICA, and the ICA managed to find similar activation in

the IPS and various parts of the parietal lobe. It is an anomaly to this researcher as

to why these results differed so extensively. On the basis that Hayter et al., (2007),

found similar activations and several significant independent components were

found in the parietal lobe during the ICA, it is proposed that these lacking data from

GLM reflect some mistake that may have been made in the preprocessing stages, for

there is no other obvious route to the discrepancy, and as such it is used an example

of how, the resource intensive manual data entry, involved in the setting up of the

Page 36: Model free approaches to investigating the neural mechanisms of working memory

model within the SPM5 programme, can allow for fairly huge differences in results,

due to human error, that is extremely difficult to trace and remedy.

Based on the research, these parietal areas probably represent how the Parietal lobe

is critically involved in VWM. Majerus et al., (2006), found that the IPS is

functionally connected to serial/temporal processing areas in the premotor and

cerebellar cortices, as well as to phonological and orthographic processing areas in

the superior temporal and fusiform gyri. This connectivity, suggest that parietal

areas (the IPS especially) act as attentional modulators of distant neural networks

that are known to deal with language and order information. The IPS, does seem to

be influenced by cognitive load and demonstrates a fronto-parietal-cerebellar

network for verbal short-term memory tasks, that is supported by our data.

Majerus et al., (2006), also note the recent discovery of increased connectivity

between IPS and medial superior frontal gyrus, cuneus and bilateral posterior

cingulate cortex involved in visual mental imagery, so these activations seem to

show that the Parietal areas of the brain, take part in managing and communicating

information from higher order processing areas in the frontal areas, to multimodal

cortex and towards the cerebellum, where automaticity of learning relieves the

demands on WM structures in the prefrontal areas to deal with novel information.

Phonological processing is known to activate the left interior parietal lobe, posterior

inferior frontal gyrus (Broca’s area), premotor cortex and the cerebellum

(D’Esposito, 2007). It is interesting to note that all parietal activations discovered

Page 37: Model free approaches to investigating the neural mechanisms of working memory

through ICA, were bilateral activations, which seem to suggest that their joint

activity, is not only task-related, but almost certainly not noise. The chances that

ICA would find perfectly bilateral activations in brain areas known to be

implemented during VWM tasks, consistently across subjects, is far more indicative

of the benefits of using ICA than the possibility that these may be uncorrelated.

Temporal

At 2nd Level, GLM finds the insular lobe (as did Hayter et al., 2007), along with

anterior and intermediate temporal gyri. ICA, on the other hand, does not find these

areas, but implicates inferior-, superior- and middle-temporal gyri, along with the

superior temporal sulcus, fusiform gyrus, subcollosal gyrus and parahippocampal

gyrus. 1st level GLM statistics for subject SD showed superior temporal gyral

activity that was not common to the ICA data, whereas subject FK, showed a

common activation there for both GLM and ICA. AM, RO and SD, all display

significant temporal lobe activations from the ICA, providing as yet, the most

commonly activated area in this comparison. Temporal activity here, reflects the

phonological component of the verbal task at hand. Majerus et al., (1996), state that

the posterior superior temporal gyrus is involved in the processing of novel

phonological information.

Subcortical

Page 38: Model free approaches to investigating the neural mechanisms of working memory

Caudate nucleus, subthalamic nucleus (putatively substantia nigra), and cerebellum,

all displayed clear activity from the 2nd level GLM statistics, whereas ICA delineated

the hippocampus, basal nucleus of the amygdala and crus cerebri, as being

significantly activated during the experiment. At 1st level GLM analysis, lobule VII

and crus II in the cerebellum were found in subjects AM and RO, caudate activations

were discovered in the results for subject AM and SD, and subject FK demonstrated

activity in crus I of the cerebellum. Overwhelmingly, ICA could not pick up any of

these findings, except for caudate activity, adding something by finding hippocampal

activity in subject RO. Clearly, there is subcortical involvement in the demands of

WM, but it seems as though GLM does a better job of retrieving these activations.

It is well established that the caudate nucleus, which is part of the striatum, is an

integral part of the fronto-striatal circuit subserving cognitive functions (Abdullaev

et al., 1997) with afferent connections from frontal and temporal-parietal cortical

areas and efferents to the basal ganglia, it would seem that these networks

contribute to the dorsolateral prefrontal and anterior cingulate pathways of

cognition. One suggestion is that the head of the caudate is implicated in the

prefrontal circuitry responsible for semantic and phonological computations in

cognitive tasks (Abdullaev et al., 1997) providing a reasonable explanation for why

caudate activity was found in this analysis, based on the phonological and higher

order cognitive demands of the original task.

CONCLUSION:

Page 39: Model free approaches to investigating the neural mechanisms of working memory

Over all ICA finds more, but there are instances where ICA misses something that

GLM finds, so using both methods is the surest way to get the fullest picture of what

is going on in the brain during this task. One caveat to the ICA is that group analysis

within stereotaxic space as opposed to individual brains is not possible at this stage.

