Modelling Executive & Attentional function in Alzheimer's Disease - 1999 MSc Thesis - M. Moutoussis

VI. Discussion

1. Rationale of the present study


This study is novel in bringing together rigorous neuropsychology, current theories of attentional control and computer modelling to bear upon the study of Alzheimer's disease. Although models of dysmnesia have been developed in AD, models of executive function have not. Rigorous neuropsychological experiments are seldom used as a basis for modelling in AD.

Even more novel and important is the use of dynamical systems theory to bridge the informational and biological domains. This is crucial as neural activity shows complex patterns, including oscillations, whose emergence is intimately linked to the brain's information processing ability. On the other hand, even simple psychological models have important, unexplored dynamical properties.

The present approach does not apply simply the minimal modifications that would fix model shortcomings, without reference to the biological substrate. Lack of such reference can be likened to trying to understand how a hummingbird flies by building helicopters: both might be able to hover, but the biological implementation may be fundamentally different to the engineering one. Models are better developed by examining hitherto neglected biological data while recognising that the relevant biology may be as yet unknown. Another objection to the 'minimal modification' approach is that it is Kuhn's 'normal science' par excellence (Notturno 1984), designed to avoid fundamental issues.

2. Critique of methods

'Executive dysfunction' in the dual task may not have the same origin or importance as the impairment of ADL related functions. Hence it may not be the most relevant starting point. The underlying attentional theory of Houghton et al is not the only possible starting point for modelling either. The information processing theory of EPIC (Meyer & Kieras, 1997) could be a different one; The methodology of Cohen et al (1992) yet another, particularly as a lot of work has examined the Stroop task in AD. The EPIC model was not used as it has very coarse graining of the biological processes involved. It involves however a detailed consideration of cognitive strategy, which has been neglected in our study. The Baddeley experiments did not consider differences in cognitive strategy and thus would be difficult to reconcile with EPIC. It would not be surprising if AD patients had difficulties inventing efficient strategies, and these would have to be considered in the modelling of more complex tasks. This is relevant to the conjecture that better strategies help the more educated preserve function in the face of AD.

Baddeley's papers provide little detail into the particular ways in which performance deteriorated under different conditions. They do not, for example, report false-positive responses or evidence of strategy failure. Such information could lead to more detailed modelling. Access to the original data might help. In addition the conclusion of an 'Executive lesion' in AD reached by Baddeley and assumed in this study is not the only possible one. An alternative could be deficient automatization. Rather than AD impairing executive control it could impair the process of automatisation (Cohen et al, 1992), so that AD subjects find dual tasks harder because it is harder to combine non-automatised tasks.

Modelling studies make many assumptions, which can be both a weakness and a strength. Unproved assumptions used in the Houghton derived models which could reduce the validity of our conclusions include that: 1. Performance is directly related to level of activation of cortical areas 2. The most relevant cortical areas are perceptual; 3. Diffuse influences (e.g. ACh) can be neglected; 4. Cortices can be treated in a 'lumped' manner; 5. Neural noise can be neglected; 6. The way in which 'match' unit performance is impaired is immaterial; 7. Cross-modality interference can be assumed to take place via descending projections of the attentional mechanism 8. Parameter exploration was adequate. The related strength of the approach is that such assumptions are inherent in much thinking about attentional control, only they remain hidden. On the other hand in this study they serve as incentives for meaningful exploration. The manner in which some of these assumptions were dealt with has been described in Methods. With respect to assumption 2., simulation of variables more directly related to behavioural performance would be highly desirable. Assumption 7 is of particular concern. It is adopted on the grounds of consistency with imaging studies showing suppression of non-attended modalities, of it providing a task-specific mechanism of attentional control and of it being a priori unbiased with respect to the study hypothesis. These assumptions will be discussed further.

