Thesis for the degree of Master
of Science in Psychiatric Theory and Research Methods
Michael Moutoussis firstname.lastname@example.org
at the Department of Psychiatry, University College London Medical School, Wolfson Building, Middelesex Hospital, London W1 8AA ,
Supervised by Dr. Martin Orrell, Dept. of Psychiatry and Dr. George Houghton, Dept. of Psychology
: In Alzheimer's disease (AD) executive and working memory deficits
often compromise the safety and independence of patients. Such executive
functions are understood to organise working memory. The latter has been
described as consisting of "slave" systems co-ordinated by a Central Executive
System (CES) that controls attention. The CES is particularly affected
in AD. Computer models have contributed greatly to the understanding of
selective attention, but not with reference to AD. We explore the potential
for an understanding of the loss of executive function in AD using computer
modelling and we review recent research from a variety of fields whose
integration is central to this approach. Research has suggested that high
educational achievement protects against the development of AD, possibly
by helping preserve such executive functions. This finding is explored
in the light of the modelling studies. Executive function in AD has been
investigated experimentally, most notably by the group of Baddeley and
colleagues. Dual task performance in normal elderly controls has been compared
with that of AD patients. In the AD group, performance in any task was
reduced by the presence of a concurrent task. With the passage of time
the effect of any concurrent task on any primary one increased for the
AD group but not for the controls. This demonstrated a specific executive
deficit in the patient group.
: It is feasible to develop rigorous models of 'central executive'
function. The main hypothesis of our study is that the application of the
computational model of attentional control of Houghton and co-workers could
account for the pattern of deficit observed in AD patients.
: Dual task performance data was simulated by using models of attentional
control derived from those of Houghton and co-workers. In these models
for each sensory modality bottom-up and top-down signals are compared in
'match-mismatch' modules. The output of these modules can excite or inhibit
lower level sensory modules towards particular input features. Two attentional
control systems, one for each modality, were combined in the present simulations
to account for dual task performance. Sensory and 'match-mismatch' modules
were linked by cross modality inhibitory attentional control whose implementation
here was guided by the neuropathology of early AD. Key building blocks
of the original Houghton model were analysed mathematically to explain
the behaviour of the overall models. Alternative building blocks and structural
modifications were explored.
1. The models adequately simulated performance of controls.
2. Reduction in single task performance in AD was also successfully simulated.
3. Reduction in dual relative to single task performance in AD was only simulated for narrow parameter ranges (was not robust).
4. Model capacity for gradual performance deterioration was limited.
5. Positive feedback loops involving units such as those used by Houghton have drawbacks such as spurious activated states and limited capacity for increases in gain.
6. These problems can be overcome by using units with a flatter response at their rest state.
7. Incorporation of direct lateral inhibition at the level of sensory cortex greatly improves model performance.
Preliminary implementation of the matching mechanism as synchronisation
in the oscillatory activities of two brain areas and of lateral inhibition
as oscillatory noise interference between modalities is successful in qualitatively
simulating single and dual task performance for normal controls and for
sufferers of Alzheimer's disease.
: The models described here were successful in simulating many features
of attentional control. However, they do not robustly predict the experimental
results without substantial modification. When Houghton's theory is reviewed
most of the conceptual essentials are retained, but many important details
of implementation are rejected. More physiological details of neural masses
help improve models. It is argued that direct lateral inhibition is unlikely
to be a biological default between modalities: its essential point is likely
to be the presence of pathway interference at specific levels (rather than
across levels) which is preserved or even augmented in AD. In the light
of recent neuroscientific findings, a modified theory is suggested that
postulates 'Match / mismatch' function to involve oscillatory synchronisation.
Interference within levels (similar to direct lateral inhibition) is hypothesised
to impair synchronisation between levels. Preliminary simulations support
this theory but more extensive exploration is required. It is envisaged
that in the future the study of executive function may aid the assessment
of the ability for independent living of patients and the understanding
of the protective role of education in AD.
