zorgdeid 1.1.0

ZorgDeID: PII detection, masking & anonymization for Dutch health tech.

ZorgDeID — Dutch PII Detection & Anonymization

Detect, mask, and anonymize Personally Identifiable Information (PII) in Dutch text and documents.
Built for Dutch healthcare, GDPR / AVG compliance, and NEN 7510 data-protection pipelines.

---

What is ZorgDeID?

ZorgDeID is a Python library that detects and anonymizes Dutch PII in plain text, documents (.pdf, .docx, .txt), and scanned files (.jpg, .png, .tiff, and image-based PDFs via OCR). It combines:

• Custom Dutch regex recognizers — hand-tuned patterns for Dutch identifiers (BSN, IBAN, Health Identifier, licence plates, …)
• spaCy Dutch NER (nl_core_news_lg) — neural named-entity recognition for persons and locations
• Algorithmic validation — elfproef for BSN, mod-97 for IBAN, Luhn for credit cards and IMEI
• Context-aware scoring — sentence-bounded keyword windows boost or penalize confidence before anonymization decisions

Use cases: de-identifying patient records, anonymizing clinical notes, sanitizing intake forms, GDPR / AVG data-minimization pipelines, NEN 7510 technical controls.

---

Key Features

Feature	Detail
18 entity types	Full Dutch PII coverage — from BSN and Health Identifier to GPS coordinates and IMEI
3 guard modes	anonymize (realistic Dutch fakes) · tag ([PERSON]) · i_tag ([PERSON_1])
Document support	Reads .pdf (pypdf), .docx (python-docx), and .txt natively
OCR support	Scanned PDFs and image files (.jpg, .png, .tiff, .bmp, .webp) via EasyOCR — no system binary required
PDF normalization	Automatically repairs pypdf extraction artifacts; falls back to OCR for image-only pages
Algorithmic validation	BSN elfproef · IBAN mod-97 · Credit card & IMEI Luhn
Context-aware scoring	Sentence-bounded keyword windows boost or penalize confidence scores
Vocabulary tiers	Strong / weak keyword distinction — partial boosts for ambiguous context
Negative context	Contradicting keywords reduce score before thresholding
Entity filtering	keep allowlist or ignore denylist per call
Custom patterns	Plug in your own regex with optional context words and fake-value pools
GDPR / AVG ready	Designed for Dutch healthcare data pipelines and NEN 7510 technical controls

---

Installation

pip install zorgdeid

Download the Dutch spaCy model (required for PERSON and LOCATION detection):

python -m spacy download nl_core_news_lg

Optional: OCR support

To process scanned PDFs and image files (.jpg, .png, .tiff, .bmp, .webp):

pip install zorgdeid[ocr]

This installs EasyOCR and PyMuPDF — no system binaries (Tesseract, poppler) required. The Dutch neural model (~120 MB) is downloaded automatically on first use and cached in ~/.EasyOCR/.

---

Quick Start

from zorgdeid import analyze, guard

text = "Mijn naam is Jan de Vries en ik woon in Amsterdam. Mijn BSN is 123456782."

# ── Detect PII ────────────────────────────────────────────────────
findings = analyze.text(text)
for f in findings:
    print(f"[{f['type']}] {text[f['start']:f['end']]} (score: {f['score']})")
# [PERSON]   Jan de Vries  (score: 0.85)
# [LOCATION] Amsterdam     (score: 0.85)
# [BSN]      123456782     (score: 0.85)

# ── Anonymize (default mode) ──────────────────────────────────────
result = guard.text(text)
print(result["guarded_text"])
# "Mijn naam is Maria Janssen en ik woon in Utrecht. Mijn BSN is 111222333."

# ── Tag mode ──────────────────────────────────────────────────────
print(guard.text(text, config={"mode": "tag"})["guarded_text"])
# "Mijn naam is [PERSON] en ik woon in [LOCATION]. Mijn BSN is [BSN]."

# ── Indexed tag mode ──────────────────────────────────────────────
print(guard.text(text, config={"mode": "i_tag"})["guarded_text"])
# "Mijn naam is [PERSON_1] en ik woon in [LOCATION_1]. Mijn BSN is [BSN_1]."