3. Interpretation of findings

A. Models directly based on Houghton's theory

The 'attentional triad' proved a useful model. It is susceptible to attentional modulation, easily allowing the implementation of excitatory and inhibitory attentional control, including lateral inhibition and the match mechanism. It was, however, found to have important shortcomings. First, it can show spuriously excited states. Second, it is a lot more responsive to inhibitory rather than excitatory modulation. Third, the 'gain' units alter the level of activation but do not, in fact, change gain itself. These problems were traced (fig. 7) to the steady-state input-output function of the specific units involved. This curve has its maximum sensitivity at the resting state. The use of excitation functions which are concave upwards at the origin can solve these problems (fig. 8; section 3A of Results). This is an illustration of why knowledge of dynamical properties of simple unit combinations can be important.

In addition, the triad involves no conduction delays. More realistic, KI-like units improve the points raised above, have physiological support and also address the issue of time dependence in a more substantial way. They provide a relatively well-explored avenue to the simulation of oscillatory cortical dynamics. Objections can be raised: The KI equations have been derived from archicortical rather than neocortical physiology; and the timescales on which KI units operate do not relate to behavioural reaction times (their response is still too fast). Therefore an indirect index of their response would still have to be used to infer behavioural performance.

The limitations of the 'attentional triad' caused problems in dual-task simulations. If cross-modality inhibition is used, then performance of each task has to overcome inhibition by the other. This is difficult to achieve if excitatory control is inherently weaker than inhibitory. It can result in attention paradoxically decreasing performance during the dual task and in damage to the attentional system dis-inhibiting perceptual units more than de-exciting them. The result is a tendency for the AD condition to perform relatively better in the dual task than normals, except within a limited parameter range. This includes: 1. low stimulus intensities, which essentially linearise perceptual unit response; 2. within-pathway, excitatory synapses from match units to on-gain units much (e.g. four times) stronger than the cross-modality, inhibitory ones. It can be seen that both these properties serve to offset the tendency of the original ('Grossberg') units to saturate and to favour inhibitory inputs. These considerations also explain why even successful models showed quantitatively small effects.

Another drawback of the initial models is that progressive impairment of the 'match' mechanism does not lead to the profound effects of advancing AD; rather, impairment shows a floor effect. This can be understood if the positive feedback loop formed by the perceptual-match connections is considered. The participation of Grossberg-type units necessitates a low gain around the loop to avoid spurious states. An improved high-gain loop (as in fig. 8) would be more suitable for the simulation of the profoundly deleterious effect of AD.

On the other hand, the Houghton-based model with suitable qualifications succeeded in simulating qualitatively the overall pattern of the Baddeley results. It seems that the essential features of this model include: 1. the matching process 2. the presence of an active description of behavioural targets and 3. change of responsiveness of perceptual units towards stimuli through both inhibitory (decreasing responsiveness) and excitatory (increasing it) attentional control. Shortcomings of the original models include use of the Grossberg units and the symbolic implementation of the match / mismatch mechanism. Features that do not keep with cortical physiology include the zero transmission delays and system operation via point attractors.

B. The success of lateral inhibition : Interpretation and novel predictions of the augmented model.

The combination of direct lateral inhibition and match-based attentional control provided a strikingly robust model, simulating a number of results in a quantitative manner. This is because it involves excitatory attentional control based at the level of the match mechanism, with cross-modality inhibition at a lower level, less prone to the ravages of AD. This model leads to some novel predictions that are physiologically and psychologically testable. Cross-modality inhibition is independent of attention for a given sensory cortical activation. As in the Baddeley experiments it takes place between procedurally unrelated tasks it is likely to operate in a general, non-specific manner: different modalities laterally inhibit each other by default. Such laterally inhibitory effects would be equivalent in AD and normals as long as subjects attended to neither of the interfering modalities, performing instead some third, neutral task. Such effects should be visible in human imaging activation studies as well as in invasive animal experiments.

What could be the biological function of cross-modality interference effects observed in the normal elderly, and more specifically of lateral inhibition? Cohen and co-workers reach the conclusion in their discussion of automatic task performance that tasks should interfere only to the extent that they utilise overlapping brain modules, as opposed to some ill defined 'attentional resource'. Similar principles underlie the dual task performance limitations of the conceptually different EPIC (Meyer & Kieran, 1997). However, Baddeley's experiments specifically minimise overlap between resources required to perform the constituent tasks. Only then does the finding of interference between the tasks, particularly between response-to-tones and tracking, lend support to the theory of a 'central executive'. This is considered as a shared resource, the homuncular manager of a limited attentional 'budget'. There is, however, poor support for such a shared resource.