The importance of executive functions in Alzheimer's disease
is one of the most frightening and costly diseases to affect humanity,
one that eventually destroys most mental abilities. Memory deficits are
prominent but not necessarily sufficient to impose severe restrictions
in the patient's lifestyle and independence. When, however, executive as
well as mnemonic functions are affected, so that the patient cannot carry
out the schemas necessary for daily activities, the patient puts herself
in danger. An example would be having difficulty coordinating the sequence
of actions required to make a cup of tea. It is therefore of great practical
importance to understand and support those mental mechanisms concerned
with executive and attentional functions.
into Alzheimer's disease (AD), the commonest cause of dementia, has therefore
explored in recent years the 'executive' deficits (Baddeley et al, 1991).
The term `executive' is used in different ways by different researchers
and in different contexts, in line with the ignorance characterising new
fields of enquiry. Jaak Panksepp (1998) defines: " 'Executive system' implies
that a neural system has a superordinate role in a cascade of hierarchical
controls ". 'Executive' is applied to planning, sequencing, control of
attention and of working memory and specific functions such as cognitive
set shifting or suppression of dominant tendencies.
Psychological theories of executive control of attention
functional modules, such as the Supervisory Attentional System or the related
Central Executive System (CES) have been proposed to malfunction in AD
in order to account for executive deficits (Baddeley 1996). According to
such psychological models a 'Supervisory Attentional System' is hypothesised
to co-ordinate the lower level, or 'slave', components of working memory,
i.e. the 'Phonological Loop' and the 'Visuospatial Scratchpad'. Since Baddeley's
pioneering work there has been active debate as to the modularity and brain
localisation, of functional modules that subserve and co-ordinate working
memory (Wickelgren, 1997).
Dual Attention and the Central Executive System
and colleagues (1991) examined executive function in a series of experiments
which form the basis for the present analysis. They attempted to test whether
AD disproportionately affects Central Executive allocation of attention,
so that dual task performance is affected more than the constituent tasks.
They compared patients with mild AD with normal elderly controls and followed
them up over one year. In these experiments a primary tracking task, following
a randomly moving white square, was combined with graded secondary tasks.
The difficulty of the tracking task was first individually adjusted in
the absence of secondary tasks so that subjects managed to stay on the
square for 40-60% of the time. The difficulty was then fixed. The simplest
concurrent secondary task was 'articulatory suppression': the subject counted
repeatedly from 1 to 5 during pursuit tracking. The next stage was reaction
time to tones : the subject had to press a foot switch as soon as an auditory
stimulus was presented. Reaction times and percentage of missed tones were
recorded. The final stage was a memory span task. In this the maximum length
of a random digit sequence that the subject could reliably repeat back
was determined in the absence of pursuit tracking prior to each of the
three testing sessions. Sequences of this 'subject-tailored' maximum length
were then presented during pursuit tracking and recall performance recorded.
results obtained were consistent with the hypothesis. In the AD group,
performance in any task was reduced by the presence of a concurrent task.
A key result was that with the passage of time the effect of any concurrent
task on any primary one increased for the AD group but not for the controls,
supporting the hypothesis of a specific executive deficit in the patient
Functions preserved in AD - Implications for Attention and Planning
a series of experiments complementing those of Baddeley, Simone & Baylis
(1997) showed that executive function was impaired in AD using a paradigm
involving suppression of a dominant response tendency. Subjects had to
quickly respond to a green light, but ignore a yellow one. AD subjects
showed many false positive responses, consistent with weak executive control.
To test for the cognitive level at which errors occurred, subjects were
asked how sure they were that they responded correctly. AD patients were
aware of their erroneous choices. The authors concluded that the executive
deficit did not involve early information processing but rather the efficient
implementation of a response.
Dementia, education and executive function
epidemiological findings in dementia are related to the above concerns.
educational achievement appears to be an independent risk factor for the
development of dementia (Orrell & Sahakian, 1995). The aged of lower
educational achievement more commonly fail tests of activities of daily
living (Zhang et al, 1990), while the better educated require greater damage
to cortical areas important for executive function to get as impaired (Alexander
et al, 1997). The mechanisms underlying the protective effect of education
are however poorly understood. Mathematical modelling could help clarify
Mathematical Modelling in Psychiatry: some difficulties
most sciences modelling has an important role between experimental studies
and theoretical analysis. It is not, however, obvious that mathematical
modelling has much place in the neurosciences in general and in psychiatry
in particular. Psychiatric theories are often conceived in qualitative
terms and expected predictions are derived on the basis of semantic inference
and common sense. However, many theories are not precise enough (they may
be underspecified) to derive accurate predictions from; or the system involved
may be too complicated for one to derive predictions by common sense (the
theory may be intractable).
rigorously specified model has many advantages : First, it forces a more
complete description of the problem. The key variables have to be defined
and theoretical assumptions become explicit. Second, explicit alternative
explanations of the data on which the model is based can be formulated.
