Prognosis in a box: Automated analysis of speech predicts transition to psychosis in high-risk patients

BEDI, GILLINDER; CARRILLO, FACUNDO; CECCHI, GUILLERMO; FERNÁNDEZ SLEZAK, DIEGO; SIGMAN, MARIANO; CHERYL CORCORAN

NY

Congreso; Computational Psychiatry; 2013

More than any other medical field, psychiatry relies on self-reported symptoms to ascertain diagnosis and prognosis. The validity of self-report data depends critically on the patient?s motivation and capacity to accurately report their introspective experiences, which may be impacted by psychiatric illness. Analyzing the content of speech using computational methods may present a direct, objective ?window into the mind? to complement existing methods (i.e. self report and clinical observation). To investigate this possibility, we employed computational analysis of speech semantic and syntactic coherence in a sample of patients at Clinical High Risk (CHR) for psychosis. We aimed to assess whether automated speech analyses could distinguish between CHR patients who would go on to transition to psychosis and those who would not. Thirty-five CHR patients underwent an open-ended interview in which they described life experiences and their feelings about their symptoms. Patients were then followed up for 2.5 years or until they transitioned to psychosis. Transcribed interviews were subjected to a novel automated analytic approach, which identified indices of semantic continuity between sentences and syntactic continuity in the text. We employed machine-learning analyses (forced choice with leave-subject-out cross-validation) to assess whether the identified features predicted transition to psychosis over the period of follow-up. For further validation, we assessed whether the classification algorithms could also differentiate between two separate cohorts of schizophrenia patients and healthy controls on the basis of free speech. Five of the 35 CHR patients transitioned to psychosis during follow-up. Automated analysis of speech at baseline predicted the development of psychosis in this high-risk cohort with a sensitivity of 0.8 and specificity of 0.93. The classifiers also accurately discriminated between schizophrenia patients and non-patients in the two independent samples. These findings indicate that computational analyses of speech semantic and syntactic coherence can capture clinically-relevant alterations in mental state, accurately predicting prognosis in a CHR cohort. This proof-of-concept study illustrates the potential for automated speech-based assessment approaches to provide an important complement to existing measures on which psychiatric clinicians base their diagnostic and prognostic decisions.