---

Document Processing

Process files directly — text extraction and PII analysis in one call:

from zorgdeid import analyze, guard

# Analyze a file
findings = analyze.doc("patient_report.pdf")
findings = analyze.doc("intake_form.docx")
findings = analyze.doc("clinical_note.txt")

# Anonymize a file
result = guard.doc("patient_report.pdf")
print(result["guarded_text"])   # clean, anonymized text
print(result["findings"])       # list of detected PII spans

# All config options work the same as with .text()
result = guard.doc("intake_form.docx", config={
    "mode": "tag",
    "score_threshold": 0.6,
    "set_entities": {"keep": ["PERSON", "BSN", "IBAN_CODE"]},
})

Supported formats:

Format	Reader	Notes
.txt	built-in open()	UTF-8 with cp1252 fallback
.pdf	pypdf	Digital PDFs extracted directly; image-only pages fall back to OCR automatically
.docx	python-docx	All paragraphs and table cells joined
.jpg .jpeg .png .tiff .bmp .webp	EasyOCR	Requires pip install zorgdeid[ocr]

Any other extension raises UnsupportedFormatError before the file-existence check.

OCR example

from zorgdeid import analyze, guard
from zorgdeid.processors.doc_processor import read as doc_read

# Extract text via OCR
text = doc_read("intake_scan.png")
print(text)

# Detect PII
findings = analyze.doc("intake_scan.png")

# Anonymize
result = guard.doc("intake_scan.png", config={"mode": "tag"})
print(result["guarded_text"])

# Works identically for scanned PDFs
result = guard.doc("patient_record_scan.pdf", config={"mode": "anonymize"})
print(result["guarded_text"])

---

Supported Entity Types

Entity	Description	Validation
PERSON	Person names	spaCy NER
LOCATION	Cities, addresses, regions	spaCy NER
DATE	Dates (numeric & Dutch month names)	—
TIME	Times (12h / 24h / Dutch "uur")	—
PHONE_NUMBER	Dutch mobile & landline, EU format	—
EMAIL_ADDRESS	E-mail addresses	—
URL	HTTP/HTTPS/FTP links	—
ZIPCODE	Dutch postal codes (1234 AB)	—
GPS_COORDINATES	Latitude / longitude pairs	—
IBAN_CODE	Dutch & international IBANs	✓ ISO 13616 mod-97
CREDIT_CARD	Visa, Mastercard, Amex, Diners, Discover, JCB	✓ Luhn
BSN	Burgerservicenummer	✓ Elfproef (11-proef)
PASSPORT	Dutch passport & driving licence numbers	—
HEALTH_IDENTIFIER	Dutch health insurance policy numbers	—
IP_ADDRESS	IPv4 and IPv6 addresses	—
MAC_ADDRESS	Ethernet MAC addresses	—
IMEI	Mobile device identifiers (15 digits)	✓ Luhn
LICENCE_PLATE	Dutch vehicle licence plates	—

---

Guard Modes

Mode	Behaviour	Output example
anonymize (default)	Replace each entity with a realistic Dutch synthetic value	Jan Bakker, 111222333, NL20 INGB 0001 2345 67
tag	Replace with [ENTITY_TYPE]	[PERSON], [BSN], [IBAN_CODE]
i_tag	Replace with [ENTITY_TYPE_N] — same entity type gets the same index	[PERSON_1] … [PERSON_2]

---

Configuration

All options are passed via a single config dict:

