Blindness, Psychosis, and the Visual Construction of the World

Abstract

The relationship between visual loss and psychosis is complex: congenital visual loss appears to be protective against the development of a psychotic disorder, particularly schizophrenia. In later life, however, visual deprivation or visual loss can give rise to hallucinosis, disorders of visual insight such as blindsight or Anton syndrome, or, in the context of neurodegenerative disorders, more complex psychotic presentations. We draw on a computational psychiatric approach to consider the foundational role of vision in the construction of representations of the world and the effects of visual loss at different developmental stages. Using a Bayesian prediction error minimization model, we describe how congenital visual loss may be protective against the development of the kind of computational deficits postulated to underlie schizophrenia, by increasing the precision (and consequent stability) of higher-level (including supramodal) priors, focusing on visual loss-induced changes in NMDA receptor structure and function as a possible mechanistic substrate. In simple terms, we argue that when people cannot see from birth, they rely more heavily on the context they extract from the other senses, and the resulting model of the world is more impervious to the false inferences, made in the face of inevitably noisy perceptual input, that characterize schizophrenia. We show how a Bayesian prediction error minimization framework can also explain the relationship between later visual loss and other psychotic symptoms, as well as the effects of visual deprivation and hallucinogenic drugs, and outline experimentally testable hypotheses generated by this approach.

Introduction: The Complex Relationship Between Visual Loss and Psychosis

Research on vision in psychosis has aimed to characterize the mechanisms of hallucinations¹ or elucidate visual deficits.^2,3 Little research has focused on visual loss and the development of psychosis.

The age of onset of visual loss may be crucial. Strikingly, the absence of vision at birth appears to protect against psychosis, whereas later-life visual loss appears to predispose to the development of psychotic symptoms. In this article, we explore this complex relationship, and the role of vision in constructing models of the world—which become dysfunctional in psychosis.

Very few other medical disorders that may be protective against psychosis have been identified: rheumatoid arthritis is the most notorious example.⁴ Why might congenital blindness be protective?^5,6 Demonstrating a negative association between two disorders with low prevalence is a challenge, but if the risk of schizophrenia and congenital blindness were independent, the probability of not finding any cases of co-occurrence in the literature is exceedingly small.⁷ Congenital peripheral (as opposed to cortical) blindness have been reported to co-occur with schizophrenia, although these reports do not utilize contemporary diagnostic criteria^8,9; furthermore, some reported cases are of early blindness (eg, earlier than 6 years of age) rather than being truly congenital.

More recently, a population-wide study of nearly half a million people found no occurrences of congenital/early cortical blindness and schizophrenia or broadly defined psychosis, and a lower-than-expected rate in individuals with congenital/early peripheral blindness.¹⁰ On the other hand, lower visual acuity in adolescence is associated with increased psychosis risk.¹¹

The lack of association is particularly surprising given that congenital blindness often results from perinatal infections or trauma, or chromosomal disorders,¹² all of which are independently associated with psychosis.^13–15 Congenital rubella syndrome, eg, predisposes to both schizophrenia and congenital blindness but there are no documented cases of their co-occurrence.¹⁶

The relationships between visual loss and the development of psychosis might be explained through the lens of predictive coding, a potentially unifying framework for behavioral and physiological data in both schizophrenia and blindness.

The Bayesian Brain

Predictive coding accounts of psychosis take as their starting point the assumption that the brain is a hierarchical Bayesian inference machine. In Bayesian inference, prior predictions about the world are represented as probability distributions of the causes of inputs lower in the hierarchy. These priors are then combined with data to form a posterior probability distribution. How much weight the prior has relative to the data is determined by the inverse variance of the prior probability distribution, or precision. The relative precision of priors and sensory likelihood (ie, the precision of prior and sensory prediction errors) determine whether sensory data are discounted. If sensory precision is high relative to prior precision, then sensory prediction errors invoke a much greater belief updating. Conversely, if the precision of prior prediction errors is greater than sensory precision, sensory evidence is effectively ignored. Noisy parties may yield a relatable example. If you are at a crowded party with someone you have just met, you may find it harder to understand what they say, because your model of their speech will not be as precise as it would be say for a beloved partner.¹⁷ When someone is well known, the precision of your prior beliefs about them is higher and you can more readily attribute prediction errors to the background and the chances of misunderstanding are lowered.¹⁷