Cross-interference, so far modelled by direct lateral inhibition, could happen for a number of reasons. First, the 'central executive' could be a brain processor which can only deal in terms of serial procedures. The interference between increasingly more demanding tasks would arise because the only way of multitasking this serial processor is timesharing - and high performance biological serial computation has not yet evolved. This is broadly consistent with the analysis of H.A. Simon (1995), who claims that human thought is largely serial because parallel processing becomes inefficient in a general purpose reasoning device. Second, it may be that the organism doesn't know a priori that two tasks won't clash. The default is mutual inhibition, as per the direct lateral inhibition paradigm. Still, this isn't a good explanation for an inhibitory setup. The cross-inhibitory 'setup' might be an example of a synergy phenomenon. Synergies are ensembles of large numbers of biological components (e.g. muscles, sensory organs, nerves, central circuits) which, during a task, are coordinated and behave as a whole. As these are all functionally bound together they lack freedom to behave relative to each other except as prescribed by the pattern of coordination (Bernstein 1967, as quoted in Kelso, 1995; Kelso et al, 1984). It may be that the ability for coordination is a ubiquitous characteristic of behaving organisms. This ability is achieved through couplings between all system levels and components. In the presence of such couplings random relative component activity (as is forced by the random relative timing of the two tasks in the dual tasks paradigm) runs contrary to the general propensity for coordination. Contrary to the dual task paradigm, organisms are good at multi-task performance (e.g. walking, breathing and talking) if constraints of random relative timing are lifted. This may be important in AD, where walking while talking becomes difficult. Failure in this routine dual task may contribute to some dangerous falls (Camicioli et al, 1997). In fact many behaviours can be considered to involve multiple tasks having arbitrary but non-random relations to each other. Cross-modality inhibition could be not necessarily inhibition per se, but a vestige of cross-modality propensity for coordination. Again, however, impaired dual task performance is just a side effect.

Finally the lateral inhibition may have another function. In normal environments animals utilise multi-sensory inputs. It would be disastrous for mice to be impaired from hearing cats simply by looking for them. On the contrary, cross-modality control could highlight in the 'interfered' modality features of survival value in a context defined by the 'primary' modality, while inhibiting known distractors. Irrelevant cross-modality stimuli would be neither highlighted nor inhibited.

Thus several questions arise: What dual tasks show cooperative, and what neutral, effects in normals ? and what happens in these tasks in AD ?

4. Next step models

It is important to run the whole series of simulations in this study, and particularly the MAS ones, by substituting units most compatible with cortical physiology for the Grossberg ones. These in the first instance should be KI sets. It is also important to simulate output processes including the 'binding' between perception and response selection. This would allow direct comparisons with experimental results such as reaction times and recall errors. It may be possible to simulate the digit-span experiments of Baddeley by using the successful models of Burgess & Hitch (1996) for short term memory for serial order (fig. 17).

Figure 16. Combination of attentional control model with one that simulates short-term memory (STM) for digits. Detailed description of the latter is beyond the scope of the present discussion, with can be found in such as Hartely & Houghton, 1996. The model relies on retrieval of memorised items from a long-term memory store (LTM) through learned connections with 'context' units. Attentional control could enhance activation of both types of unit.


5. Reflections in the light of other findings

As the discussion of the possible roles and biological substrates for cross-modality inhibition has shown, not only is it important to consider model development but also fundamental assumptions, while retaining successful general features of the present theory. This can be done by considering what ways of implementing these successful features would be most compatible with the current understanding of the neurophysiology of AD. Successful matching in the Houghton models involves an increase in activation of both the 'match / mismatch' and the perceptual neurons. While both sensory and frontal cortices are thought to be activated by attention and by concurrent tasks, it is not clear how the pattern of activation changes in AD.