# Allowlist — only detect these entity types
config = {"set_entities": {"keep": ["PERSON", "BSN", "IBAN_CODE"]}}

# Denylist — detect everything except these
config = {"set_entities": {"ignore": ["DATE", "TIME"]}}

# Full config example
config = {
    "set_entities": {"keep": ["PERSON", "BSN", "IBAN_CODE"]},

    # Minimum confidence to include a finding
    "score_threshold": 0.5,

    # Guard mode
    "mode": "anonymize",   # "anonymize" | "tag" | "i_tag"

    # Custom patterns (see below)
    "custom_patterns": [...],
}

Custom Patterns

from zorgdeid import analyze, guard, custom_pattern

emp = custom_pattern(
    name="EMPLOYEE_ID",
    regex=r"EMP-\d{4}",
    score=0.9,
    context=["medewerker", "werknemer"],        # nearby words boost score
    anonymize_list=["EMP-9999", "EMP-8888"],    # fake pool for anonymize mode
)

findings = analyze.text("Medewerker EMP-1234 heeft toegang.", config={"custom_patterns": [emp]})
guarded  = guard.text("Medewerker EMP-1234 heeft toegang.",  config={"custom_patterns": [emp]})
print(guarded["guarded_text"])
# "Medewerker EMP-9999 heeft toegang."

---

Scoring & Confidence

Every finding carries a score between 0 and 1. Scores are determined by a four-tier system:

Tier	Condition	Example score
base	Regex match only, no additional evidence	0.30 – 0.85
with_context	A relevant keyword appears in the same sentence	up to 0.95
validated	Algorithmic checksum passes (elfproef / mod-97 / Luhn)	0.65 – 0.90
high_confidence	Validation and context keyword present	0.90 – 0.95

Context scoring is sentence-aware — context keywords from other sentences do not influence the score. Negative-context keywords (e.g. factuurnummer near a phone pattern) actively reduce confidence.

Use score_threshold to filter out low-confidence results before anonymization.

---

Package Layout

zorgdeid/
├── types.py              — core data structures (RecognizerResult, Pattern, …)
├── analysis/
│   ├── analyzer.py       — PII analysis engine (GuardAnalyzer)
│   ├── context_awareness.py  — sentence-aware keyword scoring (DutchContextEnhancer)
│   └── overlap_resolver.py   — span deduplication & merging
├── anonymization/
│   ├── engine.py         — stateless anonymization dispatcher (GuardEngine)
│   └── fake_data.py      — synthetic Dutch PII pools
├── recognizers/
│   ├── base.py           — EntityRecognizer / PatternRecognizer base classes
│   ├── contact.py        — PHONE_NUMBER, EMAIL_ADDRESS, URL
│   ├── datetime.py       — DATE, TIME
│   ├── device.py         — IP_ADDRESS, MAC_ADDRESS, IMEI
│   ├── financial.py      — IBAN_CODE, CREDIT_CARD
│   ├── identifier.py     — BSN, PASSPORT, HEALTH_IDENTIFIER
│   ├── location.py       — ZIPCODE, GPS_COORDINATES
│   ├── spacy_recognizer.py  — NER recognizer (PERSON, LOCATION)
│   └── vehicle.py        — LICENCE_PLATE
├── processors/
│   ├── text_processor.py — analyze / guard pipelines for plain-text input
│   └── doc_processor.py  — file reading (.pdf / .docx / .txt / images) + OCR fallback
├── config/
│   ├── entities.py       — ALL_NL_ENTITY_TYPES list
│   └── scoring.py        — EntityScoreProfile per entity type
└── patterns/             — Dutch regex patterns & keyword lists

The public interface is exposed through zorgdeid/__init__.py:

from zorgdeid import analyze, guard, custom_pattern, ALL_NL_ENTITY_TYPES

---

Privacy & Compliance

Standard	How this library helps
GDPR / AVG	De-identifies personal data before storage or transfer; supports data-minimization obligations
NEN 7510	Provides a technical control layer for pseudonymization of Dutch patient data
Human-in-the-loop	Automated detection is probabilistic — for critical clinical datasets, always include human review of anonymized output

This library is a technical tool, not a legal guarantee. Your full pipeline architecture, access controls, and data governance policies must meet the applicable regulatory requirements.

---

Interactive Quickstart

The examples/quickstart.ipynb notebook covers:

• Text and document analysis
• All three guard modes
• Dutch healthcare identifiers (BSN, HEALTH_IDENTIFIER)
• Custom patterns with anonymization pools
• Entity filtering and score thresholds
• Error handling for unsupported file formats
• OCR on image files and scanned PDFs

---

License

MIT License — see LICENSE for details.