SCH: Training Mental Health Supporters with Virtual Patients and Automated Feedback

Under-treatment of mental health problems remains a major issue in the US, especially for youth, people of color, and individuals with low incomes. Technology may help reduce disparities in and expand the reach of mental health services. However, the newest technologies, such as generative AI, remain fraught

with perils such as hallucinations. Therefore, rather than using AI to directly interact with clients, we will harness generative AI to provide training to mental health support providers, with the ultimate goal of increasing accessibility of mental health services and improving mental health outcomes for those

receiving care. The goal of this research project is to develop and evaluate an automated, scalable system for delivering experiential training to mental healthcare providers. Specifically, we propose to develop a multi-agent training environment to provide interactive and experiential training on the micro-

skills and underlying common factors for both lay counselors and paraprofessionals. Our training environment consists of three components: Virtual Patient, Assessor Agents, and Trainer Agents. Our proposal has four aims. During Aim 1, we will develop and evaluate a set of LLM-based Virtual Patients (VPs) that (a) realistically depict a wide range of common clinical problems (e.g., depression, job-related

stress, ADHD, and suicidality), (b) engage in coherent conversations with trainees, and (c) present typical counseling challenges, such as addressing resistance to sharing problems in depth. During Aim 2, we will develop an Assessor Agent capable of automatically assessing the micro-skills used by the trainee, as

well as how the trainee accomplishes higher-level segment goals (common factors). During Aim 3, we will develop a Trainer Agent capable of interacting with trainees, the Assessor Agent module, and the Virtual Patient module to achieve optimal training goals. During Aim 4, we will recruit 7Cups supporters to use

our multi-agent training environment to evaluate its impacts on the training outcomes.