Extraction patterns across EHRs

The same handful of problems show up on almost every EHR. Each EHR page lists its own variants; this page collects the patterns.

Bulk export is rarely usable as advertised

What happens	Seen on
Works, but one resource type per job, sequentially, and slowly	eClinicalWorks
Latency from 3 minutes to 4+ hours, unrelated to data volume	athenahealth
Manifest returns 200 while its output list is still filling	ModMed
Accepts every kickoff, then crawls (1% after 34 minutes)	Nextech
Only Group/[id]/$export; without _since, only the last 24 hours are exported	Greenway
_type ignored — every type comes back every time	Practice Fusion
Requires a patient Group the customer must build	Epic, NextGen

Design for it: detect stalls by watching progress stop changing (for example the X-Progress header) rather than using a wall-clock timeout; require the manifest to be stable across several polls before accepting it; and keep a search-based fallback.

Population search is often forbidden

Many EHRs refuse searches that are not scoped to a patient — athenahealth returns 403, others return 400 “patient or _id must be supplied”. Your cohort then has to come from somewhere else: a bulk export of encounters, encounters searched by practitioner NPI, a patient list you already hold, or (as a last resort) enumerating patients by name.

_since and _lastUpdated select by edit time

They filter on meta.lastUpdated. A January visit edited in June appears in an export with _since in June; an unedited one does not. Some servers ignore _lastUpdated entirely. Always decide measurement-year membership yourself from period.start or effective[x].

Paging is the biggest source of silent data loss

Pitfall	Seen on
Default page of 10	Oracle Health (Cerner)
Default page of 20 with no next link unless you send _count	athenahealth
Per-resource page caps (50, 90, 100) whatever _count you send	athenahealth
Next-page offset computed from the _count you asked for, not what was served — the next page 404s	athenahealth
Bundle total reports the page size, not the number of matches	ModMed
_count rejected with 400; only link.next works	eClinicalWorks
page=N only, no link.next	Compulink
Vendor parameters duplicated onto next URLs cause 403	athenahealth

Rules that cover all of them: always send an explicit _count; always walk link.next to the end; cap the number of pages; and when you hit the cap or a page fails, mark the patient incomplete rather than treating the missing data as clinically absent.

HTTP 200 with zero results is not proof of absence

• A bare year in a date search (date=2026) returns 200 with nothing on some servers; the prefixed form (date=ge2026-01-01) returns data. (ModMed)
• category=vital-signs returns 200 with nothing; searching by the blood-pressure panel code works. (eClinicalWorks)
• A practice that has not enabled your app returns 200 with empty searchsets. (Greenway)
• The reverse also exists: 404 {"error":"[no-data]"} meaning an empty result. (Compulink)
• Scopes granted, but the clinical module that feeds FHIR is not populated.

References often lack type

athenahealth, Practice Fusion, Nextech and Compulink send subject: { reference: "Patient/123" } without type. Inject it from the reference prefix before linking resources to patients.

Encounter status and type are unreliable

• 25–35% of completed encounters arriving as in-progress (athenahealth).
• status absent or unknown (some bundle exports).
• Encounter.type that is a data-absent reason (eClinicalWorks), a generic “summary clinical document” code (Practice Fusion), or a proprietary code that matches no measure value set (Oracle Health).

A practical rule: only enforce the measure’s qualifying-encounter value sets when they match a meaningful share of the year’s encounters (we use 10%); otherwise fall back to counting any completed visit, and flag the result.

Coded data quirks

• Scan every coding, never just coding[0]. On Oracle Health the first coding of a blood pressure component is proprietary; the LOINC is further down. A coding[0] parser finds zero readings.
• Drug codes may exist only on a referenced Medication resource, not inline on the MedicationRequest (athenahealth).
• Screenings may be stored as Condition resources rather than reports or procedures (Practice Fusion).
• Depression screening scores may exist only inside note attachments (HTML or PDF), or only as QuestionnaireResponse (Elation).
• Duplicate lab results with different coding and units for the same order (eClinicalWorks A1c — see that page).

Attribution is not in FHIR

TIN is never in the data. NPI is often missing or published as a data-absent reason. Plan for a practice roster or mapping you maintain, and for attributing each encounter to its primary performer (participant type PPRF) who belongs to the practice.

Resilience

• Request level: a few retries on 0, 429 and 5xx with exponential backoff, honouring Retry-After.
• Job level: pause the whole cascade on sustained 429s (double the pause, cap it), and mark an endpoint dead after repeated consecutive failures.
• Daily quotas matter more than burst limits: a per-patient cascade on a 70,000-patient practice used a whole day’s API allowance while peaking at only 20% of the per-second limit.
• Prevent overlapping runs with a single-instance lock; make runs resumable.