The short answer
Before scoping anything, find out what the record system will actually let you do, at what cost, and how long approval takes. Many integration projects are delayed not by engineering but by waiting for a vendor to enable an interface.
Ask for it in writing, with a timeline. Verbal assurances about API availability have a way of becoming quarters.
The questions to ask first
- What interfaces exist: modern API, messaging standard, file exchange, or nothing?
- What does access cost, one-off and ongoing?
- Who has to approve it, and how long does that take?
- Is there a test environment, and does it behave like production?
- What are the rate limits and what happens when you hit them?
Point four is regularly the one that hurts. A test environment with different data shapes or different behaviour means the first real surprise happens in production.
Standards help, and do not settle it
Interoperability standards give you a shared vocabulary and structure. What they do not give you is consistency in how each system implements them, which fields are populated, or what local codes mean.
- Optional fields are frequently empty in practice
- Local code sets sit alongside standard ones
- The same concept can be represented several valid ways
- Extensions vary by vendor and by deployment
Plan to map against the actual data rather than against the specification. The specification tells you what is permitted, not what you will receive.
Read is easier than write
| Direction | Typical difficulty |
|---|---|
| Read demographics | Usually straightforward |
| Read clinical data | Available, mapping is the work |
| Write administrative data | Possible, needs agreement |
| Write clinical data | Hardest, governance as much as technical |
| Trigger workflow in the EHR | Often not exposed at all |
Design so that as much as possible is read-only. Anything writing into a clinical record carries governance requirements that add substantially to the project.
Where projects stall
Waiting for vendor approval, discovering the test environment is unrepresentative, finding the field you needed is not populated in this deployment, and governance approval that nobody scheduled.
All four are foreseeable. Building a throwaway proof against the real interface in the first fortnight surfaces the first three, and involving governance at scoping handles the fourth.