Skip to main content
Version: v0.1.6

Check a de-identifier with the pairing loop

@cosyte/synth makes data that looks like PHI and holds none; @cosyte/deid removes PHI. The @cosyte/synth/deid subpath runs one through the other, deterministically, from a seed:

  1. Generate an artifact with the synth generators.
  2. List the synthetic identifiers synth planted at the patient fields (the sentinels).
  3. De-identify the artifact with @cosyte/deid.
  4. Check that no sentinel is left at those fields, and that the clinical codes are still there (the over-scrub check).

This shows the pair works on synth's own output. It is not an independent check of @cosyte/deid against real-world documents. A sentinel that @cosyte/deid blocks rather than redacts counts as removed. The loop covers the five formats both packages support; NCPDP SCRIPT, ASTM, and DICOM are skipped, and named.

Install @cosyte/deid with @cosyte/hl7, @cosyte/fhir, @cosyte/ccda, @cosyte/x12, and @cosyte/ncpdp: the subpath loads all five parsers when it is imported (Installation).

Run one artifact through the loop​

import { hl7DeidLoop } from "@cosyte/synth/deid";

const r = hl7DeidLoop({ seed: 42, kind: "ORU^R01" });
r.pass; // => true
r.survivors; // => []
r.clinicalScrubbed; // => []

kind defaults to "ORU^R01". r.planted lists the synthetic tokens synth placed at the identifier fields, such as a name from the invented-name pool, an SSN with a 900-series area, a phone number in 555-0100 to 555-0199, a medical record number under the synthetic assigning authority, and the order numbers. r.survivors lists the ones still present at those fields after de-identification, and r.clinicalScrubbed the clinical codes that went missing. Both are empty on a pass.

Run every covered format and summarize​

import {
  hl7DeidLoop,
  fhirDeidLoop,
  x12DeidLoop,
  ncpdpTelecomDeidLoop,
  ccdaDeidLoop,
  summarizeDeidCoverage,
} from "@cosyte/synth/deid";

const summary = summarizeDeidCoverage([
  hl7DeidLoop({ seed: 1 }),
  fhirDeidLoop({ seed: 1 }),
  x12DeidLoop({ seed: 1, variant: "837P" }),
  ncpdpTelecomDeidLoop({ seed: 1, transaction: "B1" }),
  ccdaDeidLoop({ seed: 1 }),
]);
summary.allPass; // => true
summary.totalSurvivors; // => 0
summary.totalClinicalScrubbed; // => 0

Why the check looks only at the patient fields​

synth draws patient names, provider and organization names, addresses, and phone numbers from the same synthetic pools, and a de-identifier may rightly keep a provider's identity. So the removal check reads only the values left at the fields that held patient identifiers, not the whole document. A provider name that the de-identifier keeps, and that happens to match a patient's, is not reported as a survivor. The check reads the de-identifier's real serialized output, so a de-identifier that finds a field and fails to clear it is still caught.

The same seed gives the same result​

The loop depends only on the seed: the same seed gives the same artifact, the same de-identified output, and the same sentinels.

import { fhirDeidLoop } from "@cosyte/synth/deid";

fhirDeidLoop({ seed: 7 }).deidentified === fhirDeidLoop({ seed: 7 }).deidentified; // => true

See which formats are skipped​

DEID_LOOP_SKIPPED names each format the loop does not cover, with the reason:

import { DEID_LOOP_SKIPPED } from "@cosyte/synth/deid";

DEID_LOOP_SKIPPED.map((s) => s.format); // => ["ncpdp-script", "astm", "dicom"]

@cosyte/deid ships no NCPDP SCRIPT or ASTM adapter, and synth does not generate DICOM, so those pairings wait for the adapters.

Selectors the loops do not check yet​

x12DeidLoop({ variant }) and ncpdpTelecomDeidLoop({ transaction }) throw a plain TypeError for a value outside their union, rather than a SynthError, and ccdaDeidLoop({ documentType }) treats any value but "ccd" as a Referral Note. None of them quotes your value.