TL;DR
This paper introduces CAS, a new French corpus of clinical cases extracted from scientific literature, designed to address the scarcity of publicly available medical data for NLP research. The corpus is linguistically and semantically annotated, including concepts, negation, and uncertainty, to facilitate the development of robust NLP tools for the French medical domain.
What Problem Does This Paper Address?
Developing effective Natural Language Processing (NLP) tools for the medical domain is crucial for tasks like information retrieval, patient recruitment for clinical trials, and understanding medical discourse. However, a significant hurdle is the lack of readily accessible, representative textual data due to patient confidentiality and ethical concerns. While anonymization techniques exist, the risk of re-identification and the unique nature of medical histories make it challenging to freely share and utilize this data for research. This paper tackles this challenge by creating and releasing the CAS corpus, a French collection of clinical cases that aims to bridge this gap.
Methods
The CAS corpus is built from clinical cases published in French scientific literature and training materials. The process involves:
- Data Acquisition: Extracting text from PDF sources.
- Preprocessing: Automatic conversion to text format, followed by manual checking to correct segmentation errors, remove journal-specific paratext, and verify column/line/page conversions.
- Linguistic Annotation: The corpus is automatically PoS-tagged and lemmatized using an in-house tool.
- Semantic Annotation: Several layers of semantic annotation are applied automatically:
- Concept Identification: Mapping French terms to Concept Unique Identifiers (CUIs) from the UMLS. For multi-word terms, the IOB (Inside-Outside-Begin) format is used.
- Negation Detection: Identifying whether a disorder, procedure, or treatment is present or absent in a patient's history. This is achieved using a two-step process involving a trained CRF for negation markers and a neural network (Bi-LSTM with CRF) for scope detection.
- Uncertainty Detection: Identifying markers of uncertainty in the text. This involves manually building a set of simple and complex lexical markers, as well as morphological cues, which are then projected onto the corpus with heuristic rules for scope detection.
Datasets
The paper introduces the CAS corpus, a new collection of French clinical cases. It currently contains over 397,000 word occurrences (excluding punctuation) across 20,363 sentences. The corpus is annotated with linguistic (PoS-tagging, lemmas) and semantic (UMLS concepts, negation, uncertainty) information. The authors mention that similar corpora exist for English, such as the "Archive of Clinical Cases" (available on a dedicated website respecting the Creative Commons License), and that MIMIC-III is a large English dataset.
Key Insights and Ideas
- The Need for Domain-Specific Corpora: The paper strongly emphasizes that while general NLP tools are advancing, specialized corpora like CAS are essential for developing robust tools that can handle the nuances of medical language, especially in languages other than English.
- Challenges of De-identification: Even with de-identification tools, the risk of re-identification and the unique nature of medical data make it difficult to create truly "freely available" datasets, highlighting the importance of curated, ethically sourced corpora.
- Multi-layered Annotation for Richer Analysis: The inclusion of linguistic and multiple semantic annotation layers (concepts, negation, uncertainty) provides a rich resource for various NLP tasks, enabling deeper analysis of medical texts.
- Negation and Uncertainty as Crucial Elements: The paper highlights the importance of accurately detecting negation and uncertainty in medical discourse, as these elements significantly impact the interpretation of patient information and clinical findings.
Key Results
- The CAS corpus contains over 397,000 word occurrences and 20,363 sentences.
- The semantic annotations include:
- 47,708 UMLS concept annotations.
- 4,723 uncertainty annotations.
- 4,620 negation annotations.
- The negation detection system achieves up to 0.98 F-measure for cues and 0.86 for their scope on French and English data.
- Uncertainty detection yields about 0.90 F-measure for cues and 0.80 for their scope.
Limitations and Future Work
- Limitations:
- The current corpus is derived from scientific literature, which may differ in style and content from direct clinical notes.
- While efforts are made to correct conversion errors, the automatic conversion from PDF to text can still introduce artifacts.
- The current version of the corpus prioritizes longer, more precise CUIs when multiple are present, which might lead to the exclusion of potentially relevant shorter terms.
- Future Work:
- Enriching the corpus with more clinical cases.
- Adding more annotation layers and cross-validating them with human annotators.
- Providing a more detailed description of the corpus, including statistics on patient demographics, diseases, and publication sources.
- Building similar corpora for other languages.
- Making these annotated corpora freely available to encourage NLP tool development for medical free-text documents.
Summary updated