A textual dataset of de-identified health records in Spanish and Catalan for medical entity recognition and anonymization

TL;DR

This paper introduces CARMEN-I, a Spanish and Catalan dataset of anonymized health records for training medical NLP systems in entity recognition and anonymization.

Contribution

CARMEN-I is a new multilingual clinical dataset with annotations for anonymization and information extraction.

Findings

CARMEN-I includes 2 years of anonymized clinical records from a Barcelona hospital during the pandemic.

The dataset covers multiple comorbidities and includes annotations for six clinical concept classes in a subset of 500 records.

The resource supports training and evaluation of anonymization and information extraction systems.

Abstract

The advancement of clinical natural language processing systems is crucial to exploit the wealth of textual data contained in medical records. Diverse data sources are required in different languages and from different sites to represent global health services. To this end, we have released CARMEN-I, a corpus of anonymized clinical records from the Hospital Clinic of Barcelona written during the COVID-19 pandemic spanning a period of two years. In addition to COVID-19 cases of adult patients, CARMEN-I features multiple comorbidities such as cardiovascular conditions, oncology treatments, post-transplant complications, and infectious diseases. This resource is publicly accessible together with detailed annotation guidelines and granular text-bound annotations generated in a collaborative effort between clinicians, linguists, and engineers to enable training and evaluation of automatic anonymization systems. Moreover, for information extraction purposes, a subset of 500 records is annotated with six relevant clinical concept classes: diseases, symptoms, procedures, medications, pathogens and humans.

Figures