This page was translated automatically from Turkish. Read the original

This page is an AI-assisted Turkish summary of the study presented at the 26th World Congress of Psychiatry. The study belongs to the presenting authors; please refer to the original source when citing it.

EV1236

Clausal complexity and discourse organization in the speech of patients with schizophrenia in Croatian

CLAUSAL COMPLEXITY AND DISCOURSE ORGANIZATION IN THE SPEECH OF PATIENTS WITH SCHIZOPHRENIA IN CROATIAN

Despite preserved syntactic complexity, lexical self-corrections, repetitions, and verbless fragments distinguish patients with schizophrenia from healthy controls with high accuracy.

The aim was to characterize clausal complexity and discourse organization in Croatian-speaking patients with schizophrenia and to evaluate the classification value of discourse features that are typically filtered out in automated natural language processing (NLP) pipelines. The study included 30 patients with schizophrenia and 30 healthy controls, who completed a narrative task and a TAT picture description task as part of the DISCOURSE protocol. Discourse markers, self-corrections (conceptual, syntactic, lexical, morphological, and phonological), and verbless fragments were manually coded; MLT, clausal complexity, and the type-token ratio (TTR) were calculated using CLASSLA Stanza. Logistic regression and random forest models were developed for classification using 10-fold cross-validation and bootstrap resampling (B = 300). MLT and clausal complexity did not differ between groups; patients produced significantly more filler words (p = 0.005), syntactic (p = 0.007) and lexical self-corrections (p < 0.001), repetitions (p < 0.001), and verbless fragments (p < 0.001), and had lower TTR values (p = 0.006). Both classifiers achieved an area under the curve (AUC) of approximately 0.93–0.94 and an accuracy of 0.87–0.90, with lexical self-corrections emerging as the dominant predictor across all measures (OR = 3.18). The inconsistency in TTR's variable importance ranking was interpreted as potentially reflecting multicollinearity. In conclusion, these discourse features, which are routinely excluded from automated NLP pipelines, have diagnostic discriminative value and are important considerations in the preprocessing of speech transcripts.

All WCP 2026 summaries