Brain–large language model alignment during naturalistic language processing in attention deficit hyperactivity disorder, depression and schizophrenia

Psychiatric disorders including attention deficit hyperactivity disorder (ADHD), depression and schizophrenia are associated with characteristic alterations in semantic language processing, ranging from attentional dysregulation and reduced contextual coherence in ADHD, and negative cognitive bias and poverty of speech in depression, to formal thought disorder in schizophrenia. These transdiagnostic language alterations likely reflect underlying differences in how the brain processes and predicts semantic information. Large language models (LLMs) offer a computational framework for quantifying semantic representations as high-dimensional contextual embeddings. Recent work demonstrates that functional magnetic resonance imaging (fMRI) activation patterns during naturalistic language comprehension in healthy individuals share representational geometry with internal LLM representations [1, 2, 3]. Whether this brain–LLM alignment is altered in clinical populations remains largely unexplored, and no study to date has directly compared alignment profiles across multiple psychiatric diagnostic groups.

PSYCH-GPT (Psychiatric Symptom Characterisation Using GPT Technology) investigates whether brain–LLM alignment differs between patients with ADHD, depression or schizophrenia and healthy controls. Specifically, we test three hypotheses: (1) neural representational structures in healthy participants correspond to the geometry of contextual embeddings from pre-trained LLMs; (2) brain–LLM alignment is reduced in clinical groups, with disorder-specific divergence patterns; and (3) LLM characteristics such as architecture and training data modulate alignment with respective patient groups.

This prospective, observational, non-interventional parallel-group study recruits 120 participants (n = 30 per group: ADHD, depression, schizophrenia, healthy controls), aged 18–60 years, with native-level German proficiency. Sample size was determined via a priori power analysis (G*Power, one-way ANOVA, f = 0.30, α = 0.05, power = 0.80). Clinical groups require a confirmed diagnosis and stable medication for at least two weeks. The study comprises three assessment time points: (1) online screening including demographics, medical history and MRI safety; (2) clinical assessment using the Symptom Checklist (SCL-90-R), Thought and Language Disorder Scale (TALD), Adult ADHD Self-Report Scale (ASRS), Brief Psychiatric Rating Scale (BPRS) and the DISCOURSE protocol for speech recording; and (3) an fMRI session during which participants passively listen to a naturalistic auditory stimulus. fMRI data are preprocessed using standard pipelines including motion correction, slice-timing correction, normalisation and spatial smoothing. Representational dissimilarity matrices (RDMs) are computed from voxelwise fMRI activation patterns across target words and compared to RDMs derived from LLM contextual embeddings using Representational Similarity Analysis (RSA) [4]. Group comparisons on RSA-derived alignment scores are performed using one-way analysis of variance, supplemented by correlation analyses between alignment scores and clinical symptom severity. Multiple LLM architectures and parameter configurations are compared to evaluate model-specific alignment profiles.

Data collection commenced in 2026 following ethics approval (Ethikkommission, Universitätsklinikum Frankfurt, Ref. 2026-2788) and trial registration (DRKS00039902). First data from ongoing collection will be presented at the congress.

PSYCH-GPT provides a novel framework for translating computational models of language into neurobiological markers of psychiatric disorders. RSA-derived brain–LLM alignment scores may serve as candidate biomarkers for semantic processing alterations, advancing the development of language-based diagnostic tools in clinical psychiatry.

References

1. Caucheteux, C., King, J.-R., 2022. Brains and algorithms partially converge in natural language processing. Commun. Biol. 5, 134.
2. Caucheteux, C., Gramfort, A., King, J.-R., 2022. Deep language algorithms predict semantic comprehension from brain activity. Sci. Rep. 12 (1), 16327.
3. Goldstein, A., Zada, Z., Buchnik, E., Schain, M., Price, A., Aubrey, B., Nastase, S.A., Feder, A., Emanuel, D., Cohen, A., Jansen, A., Gazula, H., Choe, G., Rao, A., Kim, C., Casto, C., Fanda, L., Doyle, W., Friedman, D., Devinsky, O., Flinker, A., Hasson, U., 2022. Shared computational principles for language processing in humans and deep language models. Nat. Neurosci. 25, 369–380.
4. Kriegeskorte, N., Mur, M., Bandettini, P., 2008. Representational similarity analysis – connecting the branches of systems neuroscience. Front. Syst. Neurosci. 2, 4.