This means that we can only make inferences about individual subjects and not

between subjects, at least not statistically. GLM adheres to a rigid hypothesis, which

is argued, may consider some significant activity as noise, which is why a

combination of hypothesis driven and data-driven methods is forwarded.

There were many complicated steps in the analysis of this data, and SPM5 requires

considerably more manual input, allowing for more potential human error. More

checks would increase reliability in the statistics, and although, necessary reviews

were made, more would be recommended to improve statistical confidence on the

part of the researcher. Had there been more time, more space on the server and

more brain power in the case of this novice attempt to understanding results from

neuroimaging studies, stronger arguments could be made. In future, working with

fresh data would be recommended and variations in the size of the smoothing

kernel applied might highlight some interesting results. It is believed that the

preprocessing and statistical analyses computed for this study, were the same as the

original study, though it seems more likely that discrepancies in results between the

2 identical GLM analyses are either due to the use of a different SPM5 (instead of

SPM2) programme, or that this first try, inevitably lead to some mistakes being

Page 40: Model free approaches to investigating the neural mechanisms of working memory

made. The group was also limited in the storage space available on the server’s

hard-drive to compute additional analyses, not that there would have really been

time or scope for this undergraduate thesis to deal with so much more detail.

A reanalysis of the Hayter et al., (2007) GLM analysis, revealed some consistent and

some discrepant findings of neuronal activity in areas that are now well established

as being recruited during the multimodal and multi-sensory activity of performing a

PASAT. Some suggestions for these discrepancies and evidence to support the

consistencies have been made, along with more detail found from running an ICA on

the same data. Overall, both statistical approaches, together, help to support the

cascading model of hierarchical executive function (Koechlin et al., 2007). The

relatively consistent prefrontal, parietal, temporal and subcortical activity found in

this study, that to a greater extent contributes, as opposed to disputes the earlier

findings in Hayter et al., (2007), seems to suggest a postero-anterior axis for

corticosubcortical WM operations. This study supports the notion of a fronto-

parietal/temporal-cerebellar network for VSTM, which is evidenced by the work of

Petrides (2005), who found that frontal area 47/12 had strong links to rostral

inferotemporal visual association cortex and ventral limbic areas including rostral

parts of the parrahippocampal gyrus. It is argued that active judegments on stimuli,

typical to the requirements of the PASAT, are coded along a frontal mid-

ventrolateral executive system that implements posterior association cortex during

VWM tasks.

Page 41: Model free approaches to investigating the neural mechanisms of working memory

GLM results or ICA results alone, would not provide enough detail for one to make

this conclusion, so the use of both techniques, in this case seems to have provided

the clearest picture of what happens in the brain during this PASAT. Some of the

strongest findings however, came from the ICA which brought attention to the

Parietal lobe in particular, which is why this paper took the position it did, regarding

fronto-caudal pathways which implicate a lot of the areas found in this investigation

that may not have been so well understood without the knowledge about the IPS

which certainly represents the most robust finding from this study and the unique

contribution that ICA has made to this line of research.

Page 42: Model free approaches to investigating the neural mechanisms of working memory

REFERENCES

Abdullaev, Y, G. & Melnichuk, K, V. (1997). Cognitive operations in the human caudate nucleus. Neuroscience Letters, 234, 151-155

Baddeley, A. (1986). Working Memory, Oxford: University press

Blake, R. & Sekuler, R. (2006). Perception, London: McGraw Hill

Chen, A. S. H. & Desmond, J. E. (2005). Temporal dynamics of cerebro-cerebellar network recruitment during a cognitive task. Neuropsyhcologia, 43, 1227-1237

Curtis, C. E. & D’Esposito, M. (2003) Persistent activity in the prefrontal cortex during working memory. TRENDS in Cognitive Science, 7, 415-424

Desikan, R. S., Segonne, F., Fischl, B., Quinn, B. T., Dickerson, B. C., Blacker, D., Buckner, R. L., dale, A. M., Maguire, R. P., Hyman, B. T, Albert, M. S. & Killiany, R. J. (2006). An automated labelling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage, 31, 968-980.

D’Esposito, M. (2007). From cognitive to neural models of workinh memory. Philosophical Transactions, 362, 761-772

Duvernoy, H. M. & Bourgouin, P. (1999). The Human Brain; Surface, Three-Dimensional Sectional Anatomy with MRI, and Blood Supply. Springer

Friston, K. J. (1998). Modes or models: a critique on independent component analysis for fMRI. TRENDS in Cognitive Sciences, 2, 1227-1230

Fuster, J. M. (2004). Upper processing stages of the perception-action cycle. TRENDS in Cognitive Sciences, 8. 1-9

Hayter, A. L., Langdon, D. W. & Ramnani, N. (2007). Cerebellar contributions to working memory. Neuroimage, 36, 943-954

Page 43: Model free approaches to investigating the neural mechanisms of working memory

