Reducing Hallucinations in Clinical Summaries
A novel attention-routing mechanism down-weights low-confidence retrievals in long-form clinical summarization. Fact-conflicts dropped sharply.
- Fact conflicts
- -40%
- Eval coverage
- 12 specialties
- Duration
- 10 weeks
- Industry
- Healthcare
The problem
Clinicians were spending hours fact-checking generated discharge summaries. Hallucinated medication doses and conflated symptoms were the top complaints.
Our approach
- 1Mapped failure modes against 12 medical specialties.
- 2Designed a confidence-weighted attention router for retrieval steps.
- 3Built a 4,000-case eval set covering edge specialties.
- 4Shipped a shadow deployment alongside the legacy summarizer.
The outcome
Fact-conflict rate fell 40% across the eval set. Clinician review time per discharge dropped by 18 minutes on average. The router shipped to production in week 11.