Predictive Coding Approaches to the Symptoms of Schizophrenia

Recent accounts of psychotic symptoms in schizophrenia^18,19 have adopted a broadly Bayesian approach, although details of the precise deficit tend to differ between accounts (see refs. ²⁰ and ²¹ for reviews). All theories focus on a neuromodulatory deficit such that the variables in the hierarchy are inappropriately optimized. Aberrant precision of prediction errors causes a state where previously irrelevant stimuli become abnormally salient in terms of their ability to update beliefs higher in the hierarchy. In this setting, prior precision may increase to compensate for the over-weighting of sensory evidence, culminating in delusions and hallucinations. Alternatively, a primary abnormality of prior precision may underwrite false inference about the world or self.^22,23

The Primacy of Vision in Inferring the Causes of Sense-Data and Shaping Our Representation of the World

Of all sensory modalities, vision affords the perceiver the most amount of information about the world; eg, vision can convey more information (eg, spatial location, size, motion, color, and number) about someone approaching from a distance than touch, hearing, or smell.

Visual inputs have primacy in sculpting priors. There need only be a tiny change in the wavelength of light reaching my retina for me to change my mind about whether the ship on the horizon is flying pirates’ colors. A visual prior will usually affect the perception of an object in another modality (eg, the rubber-hand illusion). For a fully sighted individual, it is rare for what someone hears to trump what they see. As I walk down a busy shopping street, my priors are updated with every saccade. However, if I were blindfolded I would be immediately overwhelmed by a wealth of confusing auditory and tactile data. Vision confers a consistency²⁴ and a context to integrate data from other modalities.²⁵ Notably, there are cases of psychotic illness including schizophrenia following congenital cortical deafness,²⁶ consistent with the primacy of vision, as opposed to hearing, in organizing our multi- and supramodal world models.

Sensory Impairment As a Window Onto Psychosis

It is possible then to account for why adult visual loss can result in visual hallucinations, following visual deprivation²⁷ or in Charles-Bonnet syndrome.²⁸ If the brain has developed normally, then visual loss will decrease the precision of visual input. Higher-level predictions will thus “explain away” this noisy input, resulting in false inferences, ie, hallucinations.²⁹

In neurodegenerative disease such as Parkinson’s disease and Alzheimer’s, visual impairment is also a risk factor for the development of visual hallucinations^30,31 and indeed visual hallucinations in such disorders occur more frequently in conditions of poor ambient light.^31,32 Interestingly, the occurrence of complex (as opposed to simple) visual hallucinations in neurodegenerative disorders is relatively greater than in hallucinosis secondary to visual loss.²⁸ In the latter, the precision of higher-level priors is relatively unimpaired and so noisy sensory input mainly affects lower levels, creating shapes, flashes, etc. In neurodegenerative disease, complex hallucinations such as people or animals may be more frequent because of the combination of imprecise sensory data and abnormal modulation of the precision of priors higher in the hierarchy, secondary to abnormalities in modulatory neurotransmitters such as dopamine and acetylcholine; and, given that brain structure recapitulates model structure,³³ secondary also to structural brain changes with neurodegeneration, ie, the model itself is degenerating.

When healthy adults are blindfolded, a majority of subjects report visual hallucinations, after a day.²⁷ Full insight—awareness of the nonveridical nature of the experience—into these experiences is maintained. Likewise, in organic visual hallucinosis, insight is usually retained. Substantial imbalances between the precision of high-level prior beliefs and of sensory data may contribute to loss of insight such that an extraordinary percept can no longer be dismissed with confidence as unlikely. In Parkinson’s disease, worsening insight regarding visual hallucinations accompanies worsening neuromodulatory dysfunction³⁴ and disease progression into cortical areas and circuits that have been implicated in the specification of higher-level priors.^31,33

When individuals deprived of tactile or visual stimulation are played meaningful auditory stimuli (such as jokes), there is a resultant decrease in psychopathology compared with those played “white noise,” ³⁵ ie, there is a top-down attenuation of the effects of noisy, bottom-up signals,³⁶ which may explain the efficacy of keeping elderly patients occupied or talking to them in managing their distressing visual hallucinations.

