Artificial Intelligence Meets Ultra-High-Field Neuroimaging to Examine Psychosis

“The limits of my language mean the limits of my world.” – Ludwig Wittgenstein

The emergence of first-episode psychosis (FEP) is a profound, life-altering event marked by elevated mortality, persistent symptoms, and substantial cognitive and functional impairments that reshape life trajectories (1,2). Thus, the study of early psychosis remains one of the most consequential pursuits in psychiatry and should harness evolving capabilities of artificial intelligence (AI) to drive progress. Individuals with long-term disability secondary to FEP endure difficulties in real-world functioning, driven by deficits in attention, verbal learning, reasoning, problem solving, and social cognition, alongside negative symptoms and persistent psychosis. At the core of these illness domains lie impairments in linguistic processing that affect an individual’s recovery, manifesting through clinical features, such as reduced speech productivity, impairments in verbal learning, and abnormal patterns of expressed speech. Important for linguistic processing, context refers to the preceding elements of discourse that shape the processing and interpretation of a subsequent word. Contextual sensitivity denotes the capacity to dynamically adjust meaning based on this surrounding discourse, and its disruption represents a core computational deficit underlying disorganized speech in psychosis.

The search for pathophysiologic links between illness domains and neurobiology has long been a central challenge in psychosis research, with the aim of identifying novel interventions capable of altering illness trajectories. At the center of these efforts is glutamatergic dysfunction. A ubiquitous neurotransmitter important for complex metabolic and synaptic processes involving both neuronal and glial cells, glutamate has been extensively investigated in schizophrenia (3,4). However, findings remain heterogeneous, varying by level of inquiry and illness stage and often eluding consistent mapping onto cognitive and linguistic abnormalities (5). Closing this gap is critical as disorganized speech reflects underlying neurocognitive deficits, constrains contextual processing, and predicts real-world outcomes. Therefore, glutamate abnormalities may represent a mechanistic link between neurobiology and the linguistic disruptions that shape the expression of illness. Meanwhile, traditional clinical assessments remain hampered by subjectivity and insensitivity, missing subtle but consequential disruptions in language that profoundly shape communication and social functioning.

Recent advances in AI and ultra-high-field neuroimaging are now providing unique windows into these mechanisms. In a recent study of individuals with FEP and minimal antipsychotic treatment (n = 39), Wang et al. (6) in Biological Psychiatry: Global Open Science examined the relationships between regional glutamate levels, disorganized speech, and context-sensitive language processing, relative to a matched healthy control cohort. Study participants underwent a naturalistic picture description task that required them to describe pictures from the Thematic Apperception Test. Transcribed speech data were entered into OpenAI’s GPT-3’s large language model (LLM) to capture contextual sensitivity metrics. Noncontextual speech properties were quantified, and thought disorganization measures were extracted via clinical scales, including the Thought and Language Index. These clinical measures were coupled with ultra-high-field 7T magnetic resonance spectroscopy (MRS) from a voxel in the dorsal anterior cingulate cortex (dACC). The authors observed that a lower latent marker of contextual sensitivity from the LLM corresponded with increased thought disorganization, consistent with their findings in a larger study. In addition, lower contextual sensitivity was associated with lower glutamate concentrations in the dACC in participants with FEP. This finding, relating contextual insensitivity and glutamate in FEP, persisted after accounting for age, medication, and a range of other clinical measures, including noncontextual linguistic markers.

Wang et al.’s findings (6) are notable for several reasons. First, they provide direct evidence that glutamate dysfunction in the dACC is linked to a key linguistic feature, contextual insensitivity, such that lower glutamate levels predict a reduced ability to use global, but not local, context and that this computational deficit mediates the relationship between glutamate levels and disorganized speech. This modulatory function likely works in conjunction with GABA (gamma-aminobutyric acid)–dependent mechanisms and may also involve acetylcholine and dopamine, suggesting a broader neural system supporting context processing (6,7). Secondly, they show that AI can objectively capture speech disorganization, with GPT-based measures of lexical predictability detecting subtle disruptions that traditional clinical ratings and standard cognitive testing often miss, providing a more precise window onto the underlying neurobiology. Finally, these deficits are specific to context processing, indicating that disorganization is not a single construct but a separable dimension with distinct neurochemical underpinnings.

By anchoring language disturbances within this framework, the authors show how glutamate, and putatively its interactions with other neurotransmitters, shape the computations critical for organized speech. Given that disorganization is heritable, predictive of psychosis onset, and often resistant to dopamine antagonism, these results highlight glutamate-targeted interventions and computational linguistics as promising tools for understanding and treating this core symptom domain (8).

Overall, the integration of 7T MRS with AI-driven language analysis represents a methodological leap, enabling researchers to interrogate how neural systems shape the organization of thought. However, some limitations and future directions related to this work are worth considering. Applying AI-based methods across illness stages, especially before the onset of psychosis in individuals at high risk for psychosis, will be critical, as it remains unknown whether identifying contextual insensitivity before the onset of psychosis and developing treatments to address this critical linguistic abnormality could alter long-term illness trajectories and optimize outcomes. Although the authors used 7T spectroscopy to optimize their glutamate measures, multivoxel spectroscopic imaging with higher resolution and coverage over other salience network and cognitive control regions will be important (9). Additional multimodal 7T analyses, integrating functional and structural imaging, may provide more robust insights for understanding links with contextual processing. Finally, while the advanced models leveraged in this study represent a methodological step forward, they are constrained. The limits of LLMs reflect the broader limits of investigations, as current AI systems, such as human cognition in FEP, struggle with context, nuance, and the integration of complex information, and continue to evolve in their ability to dissect and model human behavior. In FEP, deficits in contextual processing directly restrict the ability to communicate effectively, mirroring the boundaries encountered in LLMs.

Despite the rapid and widespread utilization of LLMs across psychiatry, it remains unclear how LLMs should be graded for efficacy, what constitutes adequate training of models, and how the overall validity of outputs should be measured (10). Overcoming these and related issues will be crucial as LLMs are implemented in psychiatric care and integrated with research. Current limits of AI models must be recognized, both for interpreting findings and for developing tools that aim to bridge neurobiological mechanisms and real-world functional outcomes.

Looking ahead, longitudinal, multimodal studies integrating MRS, AI-based language metrics, other neuroimaging modalities, and behavioral assessments will be essential for understanding how glutamatergic alterations and language impairments evolve over the course of illness. Such integrative approaches exemplify the potential of precision psychiatry, wherein treatment decisions are informed by both neurobiology and functional outcomes. At the intersection of computational linguistics and ultra-high-resolution neuroimaging, we are gaining novel insight into early psychosis, paving the way for interventions that are timely, targeted, and meaningful for patients, while remaining mindful of the inherent limits imposed by both human cognition and current AI frameworks.