Third, detailed predictions can be made: therefore falsification of a rigorously
defined model is easier i.e. the theory provides better means for its own
falsification (Notturno, 1984). Fourth, counterintuitive predictions of
the theory may become apparent. These may explain already existing data
difficult to understand on the grounds of common sense. It may also be
rigorously tested whether data seemingly contradicting the theory could
in fact be compatible with it. Fifth, ignorance about the theory can be
quantified, e.g. in terms of model parametres. Results of further experiments
can be anticipated and such experiments planned. Finally, once successful
models are developed, numerical experiments can be performed that would
be too difficult to carry out in vivo.
Although mathematical modelling is thus indispensable in the hard sciences, it has not fared as well in psychiatry. There are several reasons for this. Biological systems are so complicated that people think that a lot more needs to be known about them before meaningful modelling can take place. It is often unclear that biological systems have laws possessing explanatory power independent of the system's fine structure. It is also unclear how components and their interactions at any particular level of description give rise to collective properties in a non-trivial manner. Indeed, the theory of evolution favours a top-down, 'how is this feature teleologically reasonable', view over the reductionist, bottom up modelling approach. In addition it is sometimes possible to construct several different models that explain the data. These models may be difficult to tell from each other by experiment or may be poor at making novel predictions.
Mathematical modelling is culturally alien to psychiatrists; most importantly the questions that could be investigated by modelling simply do not occur to us. Our heuristics of science do not include the values of rigour, unification of ideas, economical description of phenomena and beauty of mathematical structure that guide the hard sciences.
biological phenomena usually involve self-organising structures that consume
energy. In such systems measured variables do not change in proportion
each other, i.e. the systems are non-linear (Nicolis, 1991; Nicolis, 1989;
Kelso, 1995). The theory of nonlinear systems has only recently emerged
and its neurobiological applications are still at an early stage.
7. Modelling efforts in Alzheimer's disease
these obstacles, considerable inroads have been made in the modelling of
AD. Connectionism has aided the understanding of psychiatric disorders
and of AD in particular. A connectionist model consists of interconnected
units representing neurones or groups of neurones. Units can have simple
internal structure yet their networks can perform complex functions, made
possible by appropriate patterns of synaptic interconnections (Jeffery
& Reid, 1997). This is an anatomically inspired approach that can be
used to model brain function at the level of groups of cells (Traub et
al, 1997) or of individual cortical areas (Freeman, Yao & Burke, 1988).
The role of neurotransmitters can be elucidated. Models can be used to
investigate the transition from the level of nerve activities to that of
objects of perception and action (Freeman, 1991). Contextual meaning and
hence emotion can be taken into account (Armony et al, 1995). At the 'highest'
(symbol manipulation) level connectionist models can be used to simulate
objects of cognition such as memories (Hartley & Houghton, 1996). At
the neurotransmitter level they can help explain observed psychopathology
(Cohen & Servan-Schreiber, 1992).
first goal of modelling in AD has been the understanding of dysmnesia.
Pioneering studies (Carrie, 1993) used highly abstract connectionist models.
These explained the distributed storage of memories, their initial resistance
in the face of gradual neuronal loss (simulated by unit deletion in the
network) and their subsequent smooth decline. The use of fairly realistic
learning algorithms allowed important shortcomings of the model to be identified.
The greater impairment in new learning relative to memories laid before
'atrophy' could not be explained, and gradual deletion of neurones from
biologically more realistic networks failed to produce a gradual decrease
in recall. Performance remained intact until a large proportion of network
elements were lost, then dropped catastrophically. To simulate the gradual
course of the illness it was necessary to introduce synaptic compensation
as found in the real brain (Ruppin & Reggia, 1995). Synaptic compensation
is a biological constraint which permits the models to explain the gradual
degradation of recall performance, the differential sparing of remote memories
and the increased rate of false positive retrieval errors found in AD.
contrast to these sophisticated studies of memory little attention has
been paid to the modelling of executive function, despite its clinical
importance. Modelling loss of synapses might explain weakened executive
control while synaptic compensation might help explain the excess of false
positive responses, accounting for the findings of both Baddeley and Tipper
8. Mathematical models of Attentional Control
starting point for the present simulations is the theory of attentional
control of Houghton and colleagues (Houghton & Tipper 1994; Houghton
1995). In these models each sensory modality consists of low level sensory
modules, high level modules where the behavioural goals or targets of the
organism are stored and intermediate, 'match-mismatch', modules that compare
percepts with targets (fig 1a).