The visual hallucinogenic effects of dopamine agonists and anticholinergics can be similarly understood as affecting the relative precision of prior beliefs, as can the psychotomimetic effects of various drugs of abuse.³⁶ Psychedelic drugs like lysergic acid diethylamide (LSD), which act via serotonergic 5HT2A receptors, are potent inducers of visual hallucinations.³⁷ Blind subjects given LSD all experienced hallucinations in multiple modalities, although congenitally blind subjects did not report visual hallucinations. Of 13 late-blind subjects who experienced visual hallucinations, only two experienced complex hallucinations.³⁸ An in-depth account of LSD reported by a congenitally blind musician similarly reveals an absence of visual hallucinations but an abundance of experiences involving other senses.³⁹

Visual loss in later life is not protective against the development of schizophreniform psychosis. Indeed, in Usher syndrome, visual degeneration occurs after adolescence (although sometimes as early as in the first decade) and is associated with schizophrenia-like symptoms.^40–42

The Challenge Faced by the Developmentally Visually Deprived Brain

For a congenitally blind person, there is no rich visual signal with which to shape one’s priors about the world. Each of the other sensory modalities samples a much smaller part of the sensorium, in a noisier fashion, and priors must be built up, piecemeal, from the information contained therein. It is essential then that these hard-won priors (both supramodal and within individual modalities⁴³) remain stable, so as to enable effective interaction within the world. That is, the organism should exhibit a relatively greater top-down influence of priors because the bottom-up, sensory information samples much less of it. As a congenitally blind individual walks down that same shopping street described above, his situation must be very different from a sighted individual with his eyes closed. The same auditory and tactile information that a blindfolded person perceives as chaotic and confusing does not overwhelm a congenitally blind person.

Congenitally blind individuals show reduced integration between nonvisual modalities.^44,45 For example, sighted individuals are vulnerable to an auditory-tactile illusion, whereby multiple tones presented simultaneously with a single tactile stimulus leads to the perception of more than one touch. Congenitally blind individuals evince a significantly attenuated illusion.⁴⁵ Putzar et al have shown that in adult patients born with congenital cataracts who at least 5 months later had them removed, thereby restoring sight, there was evidence of significantly impaired audio–visual integration.⁴⁶ This is also consistent with the possibility that early visual deprivation permanently reduces multimodal integration. On a single-cell level, Carriere and colleagues found that visual deprivation altered the response properties of single neurons in the cat anterior ectosylvian sulcus, a cortical area implicated in higher-order multisensory processing: dark rearing caused a shift in the neuronal population away from neurons whose responses could be effectively driven by stimuli in a number of different sensory modalities towards neurons whose responses were primarily driven only by unisensory stimuli and which could now only be modulated by a simultaneously presented stimulus in a second modality.⁴⁷

The reason this occurs may be obvious: vision most clearly provides the spatial scene within which sensory data from other modalities can be most efficiently contextualized and integrated; it enables the construction of multimodal or supramodal representations. According to Hotting et al, “developmental visual input is essential for the use of space to integrate input of the non-visual modalities, possibly because of its high spatial resolution.” ⁴⁵ In the visually deprived brain evidence of impaired online multisensory integration suggests a less efficient development of supramodal and multimodal (ie, higher level) representations, despite evidence of substantial functional and anatomical cortical reorganization.^48,49 There is a requirement, therefore, for these hard-won higher-level representations to exhibit a stability that is not threatened by the individual’s “noisy” nonvisual sense data. The “imposition of structure on noise,” through inference to the best explanation or abduction, must be relatively greater.

