Evidence-based AI clinical decision support system for acute burn care and complex reconstruction.

Abstract

Burn surgery requires rapid, evidence-based decision-making across acute resuscitation, operative management, and reconstruction. Despite increasing interest in artificial intelligence (AI) for clinical decision support, no burn-specific system currently exists that provides transparent, citation-traceable guidance suitable for time-critical surgical care. We developed BurnRAG, a burn surgery-specific retrieval-augmented generation (RAG) clinical decision support system using a hierarchical RAPTOR architecture and a curated corpus of 4683 peer-reviewed burn surgery articles (2000-2025). Ten expert-designed clinical scenarios spanning acute management, operative decision-making, reconstruction, rehabilitation, and mass casualty triage were evaluated. Performance was characterized using semantic similarity (SEM-eval), maximum and mean similarity scores, a holistic G Score assessing response completeness, content coverage analysis, and manual citation verification. BurnRAG demonstrated consistently high evidence-alignment performance across scenarios, with mean semantic similarity of 0.94 ± 0.02 and mean G Score of 0.97 ± 0.06. The system synthesized an average of 6.8 peer-reviewed sources per query, with 100% citation accuracy on manual review. Highest performance was observed in acute surgical decision-making scenarios, including escharotomy planning, fluid resuscitation with inhalation injury, and early excision timing. Reconstructive and rehabilitation scenarios also demonstrated robust multi-source synthesis. BurnRAG is the first burn-specific, citation-traceable AI system designed to synthesize peer-reviewed evidence across the burn care continuum. By providing transparent, source-attributed guidance for emergency triage, operative planning, and reconstruction, BurnRAG establishes a foundation for specialty-focused AI integration in burn surgery.