技能 人工智能 临床日期偏移去识别化

临床日期偏移去识别化

v20260803
shifting-clinical-dates
该技能通过应用一致的日期偏移,实现临床文本的去识别化。它在使绝对日期不可知的情况下,关键地保留了医学事件之间的时间间隔。这对于需要进行纵向研究或生存分析的场景至关重要,确保在符合HIPAA安全港规则的前提下,保留时间到事件的完整性。
获取技能
181 次下载
概览

Shifting clinical dates

HIPAA Safe Harbor forbids keeping dates more specific than the year. But naively deleting dates destroys the temporal structure research depends on — time to event, length of stay, intervals between visits. Date shifting is the compromise: move every date by a single random offset so the absolute dates become meaningless while the intervals between them are preserved exactly. OpenMed does this on-device with deidentify(method="shift_dates", ...).

When to use this skill

Use it when downstream analysis needs temporal relationships (survival curves, sepsis-to-antibiotic time, readmission gaps) but the calendar dates must be de-identified. If you can throw dates away entirely, plain method="mask" is simpler — reach for shifting only when intervals matter.

Quick start

import openmed

note = (
    "Admitted 2024-03-02, started antibiotics 2024-03-04, discharged 2024-03-09. "
    "Follow-up scheduled 2024-04-02."
)

# Shift every date by the SAME offset -> intervals preserved, dates obscured.
result = openmed.deidentify(
    note,
    method="shift_dates",
    consistent=True,        # one stable offset for this run
    seed=20240519,          # reproducible per-patient offset (use a per-patient key)
    keep_year=False,        # do NOT retain the year (Safe Harbor: year-only is the max)
)
print(result.deidentified_text)
# Admit -> antibiotics is still 2 days; admit -> discharge still 7 days; etc.

How interval preservation works

All dates in the document are moved by one offset (auto-selected as a random non-zero value in roughly ±1 year, or fixed with date_shift_days=). Because the offset is identical for every date, the difference between any two dates is unchanged:

real:    Mar 2  ──2d──▶ Mar 4  ──5d──▶ Mar 9
shifted: Jul 18 ──2d──▶ Jul 20 ──5d──▶ Jul 25     (offset = +138 days, intervals intact)

That is why survival time, length of stay, and visit gaps survive de-identification while the actual calendar is destroyed.

Per-patient consistent offsets

Each patient should get their own offset, and that offset should be stable across documents and reproducible across runs. Derive a per-patient seed (e.g. from a secret keyed hash of the patient ID — never the raw MRN) and pass it as seed=:

def patient_offset_seed(patient_key: str) -> int:
    import hashlib, hmac
    # keyed so the mapping from patient -> offset is itself a secret
    digest = hmac.new(b"<vault-secret>", patient_key.encode(), hashlib.sha256).digest()
    return int.from_bytes(digest[:8], "big")

for doc in patient_documents:
    openmed.deidentify(
        doc, method="shift_dates",
        consistent=True, seed=patient_offset_seed(patient_id),
        keep_year=False,
    )

Same patient → same offset everywhere (their notes stay internally consistent); different patients → different offsets (cross-patient dates cannot be aligned).

Workflow

  1. Decide the offset policy. Per-patient is standard for clinical research. Use consistent=True + a per-patient seed. Use a fixed date_shift_days= only when a deterministic, externally-managed offset is required.
  2. Set keep_year=False for Safe Harbor. (keep_year=True retains the year, which is permissible only if dates aren't tied to an individual's care.)
  3. Cap ages over 89 separately. Date shifting moves dates; it does not aggregate ages. Safe Harbor requires ages >89 and any date implying age >89 to collapse to a single "90+" — handle AGE spans explicitly (auditing-safe-harbor-checklist).
  4. Keep names/IDs handled too. shift_dates only touches dates. Run a normal redaction pass (or a policy) for PERSON, ID_NUM, etc.
  5. Verify intervals are preserved and no calendar leaked via audit=True.

Hand-off to / from OpenMed

  • Core de-id: deidentifying-clinical-text — combine date shifting with a policy= so names/IDs are redacted in the same pipeline.
  • Safe Harbor checklist: auditing-safe-harbor-checklist — the date rule and the age-90 cap are categories C in the 18.
  • Audit: auditing-deidentification-runs records the method and per-span actions (offsets/hashes, not raw dates).
  • Other surfaces: MCP openmed_deidentify / REST POST /pii/deidentify accept the same method and parameters.

Edge cases & gotchas

  • Offset 0 is forbidden. An auto-selected offset is always non-zero — a zero shift would silently leave dates unchanged and defeat de-identification.
  • Reuse the seed per patient, not globally. A single global offset lets an attacker re-align all patients to a real anchor date; per-patient offsets break that. Derive seeds from a secret, not the plaintext MRN.
  • keep_year=True is not Safe Harbor by itself when the year reveals the age of someone >89 or ties to care episodes — pair with age capping.
  • Shifting does not cap ages. Age >89 is a separate transformation; date shifting won't fix an explicit "age 94" in the text.
  • Day-first locales. For non-English notes set lang= so 11/04/2024 is parsed in the right order before shifting (deidentifying-multilingual-text).
  • Store the per-patient seed/offset like PHI — it re-identifies the calendar if leaked. Keep it in a vault, separate from the output.

Standards & references

信息
Category 人工智能
Name shifting-clinical-dates
版本 v20260803
大小 6.51KB
更新时间 2026-08-04
语言