Neural Modeling of Brain and Cognitive Disorders

The following sections are included:

• PREFACE

• TABLE OF CONTENTS

During the last several years there has been a growing interest in developing computational models of phenomena associated with brain and cognitive disorders. Work in this area has included studies of Alzheimer's disease, aphasia and dyslexia, epilepsy, stroke, Parkinson's disease, schizophrenia, depression, and related problems. In this chapter, we give a broad overview of past work in this area as background for the specific topics addressed in the rest of this book.

A model of the hippocampus allows analysis of the role of network dynamics in the initiation and progression of neuropathology in Alzheimers disease. The model is neutral with respect to etiology, focusing on a final common breakdown in function termed runaway synaptic modification. This phenomenon could account for evidence showing that the neurofibrillary tangles associated with Alzheimers disease first appear and attain their highest concentration in subregions of the hippocampal formation, then successively spread into temporal lobe cortex and the cortex of the frontal and parietal lobes. The model demonstrates how the spread of neuropathology from the hippocampus into neo-cortical structures could result from the mechanisms of consolidation. Initial sensitivity of the hippocampus and entorhinal cortex to the development of neurofibrillary tangles is proposed to result from an imbalance of parameters regulating the influence of synaptic transmission on synaptic modification. Degeneration of cortical cholinergic innervation is proposed to result from exponentially increased demands on the feedback regulation of cholinergic modulation. Increased levels of amyloid are attributed to exponential increases in the modification and maintenance of synaptic connections, while development of paired helical filaments is ascribed to exponentially increased demands on the mechanisms of axonal transport or remodeling. Memory deficits are described as due to increased interference effects in recent memory caused by runaway synaptic modification which ultimately progresses to cause impairments of remote memory and semantic memory.

In the framework of an associative memory model, we study the interplay between synaptic deletion and compensation, and memory deterioration, a clinical hallmark of Alzheimer's disease. We show that deterioration of memory retrieval due to synaptic deletion can be much delayed by multiplying all the remaining synaptic weights by a common factor such that each neuron maintains the profile of its incoming post-synaptic current. This compensatory process can be realized either by global or local mechanisms. Our results show that neural models can account for a large variety of experimental phenomena characterizing memory degradation in Alzheimer patients. They open up the possibility that the primary factor in the pathogenesis of cognitive deficiencies in Alzheimer's disease is the failure of local neuronal regulatory mechanisms.

How can the effects of Alzheimer's disease on cognition be characterized? Experiments on confrontation naming have produced evidence implicating a wide array of underlying impairments, including visual and lexical impairments as well as impairments of semantic memory knowledge. These and similar experiments have also suggested that semantic memory is subdivided into components representing knowledge of categories and exemplars, and that AD primarily affects knowledge of exemplars. Using the concepts of distributed representation, interactivity and graded processing, we show how this evidence is also consistent with a simpler hypothesis: AD affects a single, undifferentiated semantic memory store. We also indicate the ways in which this simpler hypothesis can account for evidence initially taken to support an attentional impairment in AD, and can explain the apparently paradoxical finding of increased semantic priming in patients with AD.

Category specific language impairments have been postulated to require the existence of an explicit category organization within semantic memory. However, it may be possible to demonstrate analytically that this is not necessary. We hypothesize that category specific organization can emerge from perceptual, functional, and associative feature information about objects that is maintained in order to process language. In this paper, we conduct several experiments to test the computational validity of this hypothesis.

Physical objects were encoded in terms of semantic features, based on basic perceptual and motor modalities and higher level knowledge of function, for use in artificial neural networks. Mathematical methods were used to analyze the encodings and the neural networks. The results demonstrate the emergence of semantic categories in the networks, although such information was not pre-programmed. We conclude that category specific language organization can emerge from the inherent nature of semantic features themselves, and does not require special internal categorical organization of semantic memory.