1b. Connectivity that implements the Attentional Control mechanism.
Filled circles are inhibitory synapses, arrowheads excitatory (as in following
figures). The Target field units feed to the corresponding Match/Mismatch
unit pairs. For each target feature (see 7) there is a unit pair (6) calculating
whether the feature is present, on the basis of input by the corresponding
Property unit (e.g. unit d gives bottom up input 5). If a match is present
the corresponding Match unit activates the On unit of the Property unit
coding the same feature (1) and inhibits its sister Off unit (2), thus
increasing the responsiveness of the Property unit. If a mismatch is detected
the Mismatch unit becomes active and reduces responsiveness (3, 4). Negative
feedback tends to reduce the activation of Property units subject to mismatch.
Coactivated features then cooperate to dominate perception and activate
an appropriate response. If for example property units a and b belong to
the same object, and unit a is activated, it tends to facilitate activation
of b both directly (12) and indirectly (10,11). Finally an emergent assembly
of features, signifying an object, activates a response schema.
output of the 'match-mismatch' modules can excite or inhibit the lower
level sensory modules towards particular input features, and thus attend
or ignore these particular features: this is the crucial mechanism of attentional
control in the model. The activated features are then combined into attended
models of Houghton and co-workers can simulate a large number of experimental
data on selective attention. These include perceptual distracter processing,
negative priming, response binding in the presence of distractors and inhibition
of return. The models are built around important principles such as attentional
control of perceptual gain regulation, opponent processing and competitive-cooperative
intra-module interactions. Drawbacks of the model include first, its loose
basis on physiology. This is most evident in the match / mismatch module.
Its units perform logical computations whose physiological implementation
is quite unclear. Secondly, the whole model is constructed on the basis
of information-processing constraints. This is no bad thing in itself but
it ignores how dynamical constraints can give rise to structure. Levine,
Parks & Prueitt (1993) in their review of the methodology of simulation
of `frontal' cognitive functions expect that functionally and structurally
distinct levels of brain activity are separated by underlying dynamical
constraints. Thirdly, the Houghton model is over-stable. Once perceptual
input is terminated object assemblies tend to persist. This necessitates
the introduction of decaying synaptic weights between units of the Object
field (connections (9) and (10) in fig. 1b). Target node activities also
have to be made to decay. Freeman (1992) has discussed how positive and
negative feedback loops in neural systems need not convey excessive stability
if allowed to operate within oscillatory regimes.
- dependent allocation of attention has been studied by Cohen and coworkers
(1992) using a model of the Stroop task. In this task subjects are presented
with dual stimuli, e.g. the name of a colour spelt out and the ink colour
in which the word is written. They are asked to either read what the word
says or to name its ink colour as quickly as possible. Each 'modality'
(reading vs. colour naming) has its own 'pathway' in the model. Both are
influenced by a 'task demand' module which the authors identify with a
function of the prefrontal cortex. Cohen finds that 'attentional selection
can be thought of as the mediating effects that the internal representation
of context [here, of the instruction the subject has received from the
experimenter] has on processing' (fig. 2).
to the theory of Houghton and coworkers, the model of Cohen et al includes
a simpler 'Matching' mechanism (another way of looking at the function
of the hidden units !) capturing essential biological plausibility. It
doesn't, however, include any cross-modality interaction except at the
output level; Its simplest modification to conform to the dual task layout
would therefore completely decouple modalities. It could therefore not
predict dual-task interference effects. An important finding from the work
of Cohen and coworkers is that 'the degree to which a process relies on
attention is determined by the strength of the underlying pathway'. This
may mean that a disproportionate attentional deficit may be not because
a 'Central Executive' is particularly affected in AD but because eroded
underlying pathways would take stronger attentional modulation to perform.
