Machine learning and graph-based connectivity metrics identify language disruption in psychosis: Novel insights from dream reports in an Italian cohort

Language disorganization is a hallmark of psychosis which has traditionally been assessed through clinical interviews and standardized scales. Recent advances in Natural Language Processing (NLP) and graph theory provide innovative, objective methodologies to analyze language production in psychosis. In particular, dream reports, as a unique form of narrative, offer a valuable lens through which to examine the presence of disorganized linguistic features. This study analyzed structural speech graphs of written and oral dream reports from 193 Italian participants (115 individuals with acute psychosis and 78 healthy controls), focusing on key connectivity attributes, such as the Largest Connected Component (LCC), the Largest Strongly Connected Component (LSC) and their z-scores relative to random graph distributions. Patients with psychosis exhibited significantly lower connectivity values than controls (p < 0.0125), with their speech graphs resembling random word sequences more frequently. These results remained significant after controlling for education (p < 0.05), with LCC and LSCz surviving Bonferroni correction (p < 0.0125). A Naïve Bayes classifier using these features achieved 68 % accuracy (AUC = 0.75), demonstrating the potential for automated classification of psychosis. To our knowledge, this is the first study conducted with native Italian speakers, reinforcing the cross-linguistic validity of graph-based approaches. Also, our findings support the utility of graph analysis in detecting psychosis and reinforce the notion that speech abnormalities can be captured from a topological perspective through reductions in speech connectedness, thereby providing a novel framework for understanding thought and language impairments associated with the disorder.