DIALOG: Understanding Disorganisation: A Language-focused Global Initiative in Psychosis

Disorganisation is a multidimensional, cross-diagnostic symptom in SMD, but its quality, persistence, and severity varies markedly across diagnoses. Current knowledge on the pathophysiology of disorganisation highlights context-processing deficits, language network dysconnectivity, and aberrant beta oscillations; but this knowledge is highly fragmented - based solely on unreplicated cross-sectional data and focused predominantly on schizophrenia. DIALOG will approach these interconnected processes of disorganisation within a computational framework of predictive processing using naturalistic language modelled using Large Language Models (LLMs) - the key element for social function.

Imprecise predictive processing is a leading neurocomputational theory of disorganisation. Accordingly, the healthy brain implements a generative model of sensory causes (during speech comprehension) and consequences (during production) in the form of probabilistic predictions that propagate down the fronto-temporal cortical hierarchy (messages to words to phonemes). During conversations, our predictions are based on multiple cues (e.g., visual, emotional) and the meaning derived from the full set of preceding words (linguistic context). To be effective, predictions must be precise in their content and timing (i.e., reliable neural representations with low uncertainty). Theoretical models point to the inhibitory modulation of excitatory synapses as the biophysical basis of precision, while temporal dynamics (neural oscillations) appear to modulate this precision. Intact connectivity within the brain's language network (bilateral frontotemporal regions) is key for linguistic context to influence word choice. We propose that the computational failure of predictive language processing arising from synaptic, connectivity and oscillatory dynamics gives raise to disorganisation and social dysfunction.

Large Language Models (LLMs), trained on vast text corpora, generate accurate probabilistic predictions from preceding context (i.e., how chatGPT works). In the brain, predictions can be realised in multiple ways and predictive processing operations per se cannot be measured in-vivo, but the probabilistic effects of context at word level and the uncertainty of the predictions provide neurally valid proxies (we pioneered this in healthy individuals). By applying LLMs to natural speech and transforming probabilistic estimates into information theory-based metrics ('contextual uncertainty' or 'perplexity' and lexical surprisal; hereafter, LLM-metrics), predictive operations in daily social life and determine their neurophysiological basis can be tracked.

DIALOG will characterise the origins (i.e., the synaptic, connectivity, and temporal dynamics behind LLM-metrics) and consequences (persistent disorganisation and social dysfunction) of imprecise predictive processing and identify interventional opportunities through 5 work packages.

Lived-experience experts highlight disorganisation's impact on everyday social function, but the reported correlations are confounded by other symptoms, particularly negative symptoms. Additionally, clinical detection of disorganisation across diagnoses lacks consistency. However, we believe that objective markers of the underlying neurocomputational deficit (imprecise predictive processing) can be trait-like, offering a more consistent causal link to social dysfunction. These markers can be extracted from short segments of natural language use, obtained across time in speech samples.