User-driven design of a large language model-based solution for antipsychotic medication reviews (Preprint)

BACKGROUND Optimising antipsychotic treatment for schizophrenia requires reviewing past medication histories via antipsychotic medication reviews (AMRs). Because this data is buried in unstructured electronic health records, AMRs are labour-intensive, leaving many patients with suboptimal care. Automating text extraction with large language models (LLMs) could streamline AMRs and accelerate treatment optimisation. OBJECTIVE Conduct a user-driven design study to understand current AMR practice and produce design priorities for an LLM-based solution for AMRs. METHODS A multi-stage design study to inform the development of an LLM-based solution for AMRs was conducted. The design was user-centred, with input from clinicians, who are the intended end-users, and people with lived experience of psychosis. A survey exploring current AMR practice and barriers was disseminated to 20 specialist mental health clinicians. Clinicians participated in a co-design workshop and questionnaire to identify design priorities for the solution. A focus group was held with people with lived experience to obtain input on acceptability and utility of the solution, and potential barriers. Feedback from stakeholder engagement activities was analysed using thematic analysis, with the results used to inform the design of the solution. RESULTS Survey results revealed AMRs take on average 4.5 hours to complete (SD 3.8), and only 34.3% (SD 27.9%) of eligible patients have an AMR completed. Documentation clarity and time-burden were identified as key barriers to AMRs. Focus group feedback highlighted the importance of data security and confidentiality, and the potential concerns that the implementation of artificial intelligence techniques to clinical records could negatively affect people with psychosis. Design priorities from the co-design workshop included ensuring the solution produced a standardised output with transparency on the data used for the output. Engagement tasks were used to design a process flow for the proposed solution which incorporates feedback from the stakeholders. CONCLUSIONS AMRs are time-consuming and not completed for most patients, in part due to the difficulty of exploring electronic health records. An LLM-based solution could optimise AMR processes to ensure more service users are treated optimally.