Figure 2a. Modular structure of the model of attention in the Stroop task according to Cohen and coworkers. External input is received by the input units, C ; In order for it to activate the output units, D, activity passes through the intermediate, 'hidden' units B. These are in part activated by units A which express the context to which the organism must give priority. These are similar to the 'target' units of the Houghton models (fig. 1) in that they are independently and externally activated.
Figure 2b. Detailed structure of Cohen's Stroop task model. All the synapses in or by the 'reading pathway' are shown. The connections of the colour naming pathway are similar but weaker. Note the absence of cross modality inhibition.
2c. The 'hidden' units of the Cohen model perform a graded AND (or
'match') function by becoming activated, and therefore allowing signal
to propagate along their respective pathways, when not only bottom-up excitation
is present from the stimulus but also descending facilitation.
9. Cerebral Localisation of Supervisory Attentional / Central Executive processes
concurrent tasks involved in the Baddeley experiments were chosen so as
to minimise competition for local resources and thereby to highlight the
possibly shared 'Central Executive' requirements. Thus the primary task
- tracking - is presumed to involve the 'visuospatial scratchpad' slave
system while the secondary the 'phonological loop' for example. Of course
both are shown to depend on attention 'allocated' to them by the presumed
central executive. While earlier imaging studies showed little overlap
between the brain areas activated by verbal and non verbal working memory
tasks (Raichle, 1993), more recent studies have concluded that the dorsolateral
prefrontal cortex contributes to the maintenance of both verbal and nonverbal
information (Fiez et al, 1996). Recent imaging studies indicate that rCBF
decreases in sensory areas that are not to be attended (Drevets et al,
1995), in agreement with theoretical studies that place emphasis on inhibitory
mechanisms in attention.
deterioration of the cholinergic system in AD is thought to affect attention
significantly. Sahakian has demonstrated that when cholinergic function
enhancing drugs improve neuropsychological performance in AD this is attributable
more to an improvement of attentional rather than memory function (Lawrence
& Sahakian, 1995).Imaging studies indicate that Scopolamine, a cholinergic
antagonist, attenuates memory-task-induced increases of rCBF in the right
anterior cingulate but also bilaterally in the prefrontal cortex.
together these findings support a model of AD where elements corresponding
to particular aspects of prefrontal function are either damaged or functionally
dysconnected (Schreiter-Gasser et al, 1993; Morris, 1994) to the rest of
the model. The primary candidates for this, in Houghton's terms, would
be the 'target field' and the 'match-mismatch' (fig. 1a). Damage to either
would result in reduced attentional control on lower-level units, a priori
equally impairing excitatory and inhibitory control.
Physiological basis for modelling
most connectionist psychological models units are very simplified in biological
terms, being derived more on the basis of constraints from psychology.
This does not by itself imply that the functional approximation is crude,
as biological systems are self-organising, far from equilibrium and hence
their emergent properties at any one level may be robust within a functionally
important range of conditions. However the adequacy of the approximation
is always a matter of concern.
abstract models are used to provide tractable equations for modelling neuronal
elements. Often used is the (logistic) sigmoid neuron where output is a
sigmoid function of the sum of all inputs to the neuronal soma within a
preceding short time interval. Continuous time models usually include a
differential operator acting on neuronal state variables equated with a
sigmoid function of inputs, which may involve delays. The prototype is
the model of Wilson & Cowan (1973). The units used in the work of Houghton
et al involve first order operators and no delays. More physiological alternatives
range from the second-order-operator models of neuronal populations of
Freeman (1975) to simulations of large number of neurons with details of
their ionic currents and other biophysical properties (Traub et al, 1997).
I shall consider the logistic sigmoid neuron as a minimum requirement for
physiological relevance and more complicated models when detailed time
evolution of the system is simulated.
