Personalised Speech-Based PTSD Prediction Using Weighted-Instance Learning.

Area of Science:

• Psychiatry
• Computational linguistics
• Biomedical engineering

Background:

• Post-traumatic stress disorder (PTSD) significantly impacts mental, physical, and social well-being.
• Timely and professional treatment is crucial for individuals with PTSD.
• Speech patterns can reflect psychological states, offering potential diagnostic insights.

Purpose of the Study:

• To develop and evaluate a personalized speech-based approach for predicting PTSD.
• To assess the effectiveness of the personalized model compared to a non-personalized model.
• To analyze prediction accuracy at different stages of clinical intervention.

Main Methods:

• Collected a new dataset of speech recordings from 15 participants (PTSD and healthy controls).
• Included speech data collected before and after clinical intervention.
• Employed a personalized machine learning model for PTSD prediction from speech features.

Main Results:

• The personalized speech-based approach demonstrated superior performance in PTSD prediction.
• Achieved a best Area Under the ROC Curve (AUC) of 82%.
• Showcased a relative improvement of 7 percentage points over the non-personalized model.

Conclusions:

• Personalized speech analysis is a promising tool for PTSD prediction.
• Voice biomarkers can aid in the early detection and monitoring of PTSD.
• Further research with larger datasets is warranted to validate these findings.