An interactive activation theory of word retrieval in speaking is applied to the picture naming errors of aphasic patients. First, a model was parameterized to fit the probabilities of the major kinds of errors that normal speakers make in a picture naming task. Then, the model was fit to the error patterns of 21 fluent aphasic patients by altering its connection weight and/or decay rate parameters, each patient being fit individually. The fits were then used to derive predictions about other aspects of the patients’ behavior, in particular the influence of syntactic categories on the patients’ formal errors, the effect of phonology on semantic errors, error patterns after recovery, and patient performance on a single-word repetition task. Tests of these predictions were successful. Aphasic error patterns appear to be the result of quantitative alterations to normal processing parameters.

At least two processing routes in the brain are involved in pronouncing written words: a semantic route that derives the pronunciation via meaning, and a phonological route that derives it via spelling-sound correspondences. Simulations involving partial damage to an isolated semantic route (Plaut & Shallice, 1993) provide a comprehensive account of the rather peculiar combination of symptoms exhibited by patients with deep dyslexia, including the occurrence of semantic errors (e.g., reading RIVER as “ocean”), their co-occurrence with visual errors, and influences of imageability or concreteness on correct and error performance. Furthermore, when a version of the model is retrained after damage (Plaut, 1996), the degree and variability of its recovery and generalization are qualitatively similar to the results of some cognitive rehabilitation studies. The results challenge traditional assumptions about the nature of the mechanisms subserving word reading, and illustrate the value of explicit computational simulations of normal and impaired cognitive processes. They also suggest that connectionist modeling can provide a framework for generating specific hypotheses about strategies for rehabilitation.

The acquired reading disorder of surface dyslexia, in which lower-frequency words with atypical spelling-sound correspondences (e.g., PINT) become highly vulnerable to error, is presented in a framework based on interaction between distributed representations in a triangle of orthographic, phonological, and semantic domains. The framework suggests that low-frequency exception words are rather inefficiently processed in terms of orthographic-phonological constraints, because these words are neither sufficiently common to have much impact on learning in the network nor sufficiently consistent with the pronunciations of their orthographic neighbors to benefit from shared structure. For these words, then, the interaction between phonological and semantic representations may be especially important for settling on the correct pronunciation. It is therefore viewed as no coincidental association that all reported patients with marked surface dyslexia have also been profoundly anomic, suggesting reduced semantic-phonological activation. The chapter summarizes the simulation of surface dyslexia in the computational model of reading developed by Plaut, McClelland, Seidenberg, and Patterson (1996), and presents new data from three surface alexic patients. The graded consistency effects in the patients’ reading performance are more compatible with the distributed connectionist framework than with dual-route models maintaining a strict dichotomy between regular and exception words.

Computational models of oral reading have succeeded in simulating the general patterns of lexical and non-lexical reading disorders that occur secondary to focal brain injury. This study describes the simulation of more detailed aspects of the performance of dyslexic patients using a connectionist, dual-route model that employs competitive distribution of activation to control interaction among the model's components. Distinct simulated “lesions” are inflicted on the model to reproduce the interaction of orthographic regularity and frequency found in the word reading of Surface Dyslexics, and to model two patterns of sensitivity to the structural characteristics of non-lexical letter strings found in patients with Phonological Dyslexia. Study of the model's performance when lesioned provides a valuable source of information about how degraded information of different types combines to produce correct and incorrect reading responses.

Covert recognition, apparent knowledge without consciousness, has been described in several cognitive disorders, including pure alexia. Patients have been documented who cannot name briefly presented words, but can make semantic categorisation and lexical decisions to a degree of accuracy above chance. Here, we describe a simple connectionist model which, when lesioned, demonstrates many of the performance characteristics of pure alexia, including covert recognition. It is concluded that covert recognition can be explained in terms of the functioning of a damaged processing system, and it is not necessary to assume a disconnection of a normally functioning system from conscious expression.

Computational models have contributed to current understanding of normal brain function, and can offer new insights into the pathophysiology of neurological disease. We define some of the outstanding clinical questions in stroke, CNS injury, Alzheimer's disease, Parkinson's disease, and epilepsy from a neurologist's perspective, and discuss the potential impact of computational neurology on computational neuroscience. We consider representative examples of models constructed at several different levels of biological detail--from detailed membrane-level simulations to connectionist networks. We focus on what has been learned from several generations of models concerned with the after-effects of neural injury. In particular, we discuss how model assumptions, sometimes of ancillary importance, have constrained the ability to predict subsequent experimental results.

