Predicting Ultra-High Risk Outcomes Using Linguistic and Acoustic Measures From High-Risk Social Challenge...

Area of Science:

• Psychiatry and Mental Health
• Computational Linguistics
• Machine Learning in Healthcare

Background:

• Early detection of ultra-high risk (UHR) for psychosis is crucial for intervention, but current methods lack specificity.
• Speech-based machine learning offers potential for improved prognostic accuracy in UHR individuals.
• Existing speech tasks are often lengthy, limiting scalability; the High-Risk Social Challenge (HiSoC) is a brief, 45-second task assessing social functioning.

Purpose of the Study:

• To investigate if linguistic and acoustic features from the HiSoC task correlate with UHR outcomes.
• To determine the predictive power of these speech features for different UHR trajectories (conversion, remission, maintenance).

Main Methods:

• Collected audio recordings of the HiSoC task from 41 UHR participants.
• Analyzed linguistic features (e.g., words per minute, dysfluency) and acoustic features (e.g., fundamental frequency, intensity).
• Employed multivariate linear regression and linear support vector machines with nested cross-validation for analysis and prediction, evaluated by balanced accuracy.

Main Results:

• The group converting to psychosis showed significantly lower words per minute and higher dysfluency compared to the remitted group.
• No significant differences were observed in articulation rate, sequential coherence, or acoustic measures across outcome groups.
• Prediction models using linguistic features (BA 0.741) and combined linguistic-acoustic features (BA 0.851) outperformed random chance.

Conclusions:

• Linguistic features from the short HiSoC speech task demonstrate significant differences between UHR outcome groups.
• Findings support the feasibility of using HiSoC speech recordings for predicting remission and conversion in UHR individuals.
• The HiSoC task presents a scalable approach for identifying predictive speech signals in UHR populations.