Multi-Specialist Evidence Retrieval for Clinical Decision Support

The crisis

• A single complex clinical decision can require synthesising evidence scattered across many subspecialty journals and guidelines.
• Today that synthesis depends on individual training, textbooks, or ad-hoc PubMed searches — slow, inconsistent, and unavailable at the point of decision.
• General-purpose LLMs lack domain specificity and hallucinate without grounding in a curated corpus, so a fluent clinical answer can be confidently wrong.
• In a clinical setting an answer is only useful if every claim is linked to a source and the recommendation is auditable — accuracy without provenance is not enough.

About this research

Evidence synthesis for a real clinical decision often has to draw on several subspecialties at once, yet a single-pass search over one corpus tends to flatten that distinction and miss cross-cutting evidence. This thread investigates how an evidence-retrieval system can route a question to the right expertise, draw on multiple sources, assess the quality of what it finds, surface agreement and disagreement, and keep every claim traceable to its source, so that a clinician can verify rather than trust blindly. It is framed as decision support, not an answer engine: accuracy without provenance is not enough. The work draws on agentic retrieval, multi-source evidence synthesis, provenance tracking, and rigorous evaluation. Faculty-advised.

References

• Lewis et al. (2020) · Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (NeurIPS)
• NCBI · Entrez Programming Utilities (E-utilities) for PubMed/PMC

Related project