Using machine learning of computerized vocal expression to measure blunted vocal affect and alogia

Area of Science:

• Psychiatry
• Computational Linguistics
• Digital Health

Background:

• Negative symptoms are a core feature of serious mental illness (SMI).
• Objective vocal analysis offers potential for digital phenotyping of these symptoms.
• Previous vocal analyses showed limited convergence with clinical ratings due to restricted feature sets.

Purpose of the Study:

• To evaluate machine learning (ML) accuracy in modeling blunted vocal affect (BvA)/alogia using extensive vocal features.
• To assess associations between ML-derived BvA/alogia predictions and clinical/demographic factors.
• To identify key vocal features contributing to BvA/Alogia ratings.

Main Methods:

• Utilized a large acoustic feature set from two distinct speaking tasks (picture description, free recall).
• Applied machine learning models to predict clinically rated BvA/alogia.
• Correlated ML-derived scores with demographic data, diagnoses, symptom severity, and functioning.

Main Results:

• High prediction accuracy (>90%) for BvA/alogia, improved when analyzed per task.
• ML scores correlated with impaired cognitive and social functioning.
• Higher ML scores observed in schizophrenia patients compared to depression or mania.
• Identified predictive vocal features were not always aligned with established clinical definitions.

Conclusions:

• Machine learning with comprehensive vocal analysis shows promise for digital phenotyping of negative symptoms in SMI.
• Vocal biomarkers derived from ML correlate with clinical status and functioning.
• Further research is needed to align identified vocal features with clinical constructs for robust implementation.