From thought to language: Comparing schizophrenia spectrum disorders and Wernicke’s aphasia with machine learning and LLMs

Schizophrenia spectrum disorders (SSD) and Wernicke’s aphasia (WA) both disrupt meaningful speech, yet they arise from fundamentally different disturbances in thought and language. SSD is defined by formal thought disorder, in which disorganized thinking is inferred from abnormalities in speech, whereas WA reflects a primary breakdown of language implementation following focal brain damage. We investigated whether quantitative markers of lexical–semantic, syntactic structure, and semantic coherence in spontaneous speech can distinguish SSD, WA, and healthy controls. Using Natural Language Processing techniques, we extracted syntactic, lexical and local semantic similarity features from spontaneous speech transcripts and used them in supervised machine learning models to classify diagnostic groups. In parallel, an instruction-tuned large language model (LLM) was used in a zero-shot setting to assign transcripts to diagnostic categories and to track the severity of language disturbance. Our results showed a distinct linguistic pattern, particularly in syntactic and local semantic organization for WA, indicating a paradigmatic language disorder. By contrast, the same features were less effective in distinguishing SSD from matched controls, in line with the view that SSD reflects a more diffuse disturbance of thought that only partially manifests in surface language. Zero-shot LLM classifications approached the performance of supervised models for WA-related contrasts and were sensitive to graded language disturbance. At the same time, strong task and dataset effects underscored the need for carefully controlled speech elicitation. Together, these findings highlight both the promise and the limitations of automated language analysis for clinical diagnostics and for understanding speech and thought abnormalities.