Congenitally blind individuals are less susceptible to the somatic version of the rubber hand illusion wherein, in normally sighted but blindfolded individuals, proprioceptive and tactile information are integrated to create a false sense of bodily ownership.⁵⁰ In congenitally blind individuals, despite identical tactile and proprioceptive inputs, no illusion was experienced at all, suggesting a unique stability of their supramodal higher-level bodily representations in the face of “surprising” sensory data. There is evidence that early blindness leads to improved spatial cognition for tasks that use an egocentric reference frame,^51–53 indirectly supporting the idea that such visual deprivation necessitates the construction of a more stable representation of the world as it relates to ones interactions within it.

It may seem obvious that such an internal world would be more resilient to the kind of reality distortions that characterize the positive symptoms of schizophrenia. But how might such an invariant representation be achieved, computationally, and how might this be protective against schizophrenia?

A Computational Solution to the Challenge of Visual Deprivation: Evidence for Increased Top-Down/Neuromodulatory Drive in Congenitally Blind Individuals

In considering the neuronal instantiation of predictive inference in the brain, it is useful to distinguish between driving and modulatory signals.⁵⁴ Driving signals convey information about the presence or magnitude of prediction error and are typically thought to be mediated by strong intrinsic forward connections, relying on fast AMPA receptor-mediated glutamatergic currents; driving connections elicit a spiking response in their targets. Modulatory signals serve to modify response properties of their targets; they are thought to be mediated by slow, backward connections; they elicit small postsynaptic responses that grow larger with repeat stimuli and show nonlinear response characteristics.⁵⁵ They are implicated in fine-tuning the context-sensitivity of neuronal responses, eg, in the formation of receptive field characteristics. NMDA receptors (NMDARs) exhibit a number of properties that strongly suggest that NMDAR-mediated signaling is primarily top-down and modulatory in character. Other neurotransmitters including dopamine, GABA⁵⁶ and acetylcholine also perform modulatory roles. Dopamine might specify the precision of prediction errors⁵⁷ in the sensorimotor and interoceptive domain, namely, those involved in planning actions.⁵⁸

An increase in top-down modulatory signaling would be one way to ensure the stability of higher-level priors in the visually impaired brain. Carriere et al demonstrated that on a single-neuron level, visual deprivation shifts the responses of higher-level, multisensory neurons towards a profile indicative of an increased modulatory influence and a decreased sensory driving response.⁴⁷ This modulatory change may be NMDAR-dependent. A Transcranial Magnetic Stimulation (TMS) study has demonstrated that visual deprivation does indeed cause increases in NMDAR-dependent cortical excitability in humans.⁵⁹

This shift is explicable if the influence of prior beliefs is a result of their precision weighting relative to prediction error. In the visually deprived brain, we would expect increased precision of priors.

There is evidence that early visual impairment induces such a state. The neuromodulatory influence of NMDARs may be determined by their subunit composition. The NR2B subunit has the slowest kinetics for the release of its Mg+ ion such that those NMDARs containing the NR2B subunit are the most nonlinear and the most effective summators of EPSPs. NR2B-containing receptors are densest in layers 2 and 6 of the macaque visual cortex; these layers receive the densest termination of backward projections, consistent with the computational requirements for “top-down” signaling, ie, descending nonlinear predictions capable of negating ascending prediction errors. Thus, NR2B subunit-containing receptors may have a particularly important role in the modulation and specification of priors.⁵⁵

In the early-developing brain, the ratio of NR2A-containing to NR2B-containing NMDARs (the NR2A/NR2B ratio) is at its lowest. Postnatally there is an excess of NR2B receptors and, during development the amount of NR2A increases, increasing the NR2A/NR2B ratio in several brain regions. This coincides with a period of increased exploration of the world in most species and may mediate synaptic remodeling,⁶⁰ or regulate the threshold for long-term potentiation/depression (LTP/LTD)—known as metaplasticity.⁶¹

Dark-rearing rodents retards this shift from NR2B to NR2A in visual cortex such that higher levels of NR2B persist for longer.⁶¹ Computationally, this could be thought of as a persistence of top-down modulatory influence beyond the normal, developmentally expected period (perhaps via more efficient induction of LTP/LTD). That visual deprivation leads to an increase in the modulatory, rather than the driving, signal we hypothesize is an adaptive response to early visual loss to maintain stability of higher-level priors.

