shifting-clinical-dates

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.