Language or syndrome? Investigating the origins of the general psychopathology factor through large language model...

Psychiatric diagnostic categories are challenged by comorbidity, symptom heterogeneity, and blurred nosological boundaries, sparking interest in trans-diagnostic dimensions such as the general psychopathology factor (p-factor). Yet it remains unclear whether p-factor represents a valid psychopathological construct or merely echoes semantic overlap among questionnaire items. Leveraging recent advances in large language models (LLMs), we conduct an empirical investigation that aims to disentangle linguistic similarity from syndromic co-occurrence in the p-factor.

We analyzed 145 symptom items spanning eight diagnostic categories with two similarity metrics: (1) Semantic-cosine similarity of LLM embeddings; and (2) Syndromic-polychoric correlations of responses from 985 community participants. Exploratory factor analyses extracted latent structures, their alignment was quantified, and factor scores were tested against lifetime psychiatric diagnoses.

Both similarity matrices converged on a dominant p-factor with high cross-diagnostic loadings. Loading patterns were strongly congruent, indicating that LLM-derived semantics alone recapitulate much of the latent structure in human data. Crucially, syndromic p-factor scores predicted individual diagnostic histories more accurately than semantic p-factor scores, revealing clinically informative variance beyond language.

LLM embeddings reveal that the linguistic fabric of symptom items contains a shadow of the p-factor, yet real-world co-occurrence embeds additional, diagnosis-relevant information. Our findings position LLM-driven text analytics as a new lens for probing psychiatric constructs while underscoring the complex interplay between linguistic semantics and symptom co-occurrence in shaping our understanding of psychiatric constructs.