It is likely that there are many such adaptations, and some of the behavioral and neurophysiological differences observed in congenitally blind people,⁵ may subserve the same function. There are no data on whether dark-rearing alters the NMDAR subunit composition in areas outside of visual cortex; this a question of great interest for future studies. Furthermore, it is known that across species, early visual deprivation leads to plastic changes in cerebral cortex whereby visual cortex comes to instantiate sensory processing for other modalities.^62,63 If the relative changes in NMDAR subunit composition (along with other relevant changes in receptor function) are retained in brain areas which go on to serve nonvisual processing, then it is likely that the resulting computational changes (ie, greater influence of top-down modulation) would also affect processing in other modalities, even supramodally. This is consistent with the idea that, fundamentally, the cortex is “metamodal”: rather than being specialized for a particular kind of sensory input it is specialized for a particular kind of computation.⁶⁴

How Might Computational Changes Occurring in the Visually Deprived Brain Protect Against Schizophrenia?

While the dopamine hypothesis has hegemony, there is increasing recognition that NMDAR dysfunction may be primary to dopaminergic dysfunction,⁶⁵ with converging evidence for NMDAR antagonism and/or hypofunction in generating symptoms of psychosis.^18,66,67

There is increasing evidence for schizophrenia-associated alterations in NMDAR subunit composition,⁶⁸ consistent with NMDAR hypofunction. Mutations in the GRIN2B gene, which codes for the NR2B subunit, suggest a reduction in the number or function of NR2B-containing receptors.⁶⁹ Interestingly, administration of NMDAR antagonists, which are psychotomimetic, causes increase of synaptic NR2A-containing, but not NR2B-containing, synaptic receptors.⁷⁰ However, other alterations in glutamatergic signaling, or other neuromodulators should not be ignored.

Predictive coding theories of schizophrenia state that abnormal precision of prediction errors or priors gives rise to false inferences about one’s own thoughts, movements, and even emotions manifest as hallucinations and delusions. We suggest that increased precision of higher-level priors in the congenitally visually deprived brain protects against schizophrenia. That is, abnormalities that characterize the disorder have less impact on congenitally blind individuals because of the stability of their higher-level supramodal representations.

Patients with schizophrenia have highly variable estimates of the visual consequences of their actions. This variability correlates with the strength of delusions of control. Furthermore, they rely more on external visual information than controls in making their judgments. They may have imprecise internal predictions about the sensory consequences of action, which prompts greater reliance on external cues.⁷¹ We suggest that in visually impaired individuals this situation is reversed: relatively greater precision of internal predictions, as a consequence of the impossibility of visual calibration. This is fundamentally opposed to the low precision internal predictions posited in schizophrenia.

Altered Conscious Vision, Reality, and Psychosis

Our focus in this paper has been on the role of precision as encoding uncertainty in hierarchical predictive coding—and how this is affected by early visual experience. The explanatory scope of this formulation is appealing because it links neurodevelopment with Bayesian belief updating and the opportunity for false inference—of the sort associated with positive psychotic symptoms. A key aspect of this formulation is that it rests upon the top-down control of precision at various levels of the cortical hierarchy. In turn, this means the brain must be equipped with predictions or beliefs about precision, namely, beliefs about beliefs. This is important because it speaks to a form of metacognition, namely, beliefs about the precision of beliefs lower in the hierarchy, or “hyperpriors.” One example is “large amplitude prediction errors can only be generated by things I can’t predict” (such as external agents); such a hyperprior may give rise to erroneous external attributions of the causes of sense data in psychosis.²² Thus, hyperpriors in the hierarchies of developmentally typical individuals predispose to visual hallucinations following visual deprivation or later-life visual loss. With congenital visual loss, however, by virtue of developing within an inherently less predictable world, different hyperpriors obtain that are less likely to garner external attributions and psychosis. (There is resonance here with the autism literature also, in which aberrant precision of prediction errors likewise perturbs difficult social inferences.⁷²) On this view, the role of hyperpriors (about precision) may also inform the extent to which people have an insight into their false percepts.

