unpii

High-performance French medical text anonymization library. Rust core with Python bindings.

Designed to process millions of documents efficiently. Inspired by Micropot/incognito

Installation

pip install unpii

# With Polars support
pip install unpii[polars]

Quick Start

import unpii

text = "Dr Martin au 06 12 34 56 78, email: martin@chu-brest.fr"

# Anonymize with placeholders (default)
unpii.anonymize(text)
# → "Dr <NOM> au <TELEPHONE>, email: <EMAIL>"

# Anonymize with stars
unpii.anonymize(text, style="stars")
# → "Dr ***** au *****, email: *****"

Detection Modes

Two detection levels: standard (reliable patterns) and paranoid (aggressive).

# Standard: titles, known patterns, blacklisted names
unpii.anonymize("Dr Martin est ici")
# → "Dr <NOM> est ici"

unpii.anonymize("DUPONT Jean est ici")
# → "DUPONT Jean est ici"  (not detected in standard)

# Paranoid: also catches UPPERCASE Titlecase patterns, 5+ digit sequences, loose emails
unpii.anonymize("DUPONT Jean est ici", mode="paranoid")
# → "<NOM> est ici"

Custom Words to Mask

Pass additional words to mask per call. Useful when you know the patient's name:

unpii.anonymize("bob dylan est ici", mask=["bob", "dylan"])
# → "<PII> <PII> est ici"

Case-insensitive with word boundary checks:

unpii.anonymize("Bonjour Bob", mask=["bob"])
# → "Bonjour <PII>"

Ignore Groups

Skip specific categories:

unpii.anonymize("Dr Martin au 06 12 34 56 78", ignore_groups=["TELEPHONE"])
# → "Dr <NOM> au 06 12 34 56 78"

Inspect Detected Spans

Dry-run mode to see what would be masked:

for span in unpii.find_spans("Dr Martin au 06 12 34 56 78"):
    print(span)
# Span(start=3, end=9, category="NOM")
# Span(start=13, end=27, category="TELEPHONE")

DataFrame Integration

anonymize_dataframe anonymizes a column in a Polars DataFrame:

import polars as pl
import unpii

df = pl.DataFrame({"text": [
    "Dr Martin au 06 12 34 56 78",
    "Email: joe@chu-brest.fr",
    "Maladie de Parkinson",
]})

# Anonymize in place (overwrites the column)
df = unpii.anonymize_dataframe(df, "text")
# ┌─────────────────────────┐
# │ text                    │
# ╞═════════════════════════╡
# │ Dr <NOM> au <TELEPHONE> │
# │ Email: <EMAIL>          │
# │ Maladie de Parkinson    │  ← protected by whitelist
# └─────────────────────────┘

# Write to a new column
df = unpii.anonymize_dataframe(df, "text", new_column="text_anonymized")

# With options
df = unpii.anonymize_dataframe(df, "text", style="stars", mode="paranoid", ignore_groups=["TELEPHONE"])

Per-row words to mask (mask_from_columns)

Pass column names whose values are added as words to mask, per row. Useful when patient name/city are in structured columns:

df = pl.DataFrame({
    "text": ["bob est ici", "alice va bien"],
    "nom": ["bob", "alice"],
})

df = unpii.anonymize_dataframe(df, "text", mask_from_columns=["nom"])
# ┌─────────────────┬───────┐
# │ text            ┆ nom   │
# ╞═════════════════╪═══════╡
# │ <PII> est ici   ┆ bob   │
# │ <PII> va bien   ┆ alice │
# └─────────────────┴───────┘

Global words to mask (mask)

Words to mask on every row (e.g. the doctor who wrote all reports):

df = unpii.anonymize_dataframe(df, "text", mask=["Dupont", "Cabinet Santé Plus"])

Both combined

df = unpii.anonymize_dataframe(df, "text",
    mask_from_columns=["nom", "ville"],
    mask=["Dupont"],
    style="stars",
)

Low-level: anonymize_series

Operates on a Polars Series directly:

masked = unpii.anonymize_series(
    df["text"],
    mask_from_columns=[df["nom"], df["ville"]],
    mask=["Dupont"],
)
df = df.with_columns(masked.alias("text_anonymized"))

Batch processing: anonymize_batch

Operates on plain Python lists (no Polars dependency):

results = unpii.anonymize_batch(["Dr Martin ici", "Email: a@b.fr"])
# → ["Dr <NOM> ici", "Email: <EMAIL>"]

Threading

Control the number of threads used by anonymize_batch, anonymize_series, and anonymize_dataframe:

unpii.set_max_threads(4)     # Use 4 threads
unpii.get_max_threads()      # → 4
unpii.set_max_threads(0)     # Use all available cores (default)

Categories

Group	Placeholder	Standard	Paranoid
NOM	<NOM>	Titles + name, blacklist	UPPERCASE/Titlecase patterns, initials
TELEPHONE	<TELEPHONE>	French phone numbers	—
EMAIL	<EMAIL>	Valid emails	Anything with @
DATE	<DATE>	DD/MM/YYYY, literal months, ISO	—
BIRTHDATE	<BIRTHDATE>	né(e) le + date	—
ADRESSE	<ADRESSE>	Street number + type + name	—
CODE_POSTAL	<CODE_POSTAL>	5 digits + city name	—
NIR	<NIR>	French social security number	—
IBAN	<IBAN>	French IBAN	—
NUMBER	<NUMBER>	—	5+ consecutive digits
PII	<PII>	Custom words passed via mask=	—

Whitelist

Medical eponyms (Parkinson, Alzheimer, Verneuil...) are protected from masking by a global whitelist, regardless of which group detected them.

API Reference

# Single text
def anonymize(text, *, style="placeholder", mode="standard", ignore_groups=None, mask=None) -> str
def find_spans(text, *, mode="standard", ignore_groups=None, mask=None) -> list[Span]

# Batch (plain Python lists, no Polars needed)
def anonymize_batch(texts, *, mask_from_columns=None, mask=None, style="placeholder", mode="standard", ignore_groups=None) -> list[str | None]

# DataFrame (requires polars)
def anonymize_dataframe(df, column, *, mask_from_columns=None, mask=None, new_column=None, style="placeholder", mode="standard", ignore_groups=None) -> DataFrame
def anonymize_series(series, *, mask_from_columns=None, mask=None, style="placeholder", mode="standard", ignore_groups=None) -> Series

# Threading
def set_max_threads(n: int) -> None   # 0 = all cores (default)
def get_max_threads() -> int

# Span attributes: .start, .end, .category

Performance

Rust core with compiled regex and Aho-Corasick automata. All rules and dictionaries are embedded in the binary — zero I/O at runtime.

anonymize_batch, anonymize_series, and anonymize_dataframe use rayon for automatic parallelization across all cores.

License

MIT

See also

https://github.com/micropot/incognito https://github.com/microsoft/presidio