Some important studies (Leuchter et al, 1992; Schreiter-Gasser et al, 1993) largely gave rise to the hypothesis that intercortical functional connectivity is especially impaired in AD and that it may account for 'dysexecutive' neuropsychological deficits. These studies influenced our modelling of AD as reducing the influence of the 'match' on the perceptual units. These studies however measured levels of EEG synchronisation of cortical oscillations, rather than levels of activation, between different cortical areas. It may therefore be that the matching process depends on such oscillatory processing.

Three further lines of evidence, support this alternative view of matching. First, animal data suggest that neural structures closer to the sensory receptors receive bias, thus favouring some responses over others, through descending oscillatory signals (Kay et al, 1996; Sillito et al, 1994). This is in addition to the oscillatory signals of primary sensory cortex found during percept recognition (Gray, 1994). Second, during visuomotor tasks human EEG shows that successful trials are accompanied by increased synchronisation between different but relevant brain areas. Gevins & Cutillo (1995) describe increased evoked potential covariance between left prefrontal cortex and the relevant motor and parietal cortices preceding successful responses. Third, Bressler and coworkers (1993) showed how attention-demanding tasks involve intermittent synchronisation of oscillatory activities among multiple brain areas in the monkey. Synchronisation takes place very robustly, and in discrete time frames, over cortical areas of obvious relevance. However the findings are difficult to interpret because synchronisation epochs do not map to information processing stages in an obvious way (unlike the work of Kay et al in the olfactory system). Synchronisation may indicate 'consensual resolution of processing', whereby synchronous activity in two or more areas allows amplification of the signal transmitted to common targets that they project to (Bressler, 1995).

Thus matching may involve synchronisation and AD could disrupt it through functional disconnection, as demonstrated by the EEG studies. This could explain attentional impairment in a general way. Cross-modality interference need not become attenuated: Leuchter and coworkers (1992) clearly showed that coherence of oscillatory activity between areas connected by broad, complex networks is preserved. Most interestingly, imaging studies (Becker et al, 1996) have shown that AD subjects activate cortical areas according to task demands more diffusely than controls, activation 'spilling' into neighbouring areas. Synaptic compensatory changes that account for the excess of false-positive events during memory tasks in AD (Ruppin & Reggia, 1995) could also account for this 'spill-over' effect. With respect to dual task performance, such 'spill-over' would be consistent with a preserved or enhanced cross-task interference.

In summary, a more satisfactory model of cross-modality interference could include the physiological feature of oscillatory synchronisation and, in AD, an impairment of the balance between long-range functional connectivity and more local interference. This is in addition to the target units, the matching mechanism based on recurrent connections and the perceptual mechanism modulated by attention central to the successful psychological models.

6. New directions for further research

A. Structure of a revised model

The overall structure of the models used in this study (fig. 3) can be retained in incorporating the above improvements. However, the cross-modality inhibitory attentional projections are now omitted and local (between areas of the same level) interference is postulated. The 'match' units are again identified with association cortices. Each lumped cortical area is now capable of oscillatory activity. Activation of motor schemata is through 'consensual resolution of processing'. A diagram of the revised model is shown in fig 6a

The pattern of the Baddeley results permits however drastic simplifications to be made. In both all subjects the influence of the primary on the secondary task was much greater than vice versa. We can therefore ignore the latter influence. Also ignoring motor schema binding gives a preliminary model of only two coupled oscillators. This bare bones model is shown in fig. 6b. Here the influence of the primary task is through a non-specific broad band signal.

B. Evaluation of the revised model

Simulations showed that the model is successful in simulating the overall pattern of experimental results. Attentional activation improves its performance (indeed for certain ranges of parametres attention is necessary for model response). Simulation of AD by functional dysconnection of modules gradually impairs performance, disproportionately so in the presence of an interfering task (This is a summary of preliminary results. Please email me for  fuller info). In the normal case the presence of interference does not greatly reduce performance.