Dynamical analysis and biological models
is important to analyse the performance of the neural systems involved
in attention not just in terms of facilitation and inhibition of units,
but in terms of attractor dynamics. This involves considering what patterns
of activity are open to the network, what is the stability of such patterns
and how the 'landscape' (phase space) of all such available patterns changes
under different conditions. We may, for example, think that a stable performance
of the primary task corresponds to an attractor set of the entire network
but one with specific directions of relative instability. Perturbations
along such directions (corresponding, for example, to the auditory signal)
can 'flip' the system into another, possibly transient, 'attractor' set
(corresponding to auditory perception, match and output). The motivation
for thinking of this model in terms of attractors and their stability comes
from several sources.
the model of Houghton et al is, as discussed, over-stable. Its authors
thus introduced modifications with little psychological or physiological
support. Dynamical analysis of the model's stability could, alternatively,
clarify the causes and solutions to the problem in a less ad hoc way.
while the attractors that are used in most psychological models are point
attractors, the ones found experimentally in investigations of thalamocortical
interaction and those extensively studied in limbic structures (e.g. entorrhinal
cortex, heavily involved in AD) are par excellence oscillatory (Kay, 1996;
Gray, 1994). There is evidence for thalamic and limbic areas being involved
in attention (Forstl & Sahakian, 1993) and particularly in the matching
process. Freeman and co-workers have demonstrated dramatically the effect
that attention and behavioural significance have on the patterns of oscillatory
activation of the olfactory cortex (Grajski & Freeman, 1989, Eeckman
& Freeman, 1991, Kay,1996). An important component of the binding of
sensory features into visual percepts also relies on oscillatory attractors
(Gray, 1994; Bressler, 1996). In the present models oscillations would
arise naturally if realistic synaptic delays were incorporated in the existing
negative feedback loops. Superposition of oscillatory attractors, switching
between such attractors and global modulation of oscillatory cortical activity
differs from equilibrium (point attractor) dynamics as applied to the same
brain functions. It is thus important to consider our present models in
terms of attractor dynamics to prepare the ground for incorporation of
the above findings. Introducing oscillatory dynamics must however be necessitated
by both the probable applicability of the physiological findings and the
need to use that level of description to overcome limitations of the information
processing approach used so far.
reason to consider analysis in terms of attractors is illustrated by the
dynamical analysis of errors in a neural network model of dyslexia. This
has demonstrated that dynamical analysis can explain some counterintuitive
experimental findings that are replicated by neural network models (Hinton,
Plaut & Shallice, 1993). Burgess and Hitch (1996) claim that a model
that replicates experiment but "cannot be mapped onto a conceptual understanding
of the processes giving rise to behaviour is useless". The dyslexia model
demonstrated that this conceptual understanding, and therefore the usefulness
of such a model, may well depend on the understanding of the attractors
To develop the theory of normal 'central executive' function.
control may depend on the matching between sensory representations of perceived
objects and partial, 'target 'representations of sought objects (Houghton
& Tipper, 1994). The central hypothesis of the present work is that
models of this matching process can simulate executive attentional control
during dual tasks.
aimed to simulate the influence of attention in normal subjects during
single- and dual- task conditions by bringing together previously developed
models of attentional control ( Houghton & Tipper,1994) with the dual
pathway concept of Cohen and co-workers (1992) .
reaction time experiment could be simulated with a model of attentional
control with modality specific pathways (fig. 3) . This will be used as
the nucleus of the present study as it includes all the important experimental
findings. Once this is adequately modelled, other dual task experiments
could be simulated as further tests of the main hypothesis.
To simulate dual task performance deficits demonstrated in AD.
introducing neuropsychologically plausible 'lesions' in the above model
I aim to replicate the experimental findings in mildly demented subjects.
AD lesions can be simulated as an impairment of the match / mismatch field.
The modules situated closer to the sensory and motor interfaces will in
this approximation be treated as intact, in line with the localisation
of such modules within sensory cortices, preserved in early DAT.
in the Baddeley experiments, a greater attentional control deficit (corresponding
to more advanced AD) should lead to increasing divergence between single-task
and dual-task performance. I aim to show that progressive 'damage' would
suffice to account for the deterioration of DAT subjects over time.
to the model proposed here does not simulate the AD effects in a trivial
manner. Reducing the efficiency of cross modality inhibition might decouple
modalities, and dual task performance may improve relative to single task
performance. This possibility makes the model proposed more falsifiable.
To identify important limitations of the proposed models and to suggest
ways to overcome them.
model components or architecture are found to limit ability to simulate
experimental results, I aim to improve on these not only by introducing
the minimum sufficient modifications but also by guiding such modifications
by appropriate physiological considerations.
Click here to move on to the next
part of the thesis
Acknowledgements References Appendices
Here Endeth this page ! Home