AI
MedBeads: A New Data Structure for Grounding AI Models in Complete Patient Records
Researchers propose an immutable clinical record architecture designed to supply generative AI with complete, auditable patient context at inference time.
While large language models can store vast medical knowledge, they struggle to answer patient-specific questions without comprehensive context at the moment of inference. Electronic health records and FHIR standards help organize and share clinical data, yet they fall short of guaranteeing that an AI system receives the full, current, and verifiable information it needs. Current retrieval methods can surface related documents, but they don't ensure clinically connected records are included nor make it clear what information is missing.
What is MedBeads?#
Researchers have introduced MedBeads, a data architecture purpose-built for AI systems working with clinical information. Rather than treating grounding as a search problem, MedBeads frames it as a structural engineering challenge: assembling a complete, bounded snapshot of a patient's longitudinal record before the AI generates a response.
The system stores clinical and knowledge objects as immutable "Beads"—each identified by a SHA-256 hash of its canonical content—within append-only patient records called Pods. These Beads link together into a patient-rooted directed acyclic graph (DAG) with cryptographic properties similar to a blockchain's Merkle structure.
How Clinical Links Work#
Beyond the raw immutable facts, MedBeads separates clinical reasoning into a reconstructable interpretation layer. Typed clinical links connect related observations according to signed, versioned knowledge rules. The key innovation here is that when medical knowledge evolves, these links can be recomputed without altering the underlying clinical facts—preserving the historical record while accommodating new understanding.
When retrieving information for an AI model, the system:
- Follows authorized structural and clinical edges through the graph
- Resolves amendments and retractions as they appear in the record
- Reports when policy or token limits force truncation
- Provides explicit visibility into what is included and what is omitted
Engineering Validation#
The team built an open-source implementation in Go, using append-only storage and reconstructable SQLite projections. They converted 1,135 synthetic patient records from Synthea FHIR bundles, generating approximately one million Beads. The experiment demonstrated that clinical-link derivation and interpretation-layer reconstruction work deterministically and can be reproduced reliably.