Phantom limbs are sensations of the presence of an extremity that has been lost. A number of clinical features and recent findings of cortical map plasticity after deafferentation suggest that phantom limbs are caused by large scale cortical reorganization processes. However, paraplegics, who likewise suffer from cortical deafferentation, rarely develop phantom sensations, and if they do, these are weak, lack detail, and occur after months. This has been taken to suggest a non-cortical genesis of phantom limbs. A biologically plausible minimal neural network model is proposed to solve this apparent puzzle. In trained self-organizing feature maps, deafferentation was simulated. Reorganization is shown to be driven by input noise. According to the model, the production of input noise by the deafferented primary sensory neuron drives cortical reorganization in amputees. No such noise is generated and/or conducted to the cortex in paraplegics.

We are using computational models to study how cortical maps reorganize following sudden, focal lesions. These models are explicitly intended to simulate small cortical ischemic strokes. Two prototypical models are described here. The first model has a topographic map of the hand region of primary somatosensory cortex. Following a sudden focal cortical lesion, portions of the hand originally represented in the lesioned area reappear in the perilesion cortex, as has been observed experimentally in animal studies. The second model, currently being studied, involves feature maps in proprioceptive and motor cortex regions that control a simulated arm moving in three-dimensional space. A sudden focal cortical lesion in this model can produce a very different result: a perilesion zone of decreased cortical activity. These two models make testable predictions, including that post-lesion map reorganization occurs in two phases, and that perilesion excitability is a critical factor in map reorganization. Current work is extending these studies to more realistic models incorporating biochemical and metabolic factors important in stroke.

Understanding the effects of damage on neural networks could lead to important insights concerning neurological and psychiatric disorders. We present a simple analytical framework for estimating the functional damage resulting from focal structural lesions to a neural network model. The effects of focal lesions of varying area, shape and number on the retrieval capacities of a spatially-organized associative memory are quantified, leading to specific scaling laws that may be further examined experimentally. It is predicted that multiple focal lesions will impair performance more than a single lesion of the same size, that slit like lesions are more damaging than rounder lesions, and that the same fraction of damage (relative to the total network size) will result in significantly less performance decrease in larger networks. Our study is clinically motivated by the observation that in multi-infarct dementia, the size of metabolically impaired tissue correlates with the level of cognitive impairment more than the size of structural damage. Our results account for the detrimental effect of the number of infarcts rather than their overall size of structural damage, and for the ‘multiplicative’ interaction between Alzheimer's disease and multi-infarct dementia.

The value of using minimal biophysical computational models is addressed and illustrated. We describe with such models the dynamical behavior of seizure-like rhythms in thalamic and hippocampal slices. The Hodgkin-Huxley-like models include the essential description of the intrinsic ionic channels, synapses and network architecture but they neglect unessential complexity. In the thalamus, we model the propagation of spindle waves in a network of excitatory thalamocortical and inhibitory reticular cells. As the wave advances, cells are recruited into the population rhythm, with reticular cells bursting almost every cycle at 7–10 Hz while thalamocortical cells burst only every few cycles. When GABA_A receptors are blocked all cells burst on nearly every cycle at 3-4 Hz, reminiscent of absence seizure-like activity. For the hippocampal slice we use a 2-compartment model for a CA3 neuron, a reduction of Traub's 19-compartment model. Currents for generating fast sodium spikes are located in the soma-like compartment; slower calcium and calcium-mediated currents are located in the dendrite-like compartment. Bursting occurs only for an intermediate range of the electrical coupling conductance between the compartments. A network of hippocampal CA3 neuron models coupled by fast AMPA and slow NMDA synapses produces multiple synchronized population bursts. We distinguish between the discontinuous, “lurching” nature of the thalamic waves and the continuous nature of the hippocampal waves. The rhythms in both cases are collective phenomena, since isolated cells are quiescent.

