An open source corpus and automatic tool for section identification in Spanish health records

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Electronic Clinical Narratives (ECNs) contain vital health information but lack sufficient open-source data.
• The structural heterogeneity of ECNs, from structured headings to unstructured notes, hinders automatic system development and evaluation.
• Developing automated systems for ECNs is crucial for extracting and utilizing patient data effectively.

Purpose of the Study:

• To provide a Spanish open-source dataset for developing and evaluating automatic section identification systems for ECNs.
• To design and implement a novel evaluation metric (B2) tailored for clinical narrative section identification.
• To create a fine-tuned language model optimized for this specific task.

Main Methods:

• Annotation of a corpus of Spanish clinical progress notes into seven major section types.
• Assessment of existing metrics and definition of a new B2 metric for improved task-specific evaluation.
• Development and release of a baseline language model fine-tuned for section identification.

Main Results:

• The open-source annotated corpus, evaluation script, and baseline model are publicly available.
• The baseline model achieved an average B2 score of 71.3 on the open-source dataset.
• The model demonstrated competence in data scarcity scenarios, achieving an average B2 of 67.0.

Conclusions:

• Automatic section identification in unstructured clinical narratives is feasible with adequate data and evaluation metrics.
• This contribution aims to accelerate the development of more robust and accurate systems for processing ECNs.
• The open-sourced resources will foster further research and innovation in clinical text analysis.