The European Clinical Case Corpus (E3C) project aims at collecting and annotating a large corpus of clinical documents in ve European languages (Spanish, Basque, English, French and Italian), which will be freely distributed. Annotations include temporal information, to allow temporal reasoning on chronologies, and information about clinical entities based on medical taxonomies, to be used for semantic reasoning.
The core of the corpus is a manually annotated dataset of clinical cases. A clinical case reports statements of a clinical practice, presenting the reason for a clinical visit, the description of physical exams, and the assessment of the patient's situation (Box 1).
open-calls/ A 25-year-old man with a history of Klippel-Trenaunay syndrome presented to the hospital with mucopurulent bloody stool and epigastric persistent colic pain for 2 wk. Colonoscopy showed continuous super cial ulcers and bleeding. Subsequent gastroscopy revealed mucosa with di use edema, ulcers, errhysis, and granular and friable changes in the stomach and duodenal bulb. A diagnosis of GDUC was considered. The patient hesitated about iv corticosteroids, so he was treated with pentasa 3.2 g/d. After 0.5 mo of treatment, the symptoms achieved complete remission. Follow-up examinations showed no evidence of recurrence for 26 mo.
Box 1: Sample clinical case. 2
The main motivation of the project is creating a clinical document corpus that can be freely redistributable and that contains temporal information and clinical entity annotations. The annotation of temporal informaCopyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). tion is the core e ort of the project, so we expect the E3C corpus to be useful in tasks such as event ordering and chronology generation.
These decisions are justi ed by three common issues on Natural Language Processing (NLP) for the clinical domain, listed below.
First, the E3C corpus is formed of already published documents, mostly clinical cases in journals so as to overcome the patient privacy issues that Electronic Health Records (EHR) often convey. We have opted for selecting documents that already allow redistribution to ensure the E3C corpus will be easily usable for the research community.
Secondly, as E3C is a multilingual
corpus, it helps reducing the gap between
English and other languages in terms of
available data for research. A large dataset of
clinical cases in Spanish is already available,
SPACCC
Thirdly, E3C is centered on temporal
information in clinical narratives, which has
not been often targeted by scholars. Most
of the attention has been focused on clinical
entity extraction and classi cation
The E3C corpus is a collection of both existing corpora (e.g., the SPACCC corpus) and published texts extracted from di erent sources, such as PubMed2 (journal abstracts), SciELO3 and the PanAfrican Medi
L2 162 113 171 168 174 Layer 1: about 25K tokens per language (around 80 documents) of clinical narratives with full manual annotation of clinical entities, temporal information and factuality, for benchmarking and linguistic analysis.
Layer 1 is the core of the E3C corpus and special attention has been awarded to creating a balanced document set in terms of size. Short (<200 tokens), medium (200{ 400 tokens) and long (400{600 tokens) texts have been selected, as we presumed that text length would directly a ect the temporal information in text; the longer the text, the more complex the temporal graph.
Layer 2: 50{100K tokens per language of clinical narratives with automatic annotation of clinical entities and manual check of the annotation of a small sample (about 10%). Layer 3: about 1M tokens per language of non-annotated medical documents (not necessarily clinical narratives) to be exploited by semi-supervised approaches.
All the layers are covered for Spanish, English, Italian and French. For Basque, Layer 2 (14K token) and 3 (600K token) are covered only partially. In Table 1 we summarize the distribution of documents per layer and language. In the case of L1, the amount of texts provides the information on how many distinct temporal graphs or chronologies we will be able to build from the dataset. 3.2
The nal E3C corpus will be available for download from the ELG platform repository5. All documents will be released under
Creative Commons licenses. This is possible as a large part of the texts in the corpus were already released under Creative Commons licenses and as permission for free distribution has been requested and obtained from the original owners of some of the documents. 4
The E3C corpus will contain two types of annotations: (i) temporal information and factuality, and (ii) annotation of clinical entities, speci cally disorders.
Manual annotation has been performed by a team of NLP students and researchers. More precisely, temporal information and clinical entity annotation guidelines have been de ned by two teams of three and four experts respectively. For the manual annotation e ort, eight people have been trained and are completing the annotation of temporal information, while clinical entity annotation is being conducted by ve people. 4.1
Temporal information annotation is performed following the THYME annotation guidelines (Styler et al., 2014a) with some minor adaptations (Magnini et al., 2020). This scheme provides tags for events, time expressions, temporal relations between events and/or time expressions, and aspectual relations between events. For each tag, a set of attribute-value pairs allow to make the relevant features explicit. In order to mark information that further contributes to the clinical history of a patient, we have added three categories: measurements and test results, actors (for the patient itself, health professionals and other participants), body parts.
In Figure 1 a simpli ed temporal information annotation is displayed. Events are in dark blue and their contextual modality (ACTUAL), document time relation (BEFORE) and polarity (POS) are highlighted. The 2 wk time expression (in gray) is classi ed as a duration and the information about the actor (the patient) is represented in light blue. 4.2
Clinical entity annotation focuses on disorders. Following UMLS6 a disorder is de ned as \a de nite pathologic process with a characteristic set of signs and symptoms".
In E3C, we mark disorders mentioned in
the text and assign them an UMLS concept
unique identi er (CUI). Disorder identi
cation and coding is performed following an
adaptation of the ShARe annotation
guidelines
The E3C project aims at reducing the lack of available resources for clinical NLP, gathering a large number of clinical narratives and focusing on languages other than English. After completing the project, the E3C corpus and the associated resources (baselines, scorers, etc.) will be available for research under a Creative Commons license, which will facilitate its acquisition and reusability. More speci cally, since the corpus contains information for temporal reasoning as well as clinical entity mentions, it will be useful for works on semantic interpretation of clinical texts. The fact that the corpus contains texts in ve languages will allow linguistic comparisons as well as experimentation on transfer learning. We also consider that the E3C corpus is a resource that could be employed in a series of evaluation challenges due to its atypical contents and types of annotations.
This work was partially supported by the European Language Grid project through its
open call for pilot projects (EU grant no. 825627), and by the Basque Government post-doctoral grant POS 2020 2 0026.