💬 NLP

PARHAF, a human-authored corpus of clinical reports for fictitious patients in French

The paper introduces PARHAF, a large open-source corpus of 7,394 expert-authored, fictitious French clinical reports covering over 5,000 patient cases across 18 specialties, designed to enable privacy-preserving training and evaluation of clinical natural language processing systems while adhering to strict data protection regulations.

Original authors: Xavier Tannier, Salam Abbara, Rémi Flicoteaux, Youness Khalil, Aurélie Névéol, Pierre Zweigenbaum, Emmanuel Bacry

Published 2026-03-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xavier Tannier, Salam Abbara, Rémi Flicoteaux, Youness Khalil, Aurélie Névéol, Pierre Zweigenbaum, Emmanuel Bacry

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to teach a robot how to be a doctor. To do this, you need to show it millions of real patient stories so it can learn how diseases look, how treatments work, and how doctors write their notes.

But here's the problem: Real patient stories are top-secret. They contain private names, addresses, and medical histories. In Europe (and especially France), the rules are so strict that you can't just hand these stories to a robot or share them with researchers. It's like trying to teach someone to drive using a car that belongs to a stranger; you can't just let them take it for a spin.

For years, this has been a huge roadblock. Researchers in France were stuck, unable to build smart medical AI because they didn't have any "practice" data to use.

Enter PARHAF: The "Fake" Medical Library.

The authors of this paper came up with a brilliant solution. Instead of stealing real secrets, they decided to write brand new, completely fake patient stories.

Think of it like a massive, collaborative writing workshop.

How They Did It (The Recipe)

The Writers: They recruited 104 medical residents (doctors in training) from all over France. These weren't just random people; they were experts in everything from heart surgery to infectious diseases.
The Assignment: The researchers gave these doctors a set of "recipes" or scenarios. For example: "Write a story about a 45-year-old man who came to the ER with a broken leg, had surgery, and went home three days later."
The Twist: The doctors had to write these stories as if they were real, using the same professional language and structure they use every day. But the patients? They never existed. The names, the faces, the specific details—all made up.
The Safety Net: To make sure the stories weren't just random guesses, the team used real French government health statistics (like a giant census of hospital visits) to decide which stories to write. If 10% of real hospital visits are for pneumonia, they made sure 10% of their fake stories were about pneumonia. This ensures the "fake" library looks and feels exactly like a "real" one.

What's in the Box?

The result is PARHAF, a giant digital library containing:

7,394 medical reports (like discharge summaries, surgery notes, and lab results).
5,009 unique "fake" patients.
Stories covering 18 different medical specialties, from oncology (cancer) to obstetrics (birth).

It's like a simulator for doctors and AI. Just as a flight simulator lets a pilot practice crashing a plane without anyone getting hurt, PARHAF lets AI practice diagnosing diseases without violating anyone's privacy.

Why is this a Big Deal?

Privacy is Guaranteed: Since the patients never existed, there is zero risk of accidentally revealing a real person's secret. You can share this data with anyone, anywhere, without fear.
It's Open Source: Unlike many medical databases that are locked behind expensive passwords, this one is free to use (like a public park).
It's Realistic: Because real doctors wrote it, the language is authentic. It's not a robot guessing what a doctor might say; it's actual doctors speaking their native language.

The "Embargo" (The Secret Sauce)

The authors are so careful that they are holding back a small portion of the stories (about 1,200 documents) in a "time capsule." They won't release these yet. Why? To make sure that in the future, researchers can test new AI models on these stories without the AI having already seen them online. This ensures that when we test if a new AI is smart, it's actually learning, not just memorizing answers it found on the internet.

The Bottom Line

PARHAF is a privacy-safe playground for medical AI. It solves the "chicken and egg" problem: we need data to build better AI, but we can't share real data. By creating a massive library of high-quality, realistic, but entirely fictional medical stories, the team has given researchers the fuel they need to build smarter, safer, and more helpful medical tools for the future.

In short: They built a fake hospital to teach real robots how to be doctors, without ever risking a single patient's privacy.

1. Problem Statement

The development of Clinical Natural Language Processing (NLP) systems in France and the European Union is severely hindered by data privacy regulations.

The Bottleneck: Real electronic health records (EHRs) contain sensitive patient data. Strict privacy laws (e.g., GDPR) prevent the free sharing of these records, creating a scarcity of open, high-quality French clinical corpora.
Limitations of Existing Solutions:
- De-identification: Pseudonymizing real data (removing direct identifiers) does not guarantee anonymity against re-identification, limiting free distribution.
- Machine Translation: Translating English clinical datasets fails to capture cultural nuances and specific healthcare system structures of the French context.
- Synthetic Text (AI): Generative models can produce synthetic text, but the balance between privacy and medical utility requires further validation, and models trained on real data may memorize patient information.
- Existing Proxies: Case reports differ significantly in style from hospital EHRs, and textbook-based corpora often face copyright restrictions.