Hoover, J. E. & Strick, P. L. (1999). The organisation of Cerebellar and Basal Ganglia Outputs to the Primary Motor Cortex as revealed by retrograde Transneuronal Transport of Herpes Simplex Virus Type 1. Journal of Neuroscience,15, 1446-1463

Horwitz, B., Tagamets, M. A. & McIntosh, A. R. (1999). Neural modelling of functional brain imaging, and cognition. Trends in Cognitive Sciences, 3, 91-98

Kelly, R. M. & Strick, P. L. (2003). Cerebellar Loops with Motor Cortex and Prefrontal cortex of a nonhuman primate. The Journal of Neuroscience, 23, 8432-8444

Koechlin, E. & Summerfield, C. (2007). An information theoretical approach to prefrontal executive function. TRENDS in Cognitive Science, 11, 229-236

Koechlin, E., Ody, C. & Kouneiher, F. (2007). The Architecture of Cognitive Control in the Human Prefrontal Cortex. Science, 302, 1181-1187

Jung, T., Makeig, S., Lee, T., McKeown, M. J., Brown, G., Bell, A. J. & Sejnowski, T. J. (2000). Independent component analysis of biomedical signals. Proceedings of the 2nd International workshop on ICA and Blind source separations, 633-644.

Majerus, S., Poncelet, M., Linder, Van der., Albouy, G., Salmon, E., Sterpenich, V., Vandewalle, G., Collete, F. & Maquet, P. (2006). The left interparietal sulcus and verbal short-term memory: Focus of attention or serial order? Neuroimage, 32, 880-891

McKeown, M. J., Jung, T, Makeig, S., Brown, G, Kindermann, S. S., Lee, T. & Sejnowski, T. J (1997). Spatially independent activity patterns in functional MRI data during the Stroop color naming task. National Academy of Sciences, 95, 803-810

Middleton, F. A. & Strick, P. L. (2001). Cerebellar projections to the prefrontal cortex of the primate. Journal of Neuroscience, 21, 700-712

Pandya, D. N., Yeterian, E. H., Fleminger, S. & Dunnett S. B (1996). Comparison of prefrontal architecture and connections. Philosophical Transactions: Biological Sciences, 351, 1423-1432

Passingham, D. & Sakai, K. (2004). The prefrontal cortex and working memory; physiology and brain imaging. Current Opinions in Neurobiology, 14, 163-168

Petrides, M. (1986). The effect of periarcuate lesions in the monkey on the performance of symmetrically and asymmetrically reinforced visual and auditory go, no- go tasks. Journal of Neuroscience, 6, 2054-2063

Page 44: Model free approaches to investigating the neural mechanisms of working memory

Petrides, M. & Pandya, D. N. (2001) Dorsolateral PFC: comparative cytoarchitectonic analysis in the human and the macaque brain and corticocortical connection patterns. European journal of Neuroscience, 11, 1011-1036

Petrides, M. & Pandya, D. N. (2001). Comparative cytoarchitecture analysis of the human and macaque ventrolateral prefrontal cortex and corticocortical connection patterns in the monkey. European Journal of Neuroscience, 16, 291-310

Petrides, M. (2005). Lateral prefrontal cortex: architectonic and functional organisation.  Philosophical Transactions of the Royal Society, 360, 781-795

Ramnani, N. (2006). The primate cortico-cerebellar system; anatomy and function. Nature Reviews: Neuroscience, 7, 511-522

Ramnani, N., Behrens, T. E. J., Johansen-Berg, H., Richter, M. C., Pinsk, M. A., Andersson, J. L. R., Rudebeck, P., Ciccarelli, O., Richter, W., Thompson, A. J., Matthews, D. R., Kastner, S. & Matthews, P. M. (2006). The evolution of prefrontal inputs to the cortico-pontine sstem: Diffusion imaging evidence from macaque monkeys and Humans. Cerebral Cortex, 16, 811-818

Silveri, M. C., Di Betta, A. M., Filippini, V., Leggio, M. G. & Molinari, M. (1998). Verbal short-term store rehearsal system and the cerebellum: evidence from a patient with a right cerebellar lesion. Brain, 121, 2175-2187

Smith, E. E. & Jonides, J. (1996). Neuroimaging analyses of human working memory. National Academy of Science, 48109-48110

Tombaugh, T, N. (2006). A comprehensive review of the Paced Auditory Serial Addition Test (PASAT). Archives of Clinical Neuropsychology, 21, 53-76

Walker, A. E. (1940). A cytoarchitectural study of the prefrontal area of the macaque monkey. Journal of Computational Neurology, 73, 59-86

Wolpert, D. M., Miall, C. & Kawato, M. (1998). Internal models in the cerebellum. TRENDS in Cognitive Sciences, 2, 338-348

Yates, S. L., Barach, A., Gingell, S., Whalley, H. C., Job, D., Johnstone, E. C, Best, J. J. K. & Lawrie, S. M. Parcellating the temporal lobes from magnetic resonance images using generic software in subjects at high risk of developing schizophrenia. (2006) Psychiatry Research: Neuroimaging. 147, 197-212