PIs: Wei-Hsuan Jenny Lo-Ciganic, PhD, MS, MSPharm & Ted Park, MD
Funding Source: NIH/NIDA
September 30, 2026 - June 30, 2031
Opioid use disorder (OUD) continues to grow as a public health crisis in the United States, affecting 5.7 million individuals in 2023, more than double the number in 2021. The annual economic burden has reached approximately 1.5 trillion dollars, and opioid-related deaths totaled 54,743 in 2024. While medications for OUD (MOUD) such as buprenorphine are effective, only 25 percent of people with OUD received MOUD in 2022. Treatment is further complicated by the fact that two-thirds of patients with OUD also have a mental health disorder, with more than 65 percent experiencing depression and/or anxiety. More than half of patients discontinue buprenorphine within six months, partly due to inadequate mental health support. As primary care has become the largest and fastest-growing setting for MOUD delivery, primary care providers (PCPs) often feel unequipped to manage the complexity of co-occurring OUD and mental health conditions.
To address this gap, this project will develop PRIME-OUD, an artificial intelligence-assisted clinical decision support tool designed to help PCPs personalize treatment for patients with OUD and co-occurring depression or anxiety. Building on the team's prior experience applying machine learning to electronic health record (EHR) data, including work in sepsis care, the project will use reinforcement learning and large language models to generate personalized treatment plans for MOUD alongside depression and anxiety management. The work is organized around three aims. Aim 1 will develop and validate reinforcement learning-based treatment plans using EHR data from the PaTH Clinical Research Network. Aim 2 will refine these plans through a modified Delphi process with panels of PCPs and psychiatrists, with further validation using data from the OneFlorida+ network. Aim 3 will develop a large language model-assisted tool that translates the reinforcement learning-generated treatment plans into clinically interpretable recommendations, and will prototype a user interface through a user-centered design process.
Together, these aims will produce an AI-assisted, expert-validated tool that supports PCPs in tailoring MOUD treatment for patients with co-occurring mental health conditions. The resulting tool will be positioned for integration into EHR systems and evaluation in future large-scale pragmatic trials, advancing efforts to improve OUD care delivery in primary care settings.