AI-Driven vision large language framework for pediatric pain assessment

Area of Science:

• Artificial Intelligence in Medicine
• Pediatric Healthcare
• Clinical Decision Support Systems

Background:

• Accurate pediatric pain assessment is crucial for effective pain management and patient safety.
• Current methods rely on subjective scales and behavioral observations, which can be inconsistent.
• Automated pain assessment tools often lack the reliability of expert clinicians.

Purpose of the Study:

• To establish a proof-of-concept for a novel Pain Assessment Vision-Large Language Framework (PA-VLLF).
• To develop a consistent pediatric pain evaluation system using advanced AI.
• To investigate the potential of vision-language models in clinical pain assessment.

Main Methods:

• Developed and validated the PA-VLLF using keyframes from venipuncture videos.
• Combined ChatGPT-4o and Qwen2-VL architectures within a human-in-the-loop framework.
• Fine-tuned models with customized prompts to generate FLACC scores.

Main Results:

• The PA-VLLF achieved 86.36% accuracy in pain score assessment, aligning with expert consensus.
• Performance was comparable to senior clinicians and superior to other machine-learning approaches.
• A Clinical Pain Assessment (CPA) dataset of 1,248 video segments from 104 children was established.

Conclusions:

• PA-VLLF successfully demonstrates that vision-language models can internalize expert clinical reasoning.
• The framework offers a standardized, AI-assisted assistant to reduce rater variance and cognitive burden.
• PA-VLLF holds significant promise for enhancing pain management strategies through consistent AI-assisted decision-making.