Core Challenge: How to create a large, realistic, and fully privacy-preserving corpus of French clinical reports that is freely shareable for research and model training.

2. Methodology

The authors developed PARHAF using a novel protocol that combines human expertise with epidemiological data to generate synthetic but realistic clinical scenarios.

A. Data Generation Protocol

Authors: 104 medical residents across 18 medical and surgical specialties were recruited via national medical associations.
Task: Authors were instructed to write clinical reports for entirely fictitious patients based on predefined scenarios.
Scenario Design:
- Epidemiological Basis: Scenarios were derived from the French National Health Data System (SNDS), utilizing Diagnosis-Related Groups (DRGs), ICD-10 codes, age, sex, and length of stay.
- Sampling Strategy: To ensure representativeness while avoiding the over-representation of common conditions, the authors applied a square-root transformation to empirical frequencies ( $\sqrt{f_i}$ ) with a probability cap ( $p_{max}$ ). This ensures diversity without sacrificing statistical realism.
- Privacy by Design: Since the patients are fictitious and the data is generated from scratch, the corpus is anonymous by definition, requiring no de-identification.
Document Types: The corpus covers various hospitalization types with specific templates:
- Medical: Discharge summaries.
- Surgical: Pre-op consultations, operative reports, and discharge summaries.
- Obstetrics: Delivery and postpartum reports.
- Oncology: Pathology reports.
Quality Control:
- Human Review: Each report was written by one resident and reviewed by a second expert.
- Validation: A structured validation process resulted in a 3.6% rejection rate.
- Automated Checks: An automated pipeline verified structural compliance (document types, counts, ICD-10 codes).
- AI Restriction: The use of Generative AI tools was explicitly discouraged to prevent stylistic bias and ensure human-authored authenticity.

B. Corpus Structure

The corpus is organized into two main components:

General-Purpose Component (Core): ~3,900 patients designed to approximate real-world hospitalization distributions across diverse specialties.
Specialized Subsets: Four targeted subsets for specific Information Extraction (IE) tasks:
- ICD-10 Coding: Surgical reports (digestive, orthopedic, etc.).
- Oncology Biomarkers: Pathology reports identifying tumor mutations and protein expression.
- Oncology Treatment Response: Reports classifying RECIST-style outcomes (e.g., complete response, progressive disease).
- Infectiology: Reports detailing bacteremia, causative bacteria, and infection sites.

3. Key Contributions

PARHAF Corpus: A large-scale, open-source dataset containing 7,394 clinical reports describing 5,009 fictitious patient cases.
Privacy-Preserving Design: The first large French clinical corpus that is freely shareable under an open license (CC-BY 4.0) because it contains no real patient data.
Methodological Framework: A replicable protocol for building synthetic clinical corpora that balances medical realism, statistical representativeness (via SNDS data), and privacy.
Specialized Subsets: The inclusion of curated subsets for high-value NLP tasks (coding, biomarker extraction, treatment response classification) in oncology and infectious diseases.
Resource Availability:
- Open Release: 6,185 documents (4,254 patients) are immediately available.
- Embargoed Portion: 1,209 documents (755 patients) are temporarily withheld to allow for future benchmarking under controlled conditions, preventing Large Language Model (LLM) contamination.
- Formats: Available as JSON with metadata and raw text files, plus a Hugging Face Parquet/Arrow dataset.

4. Results and Statistics

Scale: 7,394 reports, 5,009 patients, covering 18 specialties.
Demographics: The population pyramid and specialty distribution reflect the diversity of French hospital activity, though adjusted for the specific constraints of the author pool.
Document Types: Includes discharge summaries, operative reports, pathology reports, and consultation notes.
Cost & Efficiency: The project cost approximately €495,000, with an estimated 60 minutes per document (45 min writing + 15 min review).
Validation: 3.6% of documents were rejected during human review; an additional 120 underwent minor technical corrections (formatting/typos) without altering clinical content.

5. Significance and Impact

Enabling French Clinical NLP: PARHAF fills a critical gap by providing a resource that allows researchers to train and evaluate French clinical language models without navigating complex legal barriers associated with real patient data.
Benchmarking: It establishes a standard for evaluating models on tasks like diagnostic coding, entity extraction, and summarization in a privacy-safe environment.
Educational Value: The corpus can be used for medical education to train students in structured report writing and diagnostic reasoning using realistic scenarios.
Scalability: The methodology serves as a blueprint for creating similar synthetic corpora in other languages and healthcare systems, potentially democratizing access to clinical data for NLP research globally.
Limitations: The authors explicitly state the corpus is not for clinical decision-making, regulatory approval, or epidemiological inference, as the distributions are synthetic and do not reflect real-world prevalence or operational constraints (e.g., time pressure).

In summary, PARHAF represents a paradigm shift in clinical data sharing, moving from "restricted access to real data" to "open access to high-fidelity synthetic data," thereby accelerating the development of safe and effective AI in French healthcare.