Syndeora

Hospitals rarely have one clean source of patient data. Admissions and lab results arrive as HL7 v2 messages over raw TCP sockets, discharge summaries as C-CDA XML documents, and each system uses its own codes and patient identifiers. Clinicians end up with partial records, duplicate patients and no reliable view of where a fact came from.

Syndeora had to accept all of those feeds safely, work out which records belong to the same person without ever wrongly merging two people, respect each patient's consent while still allowing emergency access, and present one trustworthy FHIR R4 record to clinician apps and decision support.

The platform is a set of Java 21 / Spring Boot 3 microservices around a HAPI FHIR R4 JPA server on PostgreSQL 16. The interface engine is a Netty server on TCP 2575 that frames MLLP messages, replies with HL7 ACK or NACK matched to the control ID, maps admission, discharge, update, merge and lab-result messages, and sends anything malformed to a RabbitMQ dead-letter queue that operators can replay. The C-CDA importer parses R2.1 documents with XPath into Conditions, Medications, Allergies and Observations, deduplicating with SHA-256 fingerprints of the clinical content.

Every inbound record passes through a terminology service ($translate and $lookup to LOINC, SNOMED CT, ICD-10 and RxNorm, with Jaccard suggestions for unmapped codes) and a Master Patient Index that applies Fellegi-Sunter weights across national ID, phone, birth date and phonetically normalised names. Scores of 70 and above link automatically, 40 to 70 go to a human review queue, and lower scores create a new identity. Inside the FHIR server, custom profiles validate resources, a provenance engine records where each fact came from, a consent interceptor enforces purpose-of-use rules, and an AuditEvent interceptor writes a tamper-evident log, including every break-the-glass override.

Keycloak 24 provides OAuth2 and SMART App Launch 2.0 with patient-compartment scopes. A CDS Hooks v1.0 service returns hyperkalemia and drug-allergy cards, and an AI assistant extracts FHIR resources from dictated notes with confidence scores and evidence quotes. Two React 18 TypeScript apps serve clinicians (timeline, lab trends, provenance, break-the-glass) and integration engineers (live MLLP monitor, message inspector and replay, terminology approval, MPI review). Fourteen containers run under Docker Compose with health probes, Prometheus and Grafana.