Using speech analysis in virtual agent conversations to differentiate PTSD patients from control participants

Area of Science:

• Psychiatry
• Computational Linguistics
• Speech Technology

Background:

• Diagnosing posttraumatic stress disorder (PTSD) is complex due to symptom overlap with other conditions and patient reluctance to disclose. Traditional diagnostic methods may lack sensitivity and specificity.
• Automated voice analysis offers a potential solution by detecting subtle speech nuances beyond human perception, potentially enhancing PTSD diagnosis.
• Virtual avatars can facilitate semi-structured interviews, creating a comfortable environment for patients to express themselves, which is crucial for data collection.

Purpose of the Study:

• To investigate if speech features extracted from virtual avatar interviews can improve the diagnosis of posttraumatic stress disorder (PTSD).
• To assess the utility of automated voice analysis in identifying linguistic and acoustic markers associated with PTSD.
• To compare the diagnostic performance of speech features against demographic data.

Main Methods:

• Utilized the DAIC-WoZ dataset, comprising 142 dialogues between participants and a virtual avatar.
• Extracted content and acoustic features from interview transcripts and audio recordings.
• Developed classification models using speech features and demographic data to differentiate between PTSD and control participants.

Main Results:

• PTSD participants exhibited significantly more negative sentiment and used less frequent words compared to controls.
• A linear classification model incorporating 21 speech features achieved a balanced accuracy of 0.70, outperforming a demographic-only model (0.53).
• No significant speech feature differences were found between PTSD participants with and without comorbid depression.

Conclusions:

• Automated speech analysis using virtual avatars can identify linguistic markers indicative of PTSD.
• This approach holds potential for aiding PTSD diagnosis and reducing barriers to symptom disclosure.
• Further research is necessary to validate these findings in clinical settings and generalize the results.