EAGER: FDASS: Designing Accountable Mental Health Large Language Model Therapy Software

The rapid development of Artificial Intelligence (AI)-enabled technologies such as Large Language Models (LLMs) has brought significant benefits across various fields, including behavioral health. However, it is important to note that LLMs are typically trained on general-purpose data, do not explicitly adhere to the guidelines and practices followed by behavioral health professionals, and lack in-the-loop feedback, which can result in inaccurate or unhelpful responses. This project addresses the increasing demand for accountable and scalable behavioral health LLM software to help individuals manage personal challenges and enhance their overall well-being. Drawing on expertise in information systems, sociology, and public health, this research team introduces an Accountable LLM software system specifically designed for behavioral health applications based on Cognitive Behavioral Therapy (CBT). The proposed LLM aims to deliver consistent, high-quality guidance and therapy aligned with established practices. The impacts of this project include improved service delivery and the development of accountable designs for future LLM-based approaches in behavioral health applications. The project aims to develop and evaluate an Accountable LLM based on principles from behavioral science and decision theory, specifically for addressing behavioral health conditions such as depression or anxiety. The proposed Accountable LLM employs an innovative fine-tuning strategy that incorporates patient-therapist interaction pairs, supervised instruction fine-tuning, and a preference alignment tuning approach featuring a novel CBT-Kahneman-Tversky Optimization (CBT-KTO) method to ensure the model adheres to prevailing psychotherapeutic practices. The resulting LLM will be integrated into a user interface, enabling sociology and public health scholars to process and produce a wide range of qualitative data that is difficult to capture using conventional methods, such as direct observations. The proposed LLM will be evaluated through program evaluations, surveys, focus groups, and interviews with psychiatrists and end-users. Results from these evaluations will not only be used to improve the LLM iteratively but will also inform relevant health theories, policy development, user engagement, and service outcomes. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.