#1Why IoMT projects stall at the EHR boundary
Vendor demos make EHR↔IoMT integration look like a single API call. Production reality is messier: heterogeneous device protocols, vendor-specific message profiles, network jitter, and consent frameworks that vary by region. Teams either underestimate the surface area and ship brittle point-to-point links, or over-engineer a heavyweight ESB that nobody enjoys operating.
- Devices that batch overnight versus devices that stream every second.
- EHRs that accept v2 HL7 messages over MLLP versus EHRs locked to a proprietary REST profile.
- Identity systems that map a device to a patient through different identifiers per organisation.
- Consent flags that gate which data flows can legally cross which hop.
#2A FHIR-first mindset, not an HL7 retrofit
FHIR is the lingua franca of modern healthcare integration. We treat it as the internal contract between subsystems, even when one endpoint speaks HL7v2 and another speaks a proprietary JSON. The mapping layer translates inbound payloads into canonical FHIR resources at the edge, runs them through validation, then routes them to downstream consumers.
#3Bringing device streams into a clinical-grade pipeline
Continuous device data is not just 'more rows in an Observation table'. We separate the high-frequency telemetry plane (Kafka or MQTT, time-series store, deterministic backpressure) from the clinical events plane (FHIR resources, audit logs, EHR sync). Aggregations move from telemetry to clinical only when they cross a clinically validated threshold.
- Edge gateway buffers and normalises raw device payloads.
- Streaming layer aggregates rolling windows and emits derived events.
- Validator confirms the event is well-formed, references a known patient, and is consented.
- EHR adapter writes a single, semantically meaningful Observation or DiagnosticReport.
#4Safety by design — fail-closed, alert-aware
Connectivity bugs in IoMT become safety bugs the moment a missed alert lengthens a hospital stay. We default to fail-closed behaviour everywhere PHI crosses a boundary, and we ensure alert pathways have a separate, independently monitored circuit so a downstream EHR outage cannot silence a critical alarm.
- Heartbeat checks across every adapter and a single observability board for clinical engineering.
- Duplicate-event suppression keyed on a deterministic content hash to prevent alarm fatigue.
- Replay tooling that lets engineers re-run a day of telemetry through a new transform without touching production EHR data.
- Documented escalation paths for partial-degradation modes so clinicians know which features are temporarily off.

