AI in Surgical Decision Support: Promise and Pitfalls

Large language models are revolutionising how clinicians access evidence. But surgical AI tools need careful design to avoid introducing new risks.

What AI Does Well

• Speed — synthesising hundreds of papers in seconds
• Comprehensiveness — searching across multiple databases simultaneously
• Summarisation — distilling long guideline documents into actionable points
• Citation — linking every claim back to its source

The Hallucination Problem

The biggest risk with LLM-based clinical tools is confabulation — generating plausible but incorrect information. This is why any serious clinical AI must:

• Always cite sources — no unsourced claims
• Grade evidence quality — distinguish meta-analyses from case reports
• Flag uncertainty — clearly indicate when evidence is limited
• Never diagnose or prescribe — remain an evidence retrieval tool

Torr Health's Approach

We use a RAG (Retrieval-Augmented Generation) architecture that:

• Retrieves real evidence from verified sources before generating any answer
• Grades every citation on a 5-level evidence strength scale
• Includes a mandatory disclaimer on every answer
• Stores no patient data (avoids SaMD classification)

Regulatory Considerations

Clinical AI tools in the UK must consider MHRA regulation. By positioning as an evidence retrieval and synthesis tool — not a diagnostic or treatment recommendation engine — Torr Health operates as a clinical reference tool rather than a Software as a Medical Device (SaMD).