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

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Informatics
• Medical Informatics

Background:

• Developing clinical decision support systems (CDSS) requires translating complex clinical protocols into executable code for Electronic Health Record (EHR) integration.
• This process is traditionally time-consuming, error-prone, and challenging, especially for protocols with intricate logic and role-specific actions.

Purpose of the Study:

• To investigate the potential of large vision-language models (LVLMs) to automate pseudocode generation from clinical protocols.
• To evaluate the performance of state-of-the-art LVLMs and finetuned models for this task, using emergency medical services (EMS) protocols as a case study.

Main Methods:

• Utilized state-of-the-art LVLMs for pseudocode generation from clinical protocols.
• Finetuned two LVLMs specifically for generating pseudocode relevant to CDSS development.
• Employed BLEU, CodeBLEU, and F1 scores for evaluating pseudocode generation and medical terminology extraction.

Main Results:

• Finetuning LVLMs resulted in a significant performance improvement of at least 15% in BLEU/CodeBLEU for pseudocode generation.
• Medical terminology extraction accuracy improved by up to 20% (F1 score) after finetuning.
• Challenges persist in interpreting complex protocols with role-specific conditional logic.

Conclusions:

• LVLMs show considerable promise for automating pseudocode generation, streamlining CDSS development.
• Model finetuning substantially enhances performance in both code generation and terminology extraction.
• Human expert evaluation remains essential to ensure the semantic and logical correctness of generated code for clinical applications.