 If information processing in the real brain involves a series of intermittent synchronisations like the one successfully simulated here, the demonstrated delays in synchronisation might be possible to directly link with prolongation of reaction time. The problem here is that the broad band nature of cortical oscillations does not clearly reveal a timescale with which the period of oscillation of the simulated areas can be identified.

The revised model is still inadequate in simulating physiology. Possible important features that are omitted are the explicitly episodic nature of brain synchronisation; the detailed description of the oscillators used; and most importantly, all structure of information processing that the real episodic synchronisations reflect. Another omitted feature is the non-negligible transmission delays of real intercortical pathways. The model however captures enough of the physiology to support the hypothesis in a preliminary manner.

C. Education & Alzheimer's disease

Many aspects of AD related to executive function particularly lend themselves to modelling. A most important example has to do with the possible effect of lack of education, an independent risk factor for the development of AD (Orrell & Sahakian, 1995). Functional imaging indicates greater deterioration for a given degree of cognitive impairment in the brain areas subserving attentional and executive functions of the better educated (Alexander et al, 1997). This supports the hypothesis that better education is associated with greater 'cognitive reserve'. The biological and psychological substrates of the protective effect of education are however little understood and could in the future be explored according to our paradigm. This might be of preventative value.

7. Summary & Conclusion

In Alzheimer's dementia research has recently explored not only the memory but also the so called 'dysexecutive' deficits (Baddeley et al, 1991). The latter are thought to reflect difficulties in the coordination of daily activities that compromise the independence of patients with AD at least as much as memory deficits. The understanding of these executive deficits is thus of great importance. Baddeley and co-workers (1991) postulated a Central Executive System (CES) coordinating attention. Using the dual-task paradigm they showed that the CES is particularly affected in DAT, accounting for the severe attentional /executive deficits (Baddeley et al, 1991). These experiments provided the necessary data to constrain the development of rigorous models of executive control of attention.

Houghton and coworkers (Houghton & Tipper, 1994) have simulated successfully many aspects of selective allocation of attention. We formulated the hypothesis that "an application of the model of attentional control of Houghton et al can account for the pattern of deficit observed by Baddeley in AD patients during the performance of dual tasks". We aimed further to improve the model heuristically but also on the basis of current neuroscientific findings.

Two attentional control systems, one for each dual task modality, were combined (fig. 3). The two systems were linked by cross modality inhibitory attentional control (Houghton & Tipper, 1994). Damage due to AD was modelled as an impairment of the influence of the attentional ('match') module on perceptual areas, guided by the neuropathological finding that in early AD association areas and corresponding long range projections are damaged (Morris, 1994).

It was found that within a small parameter range the model could reproduce qualitatively the general pattern of experimental results. However it had a tendency to show spuriously activated states, lacked robustness, had difficulty reproducing results in a quantitative way and failed to simulate the progressive deterioration of executive function in AD. Dynamical analysis showed that the units used to simulate individual cortical areas in the original models were largely to blame for these failures. The model could be dramatically improved by the introduction of lateral inhibition between sensory areas, postulated to be largely unaffected by AD. Preliminary studies showed than association function based on synchronisation of oscillations between different brain areas, known to deteriorate in AD, could also account for Baddeley's results.

The study has contributed to the understanding of psychological mechanisms of executive function particularly as involved in AD. It has demonstrated the importance of dynamical understanding of psychological models. It brought together rigorous neuropsychology and computer modelling, something not previously attempted in this field but essential for the linking of the pathology with the clinical features of AD. It has also indicated directions for future research, mainly the investigation of the 'matching' function in the context of oscillatory cortical dynamics and the disruption of intercortical coordination of neuroactivity in AD.

VII. Acknowledgements

This thesis would have been impossible without the help of Drs. Martin Orrell and George Houghton. I would further like to thank Dr. Ann Moutoussi and Dr. David Frost for precious glimpses into their world-views, and the staff of the Medical Education Centre, Whipp's Cross Hospital, Leytonstone for their help with literature searching and provision of reference papers.


Next part of thesis
     Contents  Abstract  Introduction  Aims    Methods  Results    Discussion
 Acknowledgements  References   Appendices

Here Endeth this page !      Home