Email Icon Facebook icon Twitter Icon GooglePlus Icon Contact

User Top Menu

Prediction Error: The Genesis of Delusions?

Led by Hakon Heimer Posted on 8 May 2015

Add your thoughts to a Forum Discussion on the dominant current hypothesis on the formation of delusions in schizophrenia, a discussion stimulated by recent articles in Cognitive Neuropsychiatry.

In this installment of our Forum "journal club" series, we join with Cognitive Neuropsychiatry to stimulate discussion of several recent articles in the journal. The articles under consideration can be accessed by clicking on the links to the left. We invite you to share your ideas and insights, questions, and reactions. So read on!

Read Griffiths et al., 2014

Read Corlett and Fletcher, 2014

Oren Griffiths of the University of New South Wales in Sydney, Australia, and colleagues recently reviewed the hypothesized role of faulty prediction error mechanisms in leading to delusions in psychosis. In particular, they find limitations in the attempts by Philip Corlett of Yale University in New Haven, Connecticut, and Paul Fletcher of the University of Cambridge in the United Kingdom and their colleagues to test this experimentally in humans and to detect changes in neural activity using neuroimaging. Corlett and Fletcher wrote a detailed reply and we felt that these papers could serve as the basis for a useful update and discussion topic.

Delusions and prediction error: re-examining the behavioural evidence for disrupted error signalling in delusion formation. Griffiths O, Langdon R, Le Pelley ME, Coltheart M.
Cogn Neuropsychiatry. 2014;19(5):439-67.            

Delusions and prediction error: clarifying the roles of behavioural and brain responses. Corlett PR, Fletcher PC.
Cogn Neuropsychiatry. 2015; 20(2):95-105.

Cognitive Neuropsychiatry editor Anthony David and the publisher, Taylor and Francis, have kindly made these articles open access to facilitate this discussion, and as seed for the discussion, please read this initial comment by SRF member Guillermo Horga of Columbia University in New York City.


Last comment on 25 Mar 2015 by Guillermo Horga


Submitted by Guillermo Horga on

Predictive Learning and Psychosis: A Commentary on the Review by Griffiths et al. on the Role of Disrupted Prediction-Error Signaling in Delusion Formation


A passerby on the street is talking by herself in a loud, argumentative tone. Both her hands are busy holding grocery bags, with no hint of a cell phone. Nonetheless, as you witness this scene you infer that she must be talking on a hands-free device, a hypothesis that is quickly confirmed when you catch a glimpse of an earpiece. We constantly make use of this type of cognitive exercise, termed perceptual inference, to find suitable explanations for incoming sensory information in our surroundings. Current theories in computational neuroscience posit that we find such explanations by forming internal models of the external world―by forming beliefs about the causes of sensory inputs. These internal models are, in turn, forged through associative learning based on experience (henceforth, the term "learning model" will be used for the current purposes to refer to this class of models). After experiencing a few instances of people talking by themselves and seeing them wearing a hands-free earpiece (i.e., by associating people talking by themselves with the use of a hands-free earpiece), the model that people are talking on a hands-free device given the perception of them talking by themselves may become a satisfactory explanation for this situation. The flipside is that this model of the world becomes a biasing prism through which sensory information in the environment is perceived subjectively. In the example above, what one perceived as a hands-free earpiece from the distance could actually have been an earring, or a hearing aid, but perception was biased toward a visual percept that best fit the a priori explanatory model (i.e., perception was biased toward seeing a hands-free earpiece). Thus, current theories view perceptual inference as a two-way street whereby perception and inference (or formation of beliefs about the causes of sensory events) are part of a single process. Or as Fletcher and Frith put it: perceiving is believing (Fletcher and Frith, 2009). It follows that abnormalities in the neural systems supporting perceptual inference could simultaneously impair perception and belief formation, in extreme cases leading to hallucinatory percepts and delusional beliefs such as those observed in psychotic patients.

Bayesian views of perception and inference are rooted in theories put forth in the 19th century (Helmholtz and Kahl, 1971). Theories of delusional formation inspired by Bayesian inference also have a relatively long tradition (Hemsley and Garety, 1986). However, only relatively recently have modern tools in computational neuroscience and neuroimaging begun to afford more direct ways to test these theories of normal and abnormal brain functioning. Verbal theories of delusions (and psychosis more generally) are giving way to more specific, computational models able to generate specific sets of falsifiable predictions (Maia and Frank, 2011). A growing number of studies in cognitive and clinical neuroscience are capitalizing on these models to analyze behavioral and neural data (including but not limited to fMRI) in order to find neurobiological substrates of model variables. Model-based fMRI in particular goes beyond mapping of task-responsive brain regions to further uncover computational mechanisms that the brain uses to solve a cognitive task (O'Doherty et al., 2007). For instance, learning models typically posit that two main variables or types of computational signals underlie learning: value signals and prediction-error (PE) signals. The former represents the value of a certain stimulus or action in a given situation (i.e., the expected outcome for that stimulus or action), whereas the latter represents the difference between the actual (experienced) outcome and the expected outcome, which acts as a teaching signal that prompts updating of subsequent expectations. A model-based fMRI study of reinforcement learning may thus interrogate regions that track value or PE signals as learning progresses rather than testing which regions respond more strongly to rewarding outcomes compared to non-rewarding outcomes regardless of learning stages (Daw et al., 2006; O'Doherty et al., 2004).

