延迟、责任和法学硕士：Delphyr 的工程床边工作流程助手.pdf

1、Engineering Reliability for Clinical EnvironmentsWorld Summit AI 2025Michel Abdel Malek MDFounder/CEO Delphyran agentic platform built for clinical workflowsto reduce time,regain effiency and improve well being of staff and patientsMeet Dr.Tom,PhD Studies 15+years to become a specialist Wastes 40%of

2、 his time,daily Exposure to medical errors Burn-out riskClinical Administrative TasksWith Patient10%Finding data,Generating InsightWriting notes90%We started hereAmbient ListeningRequirements/Constraints Fit in legacy electronic health record systems Fit in legacy data models(or lack thereof)Fit in

3、legacy hospital infrastructure Fit in highly regulated markets Fit in data and patient privacy Robust,predictable,repeatable outcomes Safe and ethically built Environment-friendly etc.How do you build reliable GenAI systems in given the legacy constraints?Robust Retrieval&Reduction of Hallucinations

4、Challenges in working with real world clinical dataLack of proper metadataStructured vs UnstructuredInformation is context dependentContextual terminologyMissing notesHow difficult can it be is it?SOTA Semantic Retrieval Recall 5:55-65%(EHR Benchmark)Breaking down the retrieval pipelineExpansion-Enr

5、ichment-Classification-IndexingChunking-Embeddings-GraphsQuery Understanding-Query Routing-RerankingEvalsDelphyrs Retrieval PerformanceDelphyrs clinical retrieval platform is till date still the best performing retrieval platform for clinical data.(Recall1 95%,Recall5 98%)Did we overengineer it?Prob

6、ably.Is it robust?Yes.Reducing hallucinationsHallucinations in Clinical AIHallucinationshappen hereHallucinations are a fundamental feature/bug of LLMs.Long-tail occurrencesCompression FailuresLack metacognition“Just put the temperature on 0”Delphyr mitigates