Prediction of Relapse on OUD Treatment using Machine Learning-Driven Evidence-based Clinical Decision Support Tool (PROTECT)

University Of Florida, Gainesville FL

Investigators

Abstract

Over 5.5 million Americans had opioid use disorder (OUD), and >81,000 opioid overdose deaths occurred in 2023. Buprenorphine, one of the most prescribed medications for OUD reduces opioid use, overdose risk, and saves lives. Yet, 20%-60% of patients relapse to OUD while on buprenorphine. Relapse undermines the benefits of the treatment and increases the risk of overdose-related death and healthcare utilization. To combat the opioid crisis, relapse prevention is crucial; however, practitioners are not equipped with accurate tools for identifying patients at an elevated likelihood of relapse. Currently, they rely on individual risk factors such as low adherence, younger age, presence of psychiatric comorbidity, history of other substance use, and limited social support. However, how these factors are intertwined and influence the overall relapse likelihood is unclear, making it difficult for practitioners to predict the likelihood accurately. Hence, there is a critical need to develop a clinical decision support (CDS) tool to help primary care practitioners (PCPs) identify patients with an elevated likelihood of relapse and implement timely and targeted relapse prevention interventions such as behavioral therapies, mindfulness-based approaches, contingency management, residential rehabilitation, and connection to peer recovery specialists. Machine learning (ML) can uncover hidden patterns in complex data to create precise relapse prediction algorithms and risk stratification subgroups, thereby enhancing clinical care and intervention development. Leveraging our prior work in building ML prediction models in OUD, our goal is to develop an innovative ML algorithm to predict relapse on OUD treatment and build an evidence-based clinical decision support (CDS) e-tool (PROTECT) designed for front line practitioners treating patients with OUD with buprenorphine. We will achieve this objective through the following three specific aims. In Aim 1, we will develop and validate ML algorithms to identify buprenorphine patients with an elevated likelihood of relapse using electronic health records (EHR) data from OneFlorida+ and then externally validate the algorithms using EHR data from the PaTH network. In addition to demographic and clinical conditions, we will include information of patients’ health- related social needs (e.g., homelessness, financial constraint/unemployment), extracted from unstructured clinical notes via natural language processing. In Aim 2, we will prototype the PROTECT CDS tool and identify targeted interventions for relapse prevention. Building on Aim 1’s best-performing model, we will employ a user-centered design process involving iterative user interface development and user-feedback session, collecting feedback from PCPs, addiction psychiatrists, and community partners to assess PROTECT’s usability and optimize linkage to existing relapse prevention interventions. In Aim 3, we will conduct an economic evaluation of the PROTECT tool using a simulation model to demonstrate the potential value of the tool to end-users (health systems and practitioners). Our proposed research is highly innovative and clinically relevant in its use of an ML-based CDS tool to guide clinical practice and tailor evidence-based relapse prevention interventions, thereby optimizing resource allocation and improving health outcomes in patients with OUD.