Critical review by Griffiths et al. of the literature

In a recent article published in Cognitive Neuropsychiatry, Griffiths et al. provide a thorough overview of the corpus of empirical work testing learning models of psychosis (Griffiths et al., 2014). More specifically, they review the available evidence for a role of disrupted PE signaling in the pathophysiology of delusions, a model that is heavily based on findings of abnormal dopamine transmission in relation to psychosis (Laruelle et al., 1999) and those linking PE signals to dopamine (Schultz et al., 1997). The authors nicely lay out three points where various extant hypotheses of psychosis converge, namely, 1) that psychosis is associated with abnormal signaling of PEs, 2) that these abnormal PEs increase the salience of experiences they relate to, which 3) ultimately lead to the formation and maintenance of delusional beliefs. They first describe the robust electroencephalographic evidence for a deficit in mismatch negativity (MMN) in schizophrenia (Umbricht and Krljes, 2005), which is consistent with deficient PE signaling, given theoretical work explaining MMN as a type of sensory PE (Wacongne et al., 2012). They also offer a brief mention of model-based fMRI studies on reinforcement learning that support disrupted reward PE signals in schizophrenia. Nonetheless, the authors focus on associative (stimulus-stimulus) learning rather than on sensory or reward-based learning, emphasizing the differences between these types of learning―despite other accounts that precisely emphasize the interdependence across hierarchical levels of learning and the similarities in the computational architecture of different learning processes (Fletcher and Frith, 2009; Friston, 2005; Jardri and Deneve, 2013; Schmack et al., 2013). Specifically, the shift toward associative learning and away from reinforcement learning is justified by the difficulty in distinguishing between a disruption in PE signaling per se and a disruption of the influence of reward on error signaling in reward tasks, although one may argue that this potential confound can indeed be addressed by careful testing of model assumptions (see, e.g., fMRI work by Rutledge et al. suggesting that reward-related signals in ventral striatum meet an axiomatic definition of reward PEs; Rutledge et al., 2010). This rationale thus leads the authors to focus heavily on work by Corlett, Fletcher, and colleagues, which explored abnormal signaling during associative learning in psychosis by using a task devoid of explicit rewards.

In a set of fMRI studies, Corlett et al. used a retrospective revaluation procedure instantiated as an allergist task where the goal was to learn associations between different foods (stimuli) and allergic reactions (outcomes) in fictional patients. In short, the key manipulation is to 1) present a compound of two foods (e.g., apple and orange) followed by an outcome (e.g., allergy) in a first stage, then 2) present one of the foods in the compound alone (e.g., apple alone) followed by an outcome (e.g., allergy) in a second stage, and 3) finally present the other food alone (e.g., orange alone) followed by either of two outcomes (allergy or no allergy) in the test stage. The success of this manipulation rests on the assumption that participants retrospectively infer whether the food presented in the test stage causes allergy or not based on whether the other food in the initial compound caused allergy (i.e., in the previous example participants may learn that the patient is allergic to apples and, via "backward blocking," that oranges will thus not result in an allergic reaction). By varying the outcome in the test stage, this paradigm results in either violation or confirmation of expectations (e.g., when the orange leads to allergy or not, respectively) and thus in PEs or no PEs, respectively. Corlett et al. used this paradigm to show that healthy individuals exhibit signals compatible with PEs in the right prefrontal cortex (rPFC). In a follow-up study, they exploited the same paradigm to demonstrate that delusional patients had reduced fMRI activation in the rPFC under expectancy violations (vs. conditions of confirmed expectancies when facing the expected outcome associated with a food that was presented alone in all three stages and never caused allergy) compared to healthy controls, consistent with weakened PE signals (Corlett et al., 2007). Importantly, weakening of such signals correlated with severity of delusional ideation. Furthermore, Corlett et al. recapitulated this neural phenotype in a pharmacological model of psychosis in healthy controls receiving sub-anesthetic doses of ketamine (Corlett et al., 2006).

