Measurement of schizophrenia symptoms through speech analysis from PANSS interview recordings

Speech is considered a clinically meaningful indicator of schizophrenia symptom severity and the quantification of speech measures has the potential to improve the measurement of symptoms. Speech collection for digital phenotyping is often dependent on platforms built using closed-source code and associated with patient and clinician burden. Here, we evaluate recordings of clinical interviews conducted as part of standard clinical trial procedures as reliable sources of patient speech for symptom assessment using digital phenotyping. We hypothesize that speech will be associated with schizophrenia symptom severity as measured by PANSS scores using PANSS interview recordings as a data source, in line with existing research showing these associations using dedicated speech collection platforms and proprietary processing pipelines.

Positive and Negative Syndrome Scale (PANSS) interview recordings, collected during a Phase 2 schizophrenia clinical trial, are used to calculate speech characteristics using open source code. A total of 825 PANSS recordings from 212 participants were used in this study. Mixed effects models accounting for demographic variables and time were conducted to assess the relationship between speech characteristics and PANSS scores.

Our findings show strong relationships between the calculated speech characteristics and schizophrenia symptom severity. Positive symptoms were associated with greater amount of speech, faster speech, and shorter, less varied pauses. By contrast, negative symptoms were associated with decreased amount of speech, slower speech, and longer, more varied pauses.

A large sample of PANSS recordings was successfully processed using open source methods for phenotyping and strong relationships between speech characteristics and symptoms from these recordings were observed. These observations, consistent with existing understandings of speech-based manifestations of schizophrenia, highlight the potential use of patient speech collected passively during clinical interactions for digital phenotyping and symptom assessment. Implications for clinical practice, drug development, and progress towards precision psychiatry are discussed.