Large language model-based automation of clinical workflows: A systematic framework for healthcare process optimization

Background: Large language models (LLMs) are poised to revolutionize healthcare processes, alleviate administrative burdens, and improve patient care outcomes. Although the field of natural language processing (NLP) has seen significant progress, there is a lack of a systematic approach to apply LLMs to heterogeneous clinical workflows in the literature. 
Objective: To develop, execute, and conduct an empirical study on a scalable LLM framework for end-to-end clinical workflow automation, including diagnosis support, triage classification, documentation, and discharge planning. 
Methods: A retrospective cohort design was used with 38,700 de-identified Electronic Health Records (EHRs) from three tertiary-care hospitals. The fine-grained version of GPT-3 was coupled with HL7 FHIR R4 interfaces, HL7-compliant NLP pipelines, and a human-in-the-loop (HITL) validation process. Six clinical tasks were evaluated with the metrics: Precision, Recall, F1-Score, and AUC-ROC. 
Results: Proposed framework shows the macro-average F1-Score of 0.886 and AUC-ROC of 0.937. Results of post-implementation analysis showed that clinical documentation time was reduced by 80.7%, triage processing latency was reduced by 75% and 30-day readmission rates were reduced by 39.1%. 
Conclusion: The systematic approach is clinically viable and shows substantial efficiency gains in operation and represents a blueprint for the use of LLM in clinical practice.