Teaching a Clinical Multi-Agent System to Say I Don’t Know

This talk demonstrates a clinical AI system that avoids confident errors by abstaining when unsure, separating evidence insufficiency from out-of-distribution cases.

AWS Nova Amazon Neptune SNOMED CT ICD-10 RxNorm

Overview

Serelora is a multi-agent clinical reasoning system that abstains on cases outside its calibrated zone rather than generating confident-sounding answers. The demo walks through three real cases live. In the first, every agent converges and the system commits to a coded diagnosis with a treatment plan and a patient narrative. In the second, agents disagree on the differential and the system surfaces the disagreement to the clinician rather than averaging it away. In the third, the case sits outside the homeostenotic tolerance band of our Clinical Complexity Risk score and the system formally abstains and routes the case to human review. During the talk I will show the orchestration graph, the multi-agent conflict-detection logs, the math of the CCR tolerance band, and the structured chart artifact the system produces when it does decide to commit.

Tech stack

AWS Nova

Amazon Nova is the new family of AWS foundation models: multimodal, production-grade AI offering industry-leading price-performance via Amazon Bedrock.

Nova is Amazon's next-generation suite of foundation models, purpose-built for high-performance, cost-effective generative AI applications on AWS. The lineup includes understanding models—Nova Micro (low latency), Lite, Pro, and Premier—optimized for different tasks and supporting multimodal inputs (text, image, video) with context windows up to 300K tokens. Creative models like Nova Canvas (image generation) and Nova Reel (video generation) provide high-quality, customizable content tools for media and marketing. All Nova models integrate directly with Amazon Bedrock, ensuring native AWS security, scaling, and operational efficiency for builders.

https://nova.amazon.com

View projects
Amazon Neptune

A fully managed graph database service built for high-performance applications that require querying billions of relationships with millisecond latency.

Amazon Neptune is a purpose-built engine designed to handle highly connected datasets without the overhead of maintaining complex relational schemas. It supports the three most popular graph models and query languages: Apache TinkerPop (Gremlin), openCypher, and the W3C’s SPARQL. This flexibility allows developers to build everything from real-time fraud detection systems to sophisticated recommendation engines and knowledge graphs. Because it is fully managed, Neptune handles the heavy lifting of hardware provisioning, software patching, and backups. It also offers high availability with up to 15 read replicas and automated failover, ensuring that enterprise-scale applications remain resilient while processing massive volumes of relationship-based data.

https://aws.amazon.com/neptune/

View projects
SNOMED CT

SNOMED CT is the world's most comprehensive, multilingual clinical terminology database, enabling seamless electronic health record interoperability.

Designed to standardize global healthcare communication, SNOMED CT provides a systematically organized vocabulary of medical terms, codes, synonyms, and definitions. The system contains over 350,000 active concepts (including clinical findings, procedures, and body structures) linked by more than 900,000 semantic relationships to map clinical logic. By translating diverse medical jargon into a single machine-readable language, it allows electronic health record (EHR) systems to exchange, index, and analyze patient data with clinical precision.

https://www.snomed.org

View projects
ICD-10

ICD-10 is the global standard diagnostic classification system used to code and track diseases, symptoms, and external causes of injury.

Developed by the World Health Organization (WHO), ICD-10 (International Classification of Diseases, 10th Revision) serves as the foundational medical coding language for clinical documentation, billing, and global health statistics (such as tracking mortality rates). The core framework contains over 14,000 distinct codes, while localized clinical modifications (like the United States' ICD-10-CM) expand that library to more than 70,000 codes to capture precise diagnostic details. By translating complex medical diagnoses into standardized alphanumeric codes (such as J45 for asthma), ICD-10 enables seamless data exchange, accurate insurance reimbursement, and robust epidemiological tracking across international healthcare systems.

https://icd.who.int/browse10/2019/en

View projects
RxNorm

RxNorm is the National Library of Medicine’s standardized nomenclature and clinical drug vocabulary that enables seamless interoperability across diverse pharmacy systems.

Maintained by the National Library of Medicine (NLM), RxNorm provides normalized, standard names and unique identifiers for clinical drugs and active ingredients. It acts as a universal translator, mapping and linking disparate proprietary vocabularies (such as those from Micromedex, Gold Standard, and Multum) to ensure that electronic health records, e-prescribing tools, and pharmacy management systems speak the same language. By organizing drug concepts by their precise active ingredients, strengths, and dose forms, RxNorm eliminates ambiguity in medication data exchange, reducing clinical errors and streamlining healthcare workflows.

https://www.nlm.nih.gov/research/umls/rxnorm/

View projects

Teaching a Clinical Multi-Agent System to Say I Don’t Know

Related talks