WellbeingAgent: An LLM-Driven Agentic Framework for Personalized Mental Health Support

Abstract

Mental health disorders affect more than half a billion people worldwide, with anxiety and depression among the most common. Yet access to support remains limited, and stigma continues to discourage individuals from seeking care. Recent advances in Generative AI (GenAI) and Large Language Models (LLMs) present opportunities for scalable and accessible mental health assistance. However, major challenges remain in personalization, reliability and hallucinations that undermine trust. This thesis proposes an agentic framework for personalized mental health assistance. The framework integrates fine-tuned LLMs for empathetic conversation and structured function-calling with access to patient-specific multimodal data including data from wearable sensors, which allows the system to provide context-aware and individualized responses. Reliability is improved through customized prompting strategies that reduce hallucinations by grounding responses in prior context and health records. The LLM models are trained for incorporating empathetic conversation and personalized data retrieval, which are based on synthetic datasets, including a benchmark dataset from HuggingFace platform. An agentic framework is adopted to guarantee coordination of different tasks through structured orchestration mechanism. Our main contributions are: (i) adapting open-source LLMs to the mental health domain through instruction fine-tuning, (ii) creating a synthetic dataset for wearable sensor data retrieval and integration, (iii) introducing personalization through fine-tuned LLMs and function-calling, and (iv) developing customized prompting strategies and orchestration mechanism to improve reliability. We designed and developed the proposed agentic framework and obtained consistent improvements across multiple evaluation metrics, with performance gains ranging from 12% to 87% in BLEU and ROUGE scores, and a 20% increase in semantic similarity compared to the base model. The function-calling model further achieved an F1-score of 81.3% that demonstrates reliable structured data retrieval. Overall, the proposed approach demonstrates the feasibility of an end-to-end multimodal mental health assistant that is personalized, context-aware, and reliable.