A machine learning approach to predicting psychosis using semantic density and latent content analysis.

Area of Science:

• Computational linguistics
• Psychiatry
• Machine learning

Background:

• Subtle linguistic features in everyday language may indicate future mental illness.
• Machine learning (ML) provides tools for rapid and accurate extraction of these linguistic indicators.

Purpose of the Study:

• To investigate semantic density and speech content as linguistic indicators of psychosis.
• To assess the predictive accuracy of these markers for psychosis conversion.

Main Methods:

• Utilized vector unpacking to calculate semantic density, decomposing sentence meaning into core ideas.
• Extracted latent semantic content by comparing participant speech with social media conversations (Reddit).
• Analyzed data from 40 participants in the North American Prodrome Longitudinal Study.

Main Results:

• Low semantic density and discussion of voices/sounds were identified as key indicators of psychosis conversion.
• Combined linguistic markers achieved 93% accuracy in training and 90% in holdout datasets for prediction.
• Demonstrated the efficacy of ML in analyzing language for early mental disorder detection.

Conclusions:

• Automated language analysis, particularly semantic density and content analysis, can forecast psychosis conversion.
• These findings support a broader project using ML for early detection of various mental disorders.
• Highlights the potential of digital phenotyping through language analysis in psychiatry.