Various aspects of kindling — the phenomenon of generating epileptic seizures by means of repeated electrical or chemical stimulations — are modeled using a simple neural network. The model incorporates a number of biologically relevant features such as synaptic specificity, low average activity, synaptic delays and synaptic refractoriness. We argue that kindling of epilepsy occurs due to the formation of a large number of excitatory synaptic connections through a Hebbian learning mechanism. Several experimentally observed phenomena, such as the initial rapid growth and eventual saturation of the amplitude and duration of the afterdischarge, the insensitivity of the rate of kindling to the amplitude of the stimulus, drop in afterdischarge threshold due to repeated stimulations, and frequency dependence of kindling rate, are qualitatively reproduced by simulations of the model and explained. We show that one of the reasons for the termination of epileptic after discharge is that the neuronal network gets trapped in ‘normal’, low activity attractors of the network dynamics. It is argued that the first subthreshold shock induced afterdischarge is caused by the linking of originally present low activity attractors via stimulus driven synapse formation. New experiments, which would test the validity of the proposed mechanism of kindling and shed some light on the nature of memory storage in the brain, are suggested.

A network model of basal ganglia-thalamocortical relations during movement production is used to provide a mechanistic account for the motor deficits seen in Parkinson's disease (PD) and Hungtinton's disease (HD) subjects. The model is based on the anatomical, neurophysiological and pharmacological opponent interactions seen in the basal ganglia internal and external loops. Simulations of single-joint and multi-joint arm movements in PD and HD support the notion that the basal ganglia are involved in movement initiation and execution. Simulated lesions in the globus pallidus and sub-thalamic neurons suggest that although these focal lesions may improve some PD motor deficits such as rigidity and tremor, they may further reduce the movement modulatory capabilities in these patients. It is suggested that an approach that both reduces the tonic level of pallidal activity and restores the phasic modulatory capabilities of pallidal neurons may be an optimal strategy for the management of PD.

An approach to modeling normal and altered neural networks within the neocortex is described in this chapter. The approach is motivated by accumulating evidence that groups of neurons and networks functionally cluster together in the neocortex to encode cognitive and behavioral information. The clustering traverses many levels of spatial organization and operates on a multiplicity of time scales. This chapter reviews some of the evidence for, and theoretical modeling of, nested networks, and it postulates how simulated lesions in nested networks may provide insight into some disorders affecting the neocortex.

The concept of executive function in the brain is given a new definition, as based in a distributed neural system whereby prefrontal cortex is interconnected with various other cortical and subcortical loci. Executive function is divided roughly into three interacting parts: affective guidance of responses; establishing linkage among working memory representations; and forming complex behavioral schemata. Neural network models of these parts are reviewed and fit into a preliminary theoretical framework.

Recent studies have suggested that reduced corticocortical connectivity is associated with schizophrenia. Using neural network simulations we have explored parallel, distributed processing systems with reduced connectivity. These systems often behaved in a “schizophrenic-like” manner. Excessively pruned attractor networks became functionally fragmented, suggesting “loose associations,” and produced recurrent, intrusive representations suggestive of delusions. Pruning backpropagation simulations of speech perception networks produced spontaneous outputs which provided a model of hallucinated speech or “voices.” This model also suggested how dopamine-blocking drugs might reduce positive symptoms, and why negative symptoms arise in the wake of positive symptoms.

Schizophrenic patients show a variety of deficits in cognitive functions. These deficits have been difficult to understand in terms of a common unifying hypothesis. We describe a connectionist model of the function of prefrontal cortex and of neuromodulation of processing by dopamine. This model suggests that a single deficit may underlie poor performance in a variety of tasks: an impairment in maintaining contextual information over time and using that information to inhibit inappropriate responses. We tested schizophrenic on a new variant of the continuous performance test (CPT) designed specifically to elicit deficits in the processing of contextual information predicted by the models. The results confirmed the prediction.

We describe how a neural network model based on a system of ordinary differential equations can be used to replicate and predict depression recovery in response to specific treatments. A method for examining detailed patterns of clinical recovery using systems of second order ordinary differential equations is presented. We discuss why this method was chosen and how it can be used to reveal new information about how symptom response patterns differ across treatments. These approaches may be more broadly useful in other areas of cognitive and brain disorder research.