auditing-safe-harbor-checklist

Auditing against the HIPAA Safe Harbor checklist

The Safe Harbor method (45 CFR 164.514(b)(2)) de-identifies PHI by removing 18 specific identifier categories for the individual and their relatives, employers, and household members — and requires the covered entity to have no actual knowledge that the remainder could re-identify anyone. This skill turns that legal checklist into a concrete coverage check over OpenMed output: which of the 18 categories were detected and handled, and where the gaps are.

The full mapping table lives in references/safe-harbor-identifiers.md — all 18 categories, their OpenMed HIPAA class, the matching CANONICAL_LABELS, and per-category cautions. Read it when you need the authoritative cross-walk.

When to use this skill

Use it after a de-identification run to prove coverage, or before release to decide whether Safe Harbor is even achievable for this text. If the user needs a signed, retained record of the run, hand off to auditing-deidentification-runs.

Quick start: coverage check

import openmed
from openmed.core.labels import LABEL_TO_HIPAA, HIPAA_SAFE_HARBOR_CLASSES

note = (
    "Patient John Doe (MRN 1234567), age 92, of Smalltown, seen 2024-03-02. "
    "SSN 123-45-6789, phone 617-555-0142."
)

# 1) Detect identifiers (spans only; no rewrite).
detected = openmed.extract_pii(note)

# 2) Roll each detected span up to its HIPAA Safe Harbor class.
covered = set()
for ent in detected.entities:
    canonical = openmed.normalize_label(ent.label)        # -> CANONICAL_LABELS form
    hipaa_class = LABEL_TO_HIPAA.get(canonical)            # -> one of 18 classes
    if hipaa_class:
        covered.add(hipaa_class)

# 3) Report which of the 18 classes were touched and which weren't observed.
missing = sorted(HIPAA_SAFE_HARBOR_CLASSES - covered)
print("covered:", sorted(covered))
print("not observed in this note:", missing)

"Not observed" is not the same as "absent" — a category may simply not occur in this note, or may have been missed. That is exactly what the human review step (below) is for.

Workflow

1. De-identify with a Safe Harbor profile: openmed.deidentify(note, policy="hipaa_safe_harbor"). This masks every identifier class by default and runs the mandatory structured-ID safety sweep.
2. Map detected spans to the 18 classes via LABEL_TO_HIPAA (as above). Build a table of category → detected? → action taken.
3. Walk the checklist in references/safe-harbor-identifiers.md and flag the known gaps explicitly:
   • Ages > 89 (AGE) must be aggregated to "90+"; OpenMed flags but does not auto-cap — see shifting-clinical-dates.
   • Dates keep only the year; everything else (admit/discharge/DOB) goes.
   • ZIP beyond the first 3 digits, and small-population areas → mask whole.
   • Rare geography (small towns) and rare characteristics (unusual occupation) can re-identify even when masked field-by-field.
   • Fax shares the PHONE label; biometrics and full-face photos are out of scope for text — handle in the imaging/intake pipeline.
4. Assess residual risk. Run audit=True and read residual_risk (auditing-deidentification-runs). Non-zero projected leakage → review.
5. Record the "no actual knowledge" judgment. A human must sign off that the remaining text cannot re-identify the individual. Automated coverage is necessary, not sufficient.

Hand-off to / from OpenMed

• Detect / de-id: openmed.extract_pii (spans) and openmed.deidentify (rewrite) — see deidentifying-clinical-text.
• Label mapping: openmed.CANONICAL_LABELS, openmed.normalize_label, and LABEL_TO_HIPAA / HIPAA_SAFE_HARBOR_CLASSES in openmed/core/labels.py.
• Signed record + residual risk: auditing-deidentification-runs (audit=True → AuditReport.residual_risk).
• Profile choice: configuring-privacy-policies — if you must keep dates or geography, Safe Harbor fails; use Expert Determination (hipaa_expert_review_assist) or a Limited Data Set (research_limited_dataset).

Edge cases & gotchas

• Coverage ≠ compliance. Detecting all 18 categories does not satisfy Safe Harbor on its own — the "no actual knowledge" residual-risk judgment is required and is a human decision.
• Ages over 89 are a transformation, not a detection. Masking the digits is fine; if you keep age, aggregate to "90+". OpenMed will not cap automatically.
• ZIP / date rules are transformations. Safe Harbor permits keeping 3-digit ZIP (population-gated) and the year — implement the truncation; do not assume detection handles it.
• Some categories have no text label (biometrics, full-face photos). Mark them N/A for text and ensure another pipeline stage covers them.
• Combination re-identification. Several non-identifying quasi-identifiers together (rare diagnosis + small town + outlier age) can identify someone; this is precisely why strict_no_leak exists for high-stakes data.
• No raw PHI in the checklist output — report categories, counts, offsets, and hashes, never the underlying identifiers.

Standards & references

• 45 CFR 164.514(b)(2) (Safe Harbor): https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.514
• HHS de-identification guidance (incl. the 18 identifiers and "actual knowledge"): https://www.hhs.gov/hipaa/for-professionals/privacy/special-topics/de-identification/index.html
• Local checklist: references/safe-harbor-identifiers.md
• OpenMed source: openmed/core/labels.py (LABEL_TO_HIPAA, HIPAA_SAFE_HARBOR_CLASSES, CANONICAL_LABELS, normalize_label).