In the neuropsychology of vision, there are patients whose brain damage and subsequent dysfunction help us think about vision and consciousness. Blindsight describes the covert visual abilities of brain damaged individuals who deny conscious visual perception. It is contrasted with Anton’s syndrome, wherein blind individuals claim to be sighted and behave as though they are (walking through, but colliding with, objects in the world). Both of these syndromes are characterized by acquired damage to the visual cortex in individuals whose development (and interaction with the visual world) was otherwise typical. The syndromes are consistent with a Bayesian Prediction Error Minimization Model of conscious perception, wherein candidate percepts are predicted and to some extent visual experience is consistent with those model predictions. The syndromes differ in the richness with which those predictions are experienced. Blindsight has low richness which conflicts with behavioral detection of stimuli. Anton’s has high richness that conflicts with reality.

We suggest that in blindsight, model predictions and the extent to which perception conforms to predictions are impaired. And in Anton’s syndrome, prediction error minimization is impaired such that individuals fail to infer that they are indeed blind. In Charles-Bonnet syndrome, model making and perception aligning with ones’ model is hyper-engaged, however, and prediction error processing, particularly at higher hierarchical levels (which we may term reality monitoring) is somewhat intact, since Charles-Bonnet syndrome cases often appreciate that they are hallucinating.

Psychosis and Schizophrenia

Our discussion has focused on positive psychotic symptoms of schizophrenia. Negative symptoms and thought disorder are also central features of the illness. Whilst they have received less consideration from predictive coding theorists, they can be brought into the explanatory fold.⁷³ In brief, thought disorder would arise when the contextual predictions that constrain cognition are imprecise (subtending aberrant prediction errors that derail one’s train of thought) and negative symptoms may arise from maladaptive predictions about the consequences of one’s actions: if we experience our agency unpredictably, why act at all.⁷³ There is a dearth of research pertaining to the precision of the relevant predictions in congenitally blind individuals, but we would predict that they would show similarly increased stability in the relevant domains.

Conclusion and Predictions

We have argued that visual experience is critical in the construction of our internal world model. Visual loss can disrupt the development or maintenance of that model. We have proposed a predictive coding solution to the conundrum that congenital visual loss protects against psychosis, while later-life visual loss predisposes towards it. We argue that congenitally blind individuals exhibit greater stability of higher-level priors, possibly via increased NMDAR-mediated signaling. We believe that the functional or computational core of this theory offers an attractive and testable hypothesis, leading to a number of experimentally testable predictions:

1. Congenitally blind people will show decreased psychotomimetic effects of ketamine.

2. Congenitally blind people will show a more pronounced force matching illusion in comparison with blindfolded controls; this is in contrast with schizophrenic patients who show a reduced illusion.⁷⁴

3. Congenitally blind individuals will show lower psychosis-proneness than the sighted population. If congenitally blind individuals exhibit markedly stable prediction error signaling, they should show reduced schizotypy.⁷⁵ A similar prediction has been made by Silverstein and colleagues.⁷

4. Congenitally blind individuals will show appropriate frontal cortical fMRI prediction error responses during causal learning, unlike people with psychosis whose prediction error is aberrant (per Corlett et al⁷⁶).

Our biological knowledge of the computational changes outlined above will change as our basic understanding of both blindness and schizophrenia evolve. NMDAR-mediated signaling is unlikely to represent the full picture. Nevertheless, we hope that the preceding discussion illustrates the potential for computational psychiatry to shed light on one of psychiatry’s most recalcitrant mysteries.

Funding

P.R.C. was supported by the Yale University Department of Psychiatry the Connecticut Mental Health Center (CMHC) and Connecticut State Department of Mental Health and Addiction Services (DMHAS), an IMHRO/Janssen Rising Star Translational Research Award, and NIMH R01MH12887).

Conflict of Interest Statement

The authors declare that they have no conflicts of interest.