Pseudocode Generation from Clinical Protocol Flowchart using Large Vision-Language Models

A key component of developing clinical decision support systems (CDSS) involves translating clinical protocols into executable code linking to the backend Electronic Health Record (EHR) systems - a challenging, time-consuming, and error-prone task, particularly for flowchart-based protocols with complex logic and role-specific actions. This study investigates the potential of large vision-language models (LVLMs) to automate pseudocode generation from clinical protocols as an initial step in this process. Using emergency medical services (EMS) as a case study, we evaluated state-of-the-art LVLMs and finetuned two LVLMs for generating pseudocode that can support the development of CDSS for future EHR database integration. The results indicate that model finetuning leads to a substantial performance improvement of at least 15% in terms of BLEU or CodeBLEU for pseudocode generation and up to 20% in terms of F1 for medical terminology extraction. Although these models show promise, challenges remain in the interpretation of complex protocols with conditional assessments tied to specific roles of care providers. The Pass@1 evaluation results show that human expert evaluation is necessary to assess both the semantic and logical correctness of code generation.