# Generative AI-assisted clinical interviewing of mental health

## TL;DR

AI can conduct accurate and empathetic mental health interviews, matching or outperforming traditional methods in diagnosing disorders.

## Contribution

This study introduces and validates an AI assistant for clinical mental health interviews aligned with DSM-5 criteria.

## Key findings

- AI interviews showed higher agreement, sensitivity, and specificity than established rating scales.
- Participants rated the AI as empathetic, relevant, and supportive.
- AI demonstrated lower co-dependencies between diagnostic categories.

## Abstract

The standard assessment of mental health typically involves clinical interviews conducted by highly trained clinicians. While effective, this approach faces substantial limitations, including high costs, high clinician workload, variability in expertise, and a lack of standardization. Recent progress in large language models (LLMs) offer a promising avenue to address these limitations by simulating clinician-administered interviews through AI-powered systems. However, few studies have rigorously validated such tools. In this study, we used TalkToAlba to develop and evaluat an AI assistant designed to conduct clinical interviews aligned with DSM-5 criteria. Participants (N = 303) included individuals with self-reported clinician-diagnosed mental health disorders, namely, major depressive disorder (MDD), generalized anxiety disorder (GAD), obsessive-compulsive disorder (OCD), post-traumatic stress disorder (PTSD), attention-deficit/hyperactivity disorder (ADD/ADHD), autism spectrum disorder (ASD), eating disorders (ED), substance use disorder (SUD), and bipolar disorder (BD)—alongside healthy controls. The AI assistant conducted diagnostic interviews and assessed the likelihood of each disorder, while another AI system analyzed interview transcripts to verify diagnostic criteria and generate comprehensive justifications for its conclusions. The results showed that the AI-powered clinical interview achieved higher agreement (i.e., Cohen’s Kappa), sensitivity, and specificity in identifying self-reported, clinician-diagnosed disorders compared to established rating scales. It also exhibited significantly lower co-dependencies between diagnostic categories. Additionally, most participants rated the AI-powered interview as highly empathic, relevant, understanding, and supportive. These findings suggest that AI-powered clinical interviews can serve as accurate, standardized, and person-centered tools for assessing common mental disorders. Their scalability, low cost, and positive user experience position them as a valuable complement to traditional diagnostic methods, with potential for widespread application in mental health care delivery.

The online version contains supplementary material available at 10.1038/s41598-025-13429-x.