TLDR: FRACCO is a new, expert-annotated dataset of 1,301 synthetic French clinical oncology cases, translated from the Spanish CANTEMIST corpus. It provides detailed annotations for morphology, topography, and histologic differentiation using ICD-O-3.1, including unique composite expressions. This gold-standard corpus, with over 70,000 annotations, addresses the scarcity of French oncology NLP resources, enabling the development and evaluation of advanced NLP tools for cancer information extraction.
A new and significant resource for natural language processing (NLP) in French oncology has been introduced: FRACCO, the FRench Annotated Corpus for Clinical Oncology. This dataset is designed to help develop and evaluate NLP tools that can understand and extract crucial information from French clinical texts related to cancer.
The development of FRACCO addresses a critical gap in the field. While many annotated datasets exist for English-language medical reports, high-quality, manually annotated French-language corpora, especially those focused on oncology and using specific cancer ontologies, have been scarce. FRACCO fills this void by providing a comprehensive collection of synthetic French clinical cases, meticulously annotated with the International Classification of Diseases for Oncology (ICD-O) codes.
FRACCO originates from the Spanish CANTEMIST corpus, a resource for named entity recognition and normalization in oncology-related Spanish clinical texts. The CANTEMIST documents, which are de-identified, synthetic clinical case reports focusing on cancer diagnoses, were translated from Spanish to French as part of the FRASIMED project using machine translation. These translated texts formed the foundation for FRACCO.
The annotation process for FRACCO was rigorous and multi-layered. Initially, existing morphological annotations from the Spanish CANTEMIST corpus were projected onto the French translations. This projection was then extensively refined and extended. Expert annotators manually reviewed all documents to correct any errors introduced during translation, recover omitted oncology terms, and harmonize entity boundaries.
Crucially, FRACCO introduces two new categories of ICD-O entities: topographical codes, which describe the anatomical site of a tumor, and histology differentiation codes, which indicate tumor grading. Beyond these individual entity annotations, a novel “expression_CIM” layer was added. This layer captures composite expressions that combine multiple ICD-O elements (morphology, topography, and differentiation) into unified clinical concepts. For example, a complex phrase like “carcinome épidermoïde bien différencié du poumon” (well-differentiated squamous cell carcinoma of the lung) is normalized into a single, comprehensive entity. This allows for a deeper semantic analysis and supports advanced NLP tasks.
The dataset boasts a total of 71,127 ICD-O normalizations across 1,301 texts. It includes 399 unique morphology codes, 272 topography codes, and 2,043 unique composite expressions. The most common morphology codes include “Neoplasm, uncertain whether benign or malignant” and “Neoplasm, metastatic,” while frequently occurring topography codes feature “Lung, NOS” and “Liver.” Differentiation codes are categorized into four grades, from “Well differentiated, grade I” to “Undifferentiated, grade IV.”
To ensure semantic consistency, all entity annotations were normalized using ICD-O-3 terminology. This involved a semi-automated process combining dictionary-based matching with extensive manual review by a team of five annotators. Special attention was given to complex expressions and potential mistranslations from the initial machine translation, which were systematically identified and corrected to ensure the linguistic quality and representativeness of the French corpus.
The quality of FRACCO was further validated through structured procedures for both span annotation and ICD-O code normalization. Inter-annotator agreement scores were computed, showing strong agreement for most categories, and discrepancies were resolved through consensus among experts. The dataset’s usability for machine learning applications was demonstrated by fine-tuning several pre-trained named entity recognition (NER) models, which achieved high F1 scores, confirming the coherence of the annotation scheme and the learnability of the labels.
Also Read:
- Evaluating AI Models for Cancer Diagnosis Classification in Electronic Health Records
- AI Breakthrough: Instruction-Tuning LLMs Master German Cancer Coding
FRACCO is an invaluable reference standard for named entity recognition and concept normalization in French oncology texts. It supports the development and evaluation of NLP applications in oncology, particularly for French-language clinical data, and facilitates cross-lingual research by building upon resources like the Spanish CANTEMIST corpus. Researchers can access the dataset via the Zenodo repository and find associated Python scripts for data handling on GitHub. This work was a collaborative effort by researchers from the Service des sciences de l’information médicale and the Service d’oncologie de précision at Hôpitaux Universitaires de Genève, Suisse. You can find more details about this important work in the full research paper: FRACCO: A gold-standard annotated corpus of oncological entities with ICD-O-3.1 normalisation.
