Using Laboratory Documentation to Extend a Large Language Model to Serve as a Laboratory Assistant

Abstract Demographic and structural changes in the workforce are resulting in ever smaller cohorts of laboratory medical professionals while medical advancements and an aging population has increased demand for laboratory testing. Furthermore, the inexorable aging of our technologists has resulted in a workforce with a high variation in experience. The need for new tools to help bridge the experience gap and increase productivity are readily apparent. Recently, large language models (LLMs) have emerged as powerful tools in various industries, including healthcare. However, numerous concerns have been raised regarding data security particularly with respect to proprietary and patient data. Furthermore, these models must be adapted to function effectively within the unique context of laboratories. To this end, we present a proof of concept project that uses laboratory documentation to extend a locally run LLM to serve as a laboratory assistant. To develop this tool, we collected and analyzed job aids, SOPs, multidisciplinary guidelines, and classification systems relevant to a particular laboratory at our institution. We then used retrieval augmented generation (RAG) on these documents to locally extend open source large language models to understand laboratory-specific terminology, procedures, and contexts. Preliminary results suggest that the language model can accurately process and generate responses related to specific lab tasks, including sample handling, testing protocols, and data interpretation. For example, it can suggest possible differential diagnoses based on test results and generate suggested follow-up tests even for complex workflows such as leukemia/lymphoma flow cytometry. By using existing specific laboratory documentation as training data, we have demonstrated that it is possible to securely adapt large language models to function effectively within the context of laboratories. This represents an essential step towards integrating advanced technologies into the laboratory. Future work includes evaluating the language model’s performance in different lab settings and expanding its functionality. We envision that this project could evolve into an integral component of laboratories’ digital infrastructure, hopefully enhancing training, knowledge sharing, and operational efficiency.