Automatic Verbal Analysis of Interviews with Schizophrenic Patients

Abstract

Schizophrenia is a long-term mental disease associated with language impairments that affect about one percent of the population. Traditional assessment of schizophrenic patients is conducted by trained professionals, which requires tremendous resources of time and effort. This study is part of a larger research objective committed to creating automated platforms to aid clinical diagnosis and understanding of schizophrenia. We have analyzed non-verbal cues and movement signals in our previous work. In this study, we explore the feasibility of using automatic transcriptions of interviews to classify patients and predict the observability of negative symptoms in schizophrenic patients. Interview recordings of 50 schizophrenia patients and 25 age-matched healthy controls were automatically transcribed by a speech recognition toolkit. After which, Natural Language Processing techniques were applied to automatically extract the lexical features and document vectors of transcriptions. Using these features, we applied ensemble machine learning algorithm (by leave-one-out cross-validation) to predict the Negative Symptom Assessment subject ratings of schizophrenic patients, and to classify patients from controls, achieving a maximum accuracy of 78.7%. These results indicate that schizophrenic patients exhibit significant differences in lexical usage compared with healthy controls, and the possibility of using these lexical features in the understanding and diagnosis of schizophrenia.