Poster Finalist for the abstract titled, Improved individualized identification of schizophrenia and clinical high risk for psychosis when combining cognition with natural language processing at the Schizophrenia International Research Society Congress
My abstract investigating the clinical utility of combining linguistic features with cognition to identify patients with schizophrenia and individuals at-risk for psychosis was among the top 5% of posters nominated as a Poster Finalist.
The Computational Psychiatry Course aims to provide scientists and clinicians studying mental disorders with tools to understand, build, analyze, and critically evaluate computational models. Topics range from reinforcement learning to Baesian inference to machine learning.
During the course, I had the opportunity to present my work investigating the clinical utility of combining linguistic features with cognition to identify patients across the psychosis spectrum.
Abstract
Neurocognitive pattern classification has been used to identify patients with schizophrenia (SCZ) and individuals at clinical high risk (CHR) for psychosis with high accuracy. However, few studies have assessed the contribution of linguistic features to identify patients at various stages of psychosis. We investigated the differential diagnostic value of combining linguistic features with cognition to discriminate patients with SCZ and CHR from healthy controls (HC). Automated NLP analysis was applied to speech samples obtained from 105 HC, 42 patients with SCZ, and 122 CHR. A multiclass machine learning analysis pipeline was used to classify SCZ and CHR from HC combining linguistic and cognitive features using a linear Support Vector Machine algorithm. The models distinguishing HC from SCZ and CHR were contrasted with respect to their performance and feature reliability, and applied to the held out patient group to examine their generalizability. The cross-validated classification accuracies were 75.7% (HC vs SCZ), 68.2% (HC vs CHR), and 65.5% (CHR vs SCZ). Linguistic features drove the HC-SCZ classification, while a combination of linguistic and cognitive features drove the HC-CHR model. Applying the HC-SCZ classifier to CHR resulted in 47.5% being classified as SCZ, whereas the HC-CHR model classified 78.6% SCZ patients as CHR. Decision scores of CHR significantly differed from both correctly classified SCZ and HC. These findings suggest combining linguistic features with cognition contributes to the individualized discrimination of patients at various stages of psychosis from healthy individuals, but do not generalize well to the different stages, suggesting different underlying disturbances.