auditing-part11-trails

Auditing 21 CFR Part 11 trails for OpenMed pipelines

In FDA-regulated GxP work (GCP clinical trials, GLP, GMP) any electronic record used to support a regulatory decision must meet 21 CFR Part 11: it has to be attributable (who), contemporaneous and time-stamped (when), describe what changed, be tamper-evident, and — where a signing event occurs — carry a controlled electronic signature. These map onto the ALCOA+ data-integrity expectations (Attributable, Legible, Contemporaneous, Original, Accurate, +Complete/Consistent/Enduring/Available).

OpenMed's deidentify(..., audit=True) already emits a deterministic, PHI-free AuditReport that you can .sign() (HMAC-SHA256) and later .verify(). That gives you the tamper-evidence and attribution primitives; this skill wraps them in the who/when/what/e-signature envelope Part 11 wants.

This is a compliance-enablement aid. Part 11 compliance also requires validated systems (CSV), SOPs, and access controls that live outside any single library — a QA/validation lead signs off.

When to use

• OpenMed runs inside a validated/GxP environment and each run must leave an attributable, tamper-evident record.
• You need to wrap an OpenMed AuditReport with who/when/what + an e-signature manifestation (meaning, signer, timestamp).
• You must verify a stored trail hasn't been altered, or produce CSV evidence for an inspection.

How OpenMed's AuditReport maps to Part 11

Part 11 expectation	21 CFR cite	OpenMed mechanism
Tamper-evident, accurate copies	11.10(b),(c)	AuditReport.to_json() + repro_hash over the canonical payload
Audit trail: what changed, when	11.10(e)	AuditReport.spans (action per identifier), input_hash/deidentified_text_hash, openmed_version, manifest_hash
Operational/authority checks; attribution	11.10(d),(g)	AuditSignature.key_id (signer/key identity) + your envelope's user id
Signature manifestation (name, date, meaning)	11.50	Your envelope fields signer, signed_at, meaning
Signature/record linking, non-repudiation	11.70, 11.200	HMAC-SHA256 over the canonical payload via .sign() / .verify()

The HMAC binds the signature to that exact report content: any later edit to a span, hash, or field changes repro_hash, so .verify() fails — that is the tamper-evidence.

Quick start

import openmed, json, datetime as dt

note = "Subject S-014 (DOB 1962-08-09) reported headache on 2024-05-01."

# 1) Produce the deterministic, PHI-free audit record for this processing step.
report = openmed.deidentify(note, policy="hipaa_safe_harbor", audit=True)

# 2) Sign it with a controlled release key (stored in a vault / HSM, never in code).
report.sign(b"<release-hmac-key>", key_id="omv-signer-2026")

# 3) Wrap in a Part 11 envelope: who / when / what / signature meaning.
trail = {
    "record": report.to_dict(),              # tamper-evident, no PHI
    "who": "[email protected]",        # authenticated user (your IdP)
    "when": dt.datetime.now(dt.timezone.utc).isoformat(),
    "what": "PHI de-identification of source narrative (study X, subject S-014)",
    "signature_manifestation": {             # 21 CFR 11.50
        "signer_printed_name": "Jane Smith",
        "meaning": "reviewed and approved",
        "signed_at": dt.datetime.now(dt.timezone.utc).isoformat(),
    },
    "system": {"openmed_version": report.openmed_version,
               "manifest_hash": report.manifest_hash},
}
with open("part11_trail.json", "w") as fh:
    json.dump(trail, fh, indent=2, sort_keys=True)

# 4) Later — verify integrity (optionally bind to the exact source/output text).
ok = report.verify(b"<release-hmac-key>", original_text=note)
assert ok, "AUDIT TRAIL TAMPERED OR KEY MISMATCH"

Workflow

1. Authenticate the actor in your own IdP/access system (Part 11 needs unique IDs and operational checks — outside the library). Capture the user id.
2. Run the processing step with audit=True to get the deterministic record.
3. Sign with a controlled release key from a vault/HSM; record key_id. Never embed the key in source or the trail.
4. Build the Part 11 envelope — who, when (UTC, contemporaneous), what, and the signature manifestation (printed name, meaning, timestamp) per 11.50.
5. Append, never overwrite. Store trails write-once (WORM / append-only store). The audit trail itself must be protected and retained.
6. Verify on retrieval with .verify(key, original_text=..., deidentified_text=...) to confirm neither the record nor the bound texts changed.
7. Retain per the study/retention schedule; keep keys and any reversible mapping in a separate, access-controlled store.

Hand-off to / from OpenMed

• Produce the record: auditing-deidentification-runs (deidentify(audit=True) → AuditReport) is the source of the signed, PHI-free trail this skill envelopes.
• Coverage evidence: auditing-safe-harbor-checklist documents that the 18 identifier categories were handled — useful as a CSV artifact.
• No-PHI logging: enforcing-nophi-logging ensures the surrounding application logs don't leak identifiers into the trail.
• HIPAA overlap: checking-hipaa-compliance — Part 11 audit controls and the HIPAA Security Rule audit-controls standard (164.312(b)) reinforce each other.
• OpenMed runs on-device, so the record-generating step stays inside your validated boundary.

Edge cases & gotchas

• Part 11 ≠ one library. The signed AuditReport gives tamper-evidence and attribution, but Part 11 also requires validated systems (CSV), SOPs, training, and access controls you implement around it. Don't claim "Part 11 compliant" from the audit object alone.
• Unsigned = not tamper-evident. .sign() is a deliberate step; signature is None until called. Empty/None keys are rejected.
• Key management is the crux. The HMAC is only as trustworthy as the key. Use a vault/HSM, rotate via key_id, and never store the key with the trail.
• Contemporaneous timestamps. Use a synchronized, trusted clock (UTC) at the moment of the action — back-dating breaks ALCOA "Contemporaneous".
• Append-only retention. A trail you can silently overwrite isn't an audit trail. Use WORM/append-only storage and protect it from the operators it audits.
• No PHI in the envelope. The AuditReport is hash-and-offset only; don't reintroduce identifiers in the what/who free-text fields.
• HMAC is symmetric. It proves integrity to holders of the key, not public non-repudiation. If you need third-party non-repudiation, layer an asymmetric signature over report.to_json().

Standards & references

• 21 CFR Part 11 (Electronic Records; Electronic Signatures): https://www.ecfr.gov/current/title-21/chapter-I/subchapter-A/part-11
• FDA guidance — Part 11 Scope and Application: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/part-11-electronic-records-electronic-signatures-scope-and-application
• FDA — Electronic Systems, Records, and Signatures in Clinical Investigations (2023): https://www.fda.gov/regulatory-information/search-fda-guidance-documents/electronic-systems-electronic-records-and-electronic-signatures-clinical-investigations
• FDA — Data Integrity and Compliance With Drug CGMP (ALCOA+): https://www.fda.gov/regulatory-information/search-fda-guidance-documents/data-integrity-and-compliance-drug-cgmp-questions-and-answers-guidance-industry
• GAMP 5 (ISPE) — risk-based CSV: https://ispe.org/publications/guidance-documents/gamp-5-guide-2nd-edition
• HMAC-SHA256: RFC 2104 / FIPS 198-1. OpenMed source: openmed/core/audit.py.