Conditioned hallucinations: historic insights and future directions

Perception is a process of unconscious inference, based on a model of one's surroundings which is combined with incoming sensory evidence. Pavlov suggested that this process involved classical conditioning. The associations formed by learning from experience comprise the world model. We have known since 1895 that such learned expectations may in fact penetrate perceptual inference so profoundly that they induce hallucinations1.

Dating back to J. Konorski in the 1960s, associatively retrieved internal representations have been implicated in the genesis of hallucinations in rats2. For example, a hungry rat is presented with a tone and subsequently a sweet sugar solution. The rat learns after only a few trials that the tone predicts sugar. The tone evokes a highly realistic, sensory representation of the sugar, which the rat has trouble distinguishing from reality. With extended training, rats stop having these cue‐induced hallucinations, but not in animal models that recapitulate the biology of psychosis3.

In humans, consistent pairings between the faint illumination of a bulb and a near‐threshold tone presentation caused subjects to report hearing tones, even when none were presented4. Voice‐hearers with psychosis may be more susceptible to this effect5. Auditory stimuli can also cue expectations: a salient 1‐kHz tone can, through repeated association with a faint visual stimulus, induce visual hallucinations6. These experiences even transfer out of the laboratory: subjects later reported seeing the conditioned visual stimulus on their television screen when none was presented, conditional on hearing a 1‐kHz tone6.

In an adaptation of these classic experiments, we recently recruited four groups of subjects for participation in a functional imaging experiment7. The four groups differed in having or not a diagnosis of a psychotic illness and having or not daily hallucinations, resulting in groups of those with psychosis and hallucinations, with psychosis and without hallucinations, without psychosis and with hallucinations, and without either psychosis or hallucinations. After learning the association between the visual and auditory stimuli, all groups confidently reported hearing tones that had not been presented (conditioned hallucinations). During these, they activated a network of regions previously identified during symptom‐capture of auditory hallucinations (e.g., bilateral anterior insula, association auditory cortex, inferior frontal gyrus, superior temporal gyrus, cerebellar vermis, parahippocampal gyrus, and anterior cingulate). However, those with hallucinations, whether or not they had a diagnosable psychotic illness, reported conditioned hallucinations at a much higher rate than those without.

We next employed a formal computational model of perception: a three‐tiered hierarchical Gaussian filter (HGF)8. The HGF uses participant responses and the task structure to estimate perceptual belief across three levels of abstraction. The first level of the model (X₁) represents whether the subject believes that a tone was present or not on each trial. The second level (X₂) models belief that visual cues predict tones. The third level (X₃) is the change in belief about the contingency between visual and auditory stimuli (i.e., volatility of X₂). Those with hallucinations demonstrated higher degrees of perceptual belief on the first two layers (X₁ and X₂) and an over‐reliance on prior beliefs, which correlated with activity in insula, superior temporal gyrus and other nodes in the network active during conditioned hallucinations. Those with psychosis, regardless of whether they had hallucinations or not, were less likely to detect changes in the statistical structure of the task (X₃) compared to non‐psychotic participants, activating cerebellum and parahippocampal gyrus less while encoding the volatility of the light‐sound contingency.

The model differentiated those with hallucinations from those without, as well as those with psychosis from those without. These computational metrics may hasten the detection of those at risk for hallucinations and psychosis. The computational dissection of the circuit underlying conditioned hallucinations allows for identification of nodes within that circuit that sub‐serve specific computational functions. Results indicate that insula and superior temporal gyrus are particularly involved in encoding low‐level stimulus beliefs, while cerebellar vermis and parahippocampal gyrus are critical for encoding the volatility of learned contingencies.

This dissection has important implications for the use of repetitive transcranial magnetic stimulation (rTMS) and other forms of neuromodulation as potential treatments. Different directions of modulation may be beneficial within each region: hyperactivity within superior temporal gyrus and insula may be ameliorated by slow rTMS, inducing inhibitory plasticity. Decreased cerebellar and parahippocampal activity and belief updating may be remediated by potentiating theta‐burst stimulation. Targeting superior temporal gyrus in this manner has shown efficacy in the treatment of auditory hallucinations9. Likewise, cerebellar vermis, in addition to being driven by multiple sensory modalities, has been implicated in the etiology of schizophrenia and identified as a potential target for deep brain stimulation in treatment of psychosis.

Mathematically, prior weighting is the ratio of the precision of prior knowledge to the precision of incoming sensory information. Therefore, it may potentially be normalized by either decreasing the precision of prior knowledge or increasing the precision of incoming sensory evidence. The precision of sensory evidence appears to depend critically upon cholinergic signaling: acetylcholine increases auditory discrimination abilities and biases perceptual inference toward sensory evidence. Cholinergic receptor blockade diminishes sensory sensitivity, decreases reliance on incoming sensory evidence during perceptual inference, and can both cause spontaneous hallucinations and enhance conditioned hallucinations10. By contrast, increased cholinergic signaling ameliorates psychotic symptoms in humans and rodent models of schizophrenia.

Combining long observed phenomena in humans and animals with state‐of‐the‐art computational neuroimaging, this work has yielded new insights into the biology and psychology of hallucinations, that portend new, more precise, therapeutic approaches.