🇳🇱 dutch-med-hips

A robust, highly configurable PHI anonymization and surrogate generation toolkit designed for Dutch medical and radiology reports.
It replaces PHI tokens (e.g. <PERSOON>, <Z-NUMMER>, <DATUM>) with realistic surrogate values such as names, ages, dates, hospitals, study names, IDs, BSNs, IBANs, phone numbers, URLs, emails, and more.

dutch-med-hips uses:

• Faker for Dutch-language surrogate generation
• Configurable templates for ID-like fields
• Locale dictionaries for hospitals, months, cities, study names
• A combined regex engine for fast and safe substitution
• Document-level deterministic seeding (optional)

---

Installation

You can install dutch-med-hips via pip:

pip install dutch-med-hips

or from source:

git clone https://github.com/DIAGNijmegen/dutch-med-hips.git
cd dutch-med-hips
pip install -e .

---

Quickstart

Python API

Running HideInPlainSight in your Python code is straightforward:

from dutch_med_hips import HideInPlainSight

text = """
Patiënt <PERSOON> werd opgenomen in <HOSPITAL_NAME> op <DATUM>.
Z-nummer: <Z-NUMMER>, BSN: <BSN>, Email: <EMAIL>.
Rapport ID: <RAPPORT_ID.T_NUMMER>.
""" 

hips = HideInPlainSight()
result = hips.run(text)

print(result["text"])
print(result["mapping"])  # Shows original -> surrogate mapping

Command-Line Interface

You can also use dutch-med-hips directly from the command line:

dutch-med-hips [OPTIONS]

Common Options

Option	Meaning
-i, --input PATH	Input file (UTF-8). Use - or omit to read from stdin.
-o, --output PATH	Output file (UTF-8). Use - or omit to write to stdout.
--mapping-out PATH	Write the JSON mapping (original → surrogate) to a file.
--seed N	Use a fixed seed for deterministic surrogate generation.
--no-document-hash-seed	Disable automatic seeding based on the document hash.
--no-header	Disable the anonymization disclaimer header.
--disable-typos	Disable random typo injection in surrogates.

---

Features

🔐 PHI Surrogates

👤 People & Demographics

• Person names
  • Dutch-style names (first/last), tussenvoegsels (van, de, …)
  • Variants: first-only, last-only, full, initials (J. Jansen, J.S. Jansen)
  • Randomized casing (jan jansen, JAN JANSEN, Jansen, Jan)
• Person initials
  • Derived from full fake names: Jan Steen → JS, Vincent van Gogh → VvG
• Age
  • Sampled from a hospital-like Gaussian mixture model (more 40–85 year olds)

🧾 Identifiers & Numbers

• Patient IDs / Z-numbers / generic PHI numbers
  • All driven by simple templates per tag (e.g. <Z-NUMMER> → Z######)
  • Template mini-language (# = digit, etc.)
• Document IDs & sub-IDs
  • Main report IDs from templates
  • Sub-IDs like <RAPPORT_ID.T_NUMMER> → T123456
• BSN
  • Dutch BSN-like numbers via Faker ssn()
• IBAN
  • Dutch IBANs via Faker, compact or grouped (NL91ABNA0417164300, NL91 ABNA 0417 1643 00)
• Accreditation number
  • Always M + 3 digits (e.g. M007, M123)

🏥 Hospitals, Locations & Studies

• Hospital names
  • Realistic Dutch hospital pool with full names and abbreviations
    e.g. Amsterdam UMC locatie AMC, AMC, Radboudumc, LUMC, ADRZ
  • Sometimes uses only the city as shorthand (e.g. Amsterdam, Nijmegen)
• Locations
  • Dutch cities and place names drawn from hospital/location data
• Study names
  • Curated list of real-looking study labels and variants e.g. LEMA, Donan, M-SPECT/mSPECT, Alpe d'Huzes MRI, TULIP, PRIAS
• Company names
  • Dutch company names via Faker (nl_NL) e.g. Van der Berg BV, Jansen en Zonen, De Vries Holding

