Speech-based recognition and estimating severity of PTSD using machine learning.

Area of Science:

• Computational linguistics
• Psychiatry
• Biomedical engineering

Background:

• Traditional PTSD diagnosis relies on costly interviews and lacks objective measures.
• Biomarkers and machine learning (ML) offer potential for accurate PTSD assessment.
• Speech analysis presents a novel avenue for objective PTSD evaluation.

Purpose of the Study:

• To investigate the efficacy of acoustic features and ML in diagnosing PTSD.
• To develop and validate classification and regression models for PTSD detection and severity estimation.
• To explore the potential of speech biomarkers in differentiating PTSD patients from healthy controls.

Main Methods:

• A dataset of 76 PTSD patients and 60 healthy controls was analyzed.
• Acoustic features were extracted using the openSmile framework.
• Random forest algorithm was used for feature selection, followed by six classification and one regression model.

Main Results:

• Classification models with 18 acoustic features achieved high accuracy (0.975) and AUC (1.0) in binary PTSD prediction.
• The random forest (RF) model demonstrated peak performance.
• The regression model showed significant predictive capability for PTSD severity scores (PCL-5), with notable correlation (R²=0.10, r=0.33).

Conclusions:

• Distinct speech patterns reliably differentiate PTSD patients from controls.
• ML classification models accurately distinguish between these groups.
• While the regression model shows promise for gauging PTSD severity, further validation on larger datasets is recommended.