Full Program »
151. The Use of Automated Speech Analysis to Identify Veterans with Schizophrenia
Language and speech provide a window into cognitive processes that are relevant to psychiatric illness. Particularly for schizophrenia given its characteristic features of cognitive dysfunction and disordered thought, automated natural language processing methods hold promise for objective diagnostic identification. We analyzed transcripts of biographical narratives from 17 Veterans with schizophrenia (SZ) and 15 Veteran healthy controls (HC) who responded to the same interview prompt. From the de-identified transcripts, automatic analysis extracted word stems, categorized them into six major part-of-speech (POS) categories, counted the relative frequency of each POS, number of word stems, and unique word stems, and measured the dispersion of the words among the corpus. Exploratory t-tests showed that SZ had significantly less “W” speech feature, that is, wh-determiners, wh-pronoun, and wh-adverb than HC. A decision-tree classifier using 10 features of speech obtained 72% classification accuracy in discriminating between SZ and HC. The “W” speech feature can be interpreted as an index for greater syntactic and logical complexity. This feature may be particularly relevant to psychosis as it has been previously found to also be less common in clinical high risk individuals who later convert to psychosis as compared with non-converters. Findings support the utility of automated speech analysis to identify individuals with psychosis. Additional speech parameters will be analyzed and associated with clinical features.