Quantifying word salad: The structural randomness of verbal reports produced during First Episode Psychosis predicts negative symptoms and Schizophrenia diagnosis 6 months later

Background: The precise quantification of negative symptoms is necessary to improve differential diagnosis and prognosis prediction in Schizophrenia. In chronic psychotic patients, the representation of verbal reports as word graphs provides automated sorting of schizophrenia, bipolar disorder and control groups based on the degree of speech connectedness. Here we aim to use machine learning to verify whether speech connectedness can predict negative symptoms and Schizophrenia diagnosis during first episode psychosis. Methods: PANSS scores and memory reports were collected from 21 patients undergoing first episode psychosis followed for 6 months to establish DSM-IV diagnosis, and 21 healthy controls. Each report was represented as a graph in which words corresponded to nodes, and node temporal succession corresponded to edges. Three connectedness attributes were extracted from each graph, z-scores to random graph distributions were measured, correlated with the PANSS negative subscale, combined into a single Fragmentation Index, and used for predictions. Findings: Random-like speech was prevalent among Schizophrenia patients (64% x 5% in Control group, p=0.0002). Connectedness explained 92% of the PANSS negative subscale variance (p=0.0001). The Fragmentation Index classified low versus high scores of PANSS negative subscale with 93% accuracy (AUC=1), and predicted Schizophrenia diagnosis with 89% accuracy (AUC=0.89). Interpretation: The structural randomness of speech graph connectedness is increased in Schizophrenia. It provides a quantitative measurement of “word salad” as a Fragmentation Index that tightly correlates with negative symptoms and predicts Schizophrenia diagnosis during first episode psychosis.