Abstract for: Leveraging large language models to identify acute mental health patient flows in interview data

Models of patient flow dynamics are highly useful for identifying strategic investments in capacity and other interventions to better meet patient needs, particularly for the overburdened behavioral health system, but current approaches to developing, specifying, and updating models are time-consuming. Artificial intelligence (AI) could be used to streamline identification of model-relevant information about patient flows in text data. Using a large language model, we tested prompts to extract data relevant to patient flows in interviews with staff from emergency departments and inpatient behavioral health facilities. These interviews were conducted as part of an effort to guide strategic investments in facility capacity using a simulation model of patient flow among behavioral health facilities in Oregon. AI-identified quotations were compared with manual coding conducted by a trained qualitative analyst. We anticipate that our prompts will successfully enable the AI model to abstract quotations relevant to patient flow, including patient census, routes and rates of flow between facilities, average length of stay (delineated by treatment time and boarding time when possible), and characteristics of patients who experience delays in disposition (e.g., civilly committed). We anticipate that our approach will show a satisfying degree of accuracy compared with manual coding. AI-based approaches have the potential to transform how modelers gather and analyze data, particularly text-based data. We anticipate demonstrating a successful approach for identifying data relevant to patient flow modeling in a project related to inpatient mental health treatment. Future research should further refine and test AI-based approaches.