A Unified XAI-LLM Approach for EndotrachealSuctioning Activity Recognition

LLMs can improve medical training safety: The proposed system achieved 15-20% better accuracy than traditional machine learning approaches in recognizing endotracheal suctioning techniques from video, suggesting AI systems could help standardize training for this high-risk clinical procedure, particularly in settings with limited supervision like home care.

Explainability matters for clinical feedback: The framework prioritizes interpretable AI (XAI) by having the LLM explain its reasoning and provide natural language feedback to trainees rather than just making predictions, addressing a key responsible AI principle that decisions affecting patient safety should be understandable to end users.

Natural language makes AI guidance more accessible: By translating complex technical insights into human-understandable feedback, the system demonstrates how LLMs can democratize access to expert-level guidance for medical trainees who may lack consistent access to experienced supervisors.

Automated anomaly detection enables targeted improvement: The pilot feedback module identifies both correct actions and specific areas for improvement, suggesting that responsible AI systems can be designed not just to evaluate performance but to constructively guide learning and skill development.

Scalability potential for underserved settings: The framework's ability to provide consistent, interpretable feedback at scale could help address training gaps in home care and educational settings where expert supervision is limited, though implementation would require validation in real clinical environments.