Griffiths et al. praised the fMRI studies from Corlett et al. with regard to their sophisticated design and acknowledged the support they provide for a link between disrupted PE signaling and delusions. Nonetheless, Griffiths et al. raised several concerns with these studies, including an incomplete integration of behavioral evidence into the analysis and interpretation of fMRI data, which relied, at least in part, on "reverse inference" (or inferring a cognitive process given a spatial pattern of neural activation; see Poldrack, 2011). They also called into question the fMRI analysis for post-hoc selection of contrasts and use of an inappropriate "low-level" control condition. Particularly important is that the interpretation of the fMRI signals as PEs implies that expectations were formed, based on retrospective revaluation, that were then violated in the test phase. However, Griffiths et al. note that the lack of behavioral evidence for this assumption makes the fMRI data hard to interpret. Further, Griffiths et al. go on to offer alternative explanations that make the reverse assumption of no retrospective revaluation (although it is reasonable to assume that at least some individuals did form expectations based on retrospective revaluation). The review concludes by advocating for simpler associative learning paradigms to investigate the role of deficient PE signaling in psychosis.

Corlett and Fletcher's reply

Corlett and Fletcher's response to the abovementioned issues (Corlett and Fletcher, 2015) argues for the interpretability of neuroimaging findings in the absence of behavioral effects, invoking the possibility that behavioral measures might be less sensitive than neuroimaging measures in some cases. They argue that reverse inference is not inherently flawed, particularly when a cognitive function is inferred from brain activation in the context of a cognitive paradigm in which the same pattern of brain activation was previously related to the cognitive function in question. In sum, while they defend their approach, they acknowledge some of its limitations and suggest that their findings be taken as an emerging narrative that supports a PE model of psychosis rather than definitively demonstrating it.

What next?

Formal models of learning are proving extremely influential in our conceptualization of brain functions in normal conditions as well as their dysfunctions in pathological conditions such as schizophrenia. By providing a concrete framework to understand associative learning and belief formation, the application of these models in combination with modern neuroimaging tools may afford an unprecedented understanding of the cognitive mechanisms underlying psychosis. Work by Corlett et al. in this area capitalized on elegant, well-established psychological phenomena in combination with an fMRI-based cognitive subtraction paradigm to provide initial evidence consistent with a PE model of psychosis, specifically in relation to delusional ideation. Such a learning model of psychosis similarly finds support in electroencephalographic findings linking MMN and responses to altered auditory feedback (both of which can be interpreted as proxies for PE signals) to auditory hallucinations in schizophrenia (Fisher et al., 2012; Heinks-Maldonado et al., 2007). Thus, despite a growing body of literature supporting a disruption of learning signals in psychosis (see Deserno et al., 2013, for a systematic review), the critiques to the work by Corlett et al. reviewed here reflect the fact that conclusive evidence linking abnormal PE signaling and psychosis is still lacking (although in all fairness, these critiques apply similarly to the vast majority of fMRI research into the cognitive mechanisms of psychiatric symptoms and disorders more generally).

A possible way forward, as hinted above, is the use of model-based fMRI to formally demonstrate a direct link between abnormal PE signaling and psychosis. A precondition for this is proof in the sample under examination that a purported PE signal actually conforms to the mathematical definition of a PE (Rutledge et al., 2010) by examining the profile of neural activity under different conditions. Abnormalities in PE signals meeting those criteria can then be tested between psychotic patients and non-psychotic controls, or even preferably within individuals undergoing controlled manipulations that induce transient worsening of psychotic symptoms. Model-based analyses of trial-to-trial changes in neural activity should, in most cases, be preferable to contrast-based approaches, given that PE signals are, by definition, dynamic signals that evolve with each outcome, and the latter implicitly assumes constancy within conditions. Detailed analyses, however, do not require complex tasks. As Griffiths et al. allude to, relatively simple paradigms may suffice and be indeed more appropriate for patients who may suffer from cognitive deficits. Some significant work along these lines has already proven the utility of model-based fMRI in this area (for a review, see Deserno et al. 2013). Model-based fMRI has been used to show weakened PEs during reward-based learning in schizophrenia (Murray et al., 2008), including in never-medicated patients (Schlagenhauf et al., 2014), as well as during Pavlovian conditioning with neutral and aversive visual stimuli (Romaniuk et al., 2010) and during an auditory discrimination paradigm in hallucinating patients with schizophrenia (Horga et al., 2014). Nonetheless, the relationship between weakened (formally ascertained) PE signals and severity of psychosis has been more elusive thus far. Indeed, a counterpoint to the case for model-based analyses is that non-model-based approaches (such as those used by Corlett et al.) have been more successful in finding neural correlates of delusions, which may argue for combining model-based and non-model-based approaches (see, e.g., Romaniuk et al., 2014).

The distinction made by Griffiths et al. in terms of the respective roles of associative (stimulus-stimulus) learning, reward-based learning, and sensory learning in the genesis of psychosis, and specifically with respect to delusional ideation versus hallucinatory percepts and positive versus negative symptoms, should not be taken for granted. Instead, empirical research should aim to elucidate the commonalities and differences among these types of learning and test their potentially distinct roles in the pathophysiology of positive, negative, and cognitive symptoms. Computational models of perceptual inference and learning may thus provide a holistic framework not only for understanding the complexities of psychosis, but also the puzzling syndrome that is schizophrenia.