📅 Dates & Times

• Dates
  • Dutch-style formats:
    • Numeric: D-M, DD-MM, with/without year (03-02-2025, 3-2-12)
    • Named months: 3 februari, 3 feb 2025
  • Mix of year/no-year, numeric vs month-name
  • Start/end date range configuration (e.g. last 10 years)
• Times
  • 24h clock formats: 13:45, 13:45 uur, 13.45, 13u45
  • Natural Dutch phrases: kwart voor zes, kwart over drie, half vier

📞 Contact & Online

• Phone numbers
  • Dutch mobile numbers (06-12345678, +31 6 12345678)
  • Landlines / hospital numbers (020-5669111, 088-…)
  • Internal SEIN/pager numbers (4–5 digit codes)
• Email addresses
  • Fake but valid emails via Faker (customizable domains)
• URLs
  • Fake but valid http(s) URLs via Faker (can be styled to look like portals/EPD endpoints)

🏠 Addresses & Misc

• Addresses
  • Dutch-style street + number + postcode + city via Faker (nl_NL)
• Other PHI
  • Any additional tag-based IDs or tokens configured via templates can be mapped to surrogates in the same way.

🔧 Flexible Configuration

All defaults live in settings.py and can be overridden at runtime:

from dutch_med_hips import settings

settings.ID_TEMPLATES_BY_TAG["<Z-NUMMER>"] = "Z-###-###"
settings.PERSON_NAME_REUSE_PROB = 0.15
settings.ENABLE_TYPOS = True

🧪 Deterministic Output

The system automatically hashes the document to generate a seed to stabilize output. This can be turned off, or you can provide your own fixed seed:

hips = HideInPlainSight(default_seed=123)

!!! Note Using a fixed seed means the same input document will always yield the same output document and same surrogate mappings.
Different documents will still produce different outputs.

✏️ Optional Typo Injection

Using the typo Python package, some surrogates can receive:

• Adjacent-key typos
• Insertions
• Deletions
  You can enable/disable this globally.

⚠️ Automatic Disclaimer Header

Every anonymized document can automatically receive an anonymization disclaimer at the top.
You may customize or disable it.

---

Defining Custom PHI Tags

dutch-med-hips allows users to extend PHI detection and surrogate generation without modifying the library.
Below are the three most common customization patterns, with clear examples:

1. Add a new regex to an existing PHI category (easiest)
2. Create a new ID-style tag with its own format (advanced)
3. Extend surrogate pools (e.g. add a new study name)

---

1️⃣ Adding a New Regex to Person Names (Simple Example)

If your reports use additional person‑name markers such as <NM> or <PERSOON_NAAM>, you can plug them into the existing system.

Just add them to the existing PHI type:

from dutch_med_hips import schema
from dutch_med_hips.schema import PHIType

schema.DEFAULT_PATTERNS[PHIType.PERSON_NAME].extend([
    r"<NM>",
    r"<PERSOON_NAAM>",
])

These new tags will:

✔ Be recognized as person names
✔ Use the existing name surrogate logic
✔ Appear in the mapping as phi_type="person_name"

Example:

text = "Patiënt <NM> en begeleider <PERSOON_NAAM> kwamen binnen."
print(HideInPlainSight(seed=1).run(text)["text"])

---

2️⃣ Creating a New ID Tag with a Custom Format (Advanced)

Suppose your system uses <CENTER_ID> and you want outputs like:

CEN-123456

Step 1 — Add a detection pattern

from dutch_med_hips import schema
from dutch_med_hips.schema import PHIType

schema.DEFAULT_PATTERNS.setdefault(PHIType.GENERIC_ID, []).append(
    r"<CENTER_ID>"
)

Step 2 — Assign a surrogate template

Use the template mini‑language (# = digit):

from dutch_med_hips import settings

settings.ID_TEMPLATES_BY_TAG["<CENTER_ID>"] = "CEN-######"

Step 3 — Use it

from dutch_med_hips import HideInPlainSight

text = "Centrum: <CENTER_ID>."
print(HideInPlainSight(seed=42).run(text)["text"])
# Centrum: CEN-123456.

