AI-based decision support system for Addressing Emergency Department return Among Mental Health Patients

About This Grant

AI-based decision support system for Addressing Emergency Department return Among Mental Health Patients Project Summary In 2022, more than 1 in 5 U.S. adults (59.3 million, 23.1% of the adult population) live with mental health (MH) illness. About 1 in 8 Emergency Department (ED) visits involve MH and/or substance use diagnoses. A study indicated MH patients are 4.7 times more likely to be frequent ED users, defined as those with three or more in the previous 3 months, compared to non-MH patients. Each preventable ED return visit contributes to the $32 billion preventable ED costs to the healthcare system annually. While prediction models exist for ED returns among general patient populations or specific groups such as elderly patients, there is a lack of validated models developed to predict ED returns among MH patients. The challenge stems from the complex and unique factors affecting the ED return of MH patients, including clinical, operational, and social determinants of health (SDoH) factors. This complexity makes it challenging for clinicians to anticipate ED return for MH patients using their clinical judgment. Therefore, there is a need for advanced analytical approaches that can process complex patient data and integrate it through real-time clinical decision support (CDS). This project will: 1) Develop a Large Language Model (LLM)-based system for to automatically extract MH ED return risk factors (MERRF) from clinical notes. We will develop a system using state-of-the-art LLM techniques, including Retrieval-Augmented Generation, to accurately extract key risk factors (e.g., SDoH, medication adherence, and poor outpatient follow-up) from unstructured clinical notes. The system’s accuracy will be validated through manual chart reviews by clinical experts and multisite validation. 2) Develop explainable Machine Learning (ML) models to predict ED returns for MH patients. We will develop and validate multisite explainable ML models that integrate structured clinical data with extracted MERRF to identify high risk of ED return among MH patients. We will create an explainability framework that translates ED return risk factors into natural language, making it easy for social workers and providers to understand both the ML-generated ED return risk scores and their contributing factors for each MH patient. 3) Develop and Integrate Artificial Intelligence (AI)- MH ED Return Risk Assessment (MERRA) CDS Tool in the EHR. Using human-centered design (HCD) design principles, we will engage ED stakeholders (e.g., social workers) to develop the proposed AI-MERRA CDS and investigate its integration into ED workflow. Through iterative design and testing, we will optimize the CDS usability and clinical relevance. 4) Evaluate the AI-MERRA CDS Implementation and Feasibility. We will implement a quasi-experimental pilot study at UABHS to evaluate AI-MERRA CDS. We will assess implementation outcomes (e.g., adoption) through quantitative metrics and semi-structured interviews with ED clinicians. Findings will inform a future large scale, multisite implementation of AI-MERRA through a randomized clinical trial. This project is highly feasible and will advance mental health ED care through innovative AI solutions.