Conversational assessment using artificial intelligence is as clinically useful as depression scales and preferred by...

Area of Science:

• Mental Health
• Artificial Intelligence
• Digital Health

Background:

• Depression is a widespread, chronic condition often remaining undiagnosed due to limited screening access.
• Artificial intelligence (AI) models analyzing spoken responses offer a potential solution for efficient and accessible depression screening.

Purpose of the Study:

• To validate a previously trained AI model on bot-administered interviews using a diverse sample.
• To assess the AI model for algorithmic biases related to age, sex, race, and ethnicity.

Main Methods:

• Adults (N=393) completed a bot-administered interview and depression self-report via the Aiberry app.
• An AI model predicted depression scores from interview responses.
• Clinicians reviewed discrepancies between AI predictions and self-reports.

Main Results:

• Strong concurrent validity was found between AI predictions and self-report scores (r=0.73).
• AI predictions agreed with self-report or clinical judgment in 90% of cases.
• No differential model performance was observed across demographic groups (age, sex, race, ethnicity).

Conclusions:

• The AI model accurately predicts depression severity from remote spoken responses.
• AI shows promise as a mental health screening tool, performing on par with self-report measures.