---

3️⃣ Adding New Items to Surrogate Pools (e.g., Study Names)

Some surrogate categories use pools, such as the curated list of Dutch medical study names in locale.py:

STUDY_NAME_POOL = [
    "LEMA",
    "Donan",
    ["M-SPECT", "mSPECT"],
    ...
]

To add your own study:

from dutch_med_hips import settings

settings.STUDY_NAME_POOL.append("MY-NEW-STUDY")

Or add multiple variants:

settings.STUDY_NAME_POOL.append([
    "CUSTOMTRIAL",
    "Custom Trial",
    "CT-Study"
])

The generator will randomly pick one of the variants.

Example

text = "Onderzoek: <STUDY_NAME>."
hips = HideInPlainSight(seed=123)
print(hips.run(text)["text"])
# Onderzoek: MY-NEW-STUDY.

---

Optional: Full Custom Generator

For more complex surrogate rules, you can define your own PHI type and generator:

import re
from dutch_med_hips import schema, surrogates

def generate_center(match: re.Match) -> str:
    return "CENTER-" + "123456"

schema.DEFAULT_PATTERNS["center"] = [r"<CENTER_SPECIAL>"]
surrogates.DEFAULT_GENERATORS["center"] = generate_center

Now <CENTER_SPECIAL> maps through your custom function.

---

Summary

Task	Best Method
Add a new tag to an existing PHI category	Add regex to schema.DEFAULT_PATTERNS[...]
Create a new ID-like tag	Add regex + assign template via settings.ID_TEMPLATES_BY_TAG
Add new study/hospital/name variants	Append to the appropriate pool in settings
Create fully custom surrogate logic	Register generator in surrogates.DEFAULT_GENERATORS

Customizing dutch-med-hips is:
Add regex → (optional) assign template or generator → done.

---

NER Label Position Tracking

When working with NER (Named Entity Recognition) pipelines, you often have label positions that need to stay aligned after surrogate substitution. dutch-med-hips can automatically update these positions for you.

Usage

Pass your NER labels as a list of (start, end, label) tuples:

from dutch_med_hips import HideInPlainSight

text = "Patient <PERSOON> geboren op <DATUM> in Amsterdam."
ner_labels = [
    (8, 17, "<PERSOON>"),   # Position of <PERSOON> tag
    (29, 36, "<DATUM>"),    # Position of <DATUM> tag
    (40, 49, "LOCATION"),   # Position of "Amsterdam" (non-PHI)
]

hips = HideInPlainSight(enable_header=False)
result = hips.run(text, ner_labels=ner_labels)

print(result["text"])
# "Patient Jan de Vries geboren op 15-03-1985 in Amsterdam."

print(result["updated_labels"])
# [(8, 20, "<PERSOON>"), (32, 42, "<DATUM>"), (46, 55, "LOCATION")]

How It Works

• Labels before a replacement remain unchanged
• Labels after a replacement are shifted by the length difference (surrogate length - original length)
• Labels that overlap with a replacement have their end position adjusted
• When the header is enabled, all label positions are shifted by the header length

This is useful when you need to:

• Preserve NER annotations through anonymization
• Track entity positions for downstream processing
• Maintain alignment between original and anonymized text annotations

---

Mapping Output Structure

result = hips.run(text) returns:

{
    "text": "anonymized text...",
    "mapping": [
        {
            "original": "<PERSOON>",
            "surrogate": "Jan Steen",
            "phi_type": "person_name",
            "start": 10,
            "end": 18
        },
        ...
    ],
    "updated_labels": [...]  # Only present if ner_labels was provided
}

---