What question did this study set out to answer?

To develop an AI framework that improves the accuracy of mental health diagnosis through conversational assessment.

February 26, 2026Open Access

AI-powered conversational framework for mental health diagnosis

Key Points

To develop an AI framework that improves the accuracy of mental health diagnosis through conversational assessment.
Utilized GPT-3.5 for adaptive, human-like conversations to gather user responses.
Applied a fine-tuned DistilRoBERTa model for multi-class classification from responses.
Employed t-SNE and BERT embeddings for strategic data sampling from a public Reddit dataset.
Focused on achieving high performance metrics, including accuracy and ROC-AUC scores.
Achieved classification accuracy of 96.27% across various mental health conditions.
ROC-AUC scores were consistently above 0.91 for all classes, indicating strong model discriminative ability.
Average inference time for the model was 1.67 milliseconds per sample, allowing for real-time application.

Abstract

In the domain of mental health care, conditions such as anxiety, depression, bipolar disorder, and borderline personality disorder (BPD) affect millions globally, often going undetected due to stigma, limited access to specialists, and the complexity of diagnosis. In this study, a hybrid AI framework is proposed that combines conversational intelligence with deep learning-based classification to assist in mental health screening. The system utilizes GPT-3.5 to conduct adaptive, human-like conversations for gathering user responses, which are then analyzed by a fine-tuned DistilRoBERTa model for accurate multi-class classification. Further, a strategic data sampling technique is employed using t-distributed stochastic neighbor embedding (t-SNE) and Sentence-Bidirectional Encoder Representations from Transformers (BERT) embeddings to select the most representative samples per class from a public Reddit mental health dataset. The classification model achieved high performance, with an accuracy of 96.27%, and Area Under the Receiver Operating Characteristic Curve (ROC-AUC) scores consistently above 0.91 across all classes, indicating strong discriminative capability. The system is computationally efficient, with an average inference time of 1.67 milliseconds per sample, making it suitable for real-time applications. This work offers a lightweight, scalable, and explainable solution that can assist professionals or be integrated into virtual mental health assistants.

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper