Named entity recognition for de-identifying Spanish electronic health records

Area of Science:

• Natural Language Processing
• Machine Learning
• Health Informatics

Background:

• Electronic Health Records (EHRs) are valuable for clinical decision-making but require de-identification for data sharing.
• Prior research on EHR de-identification primarily used English data due to limited Spanish resources.
• Automatic de-identification of Spanish EHRs is essential for broader data dissemination and research.

Purpose of the Study:

• To explore automatic de-identification of medical documents in Spanish.
• To develop and compare deep learning models for this task.
• To create a practical tool for de-identifying Spanish clinical notes.

Main Methods:

• A private corpus of 599 Spanish clinical cases was annotated for protected health information.
• Named Entity Recognition (NER) was used as the core prediction task.
• Two deep learning approaches were implemented: Recurrent Neural Networks (RNNs) and Transformers (specifically XLM-RoBERTa).
• A text expansion procedure was used to augment training data.

Main Results:

• Transformer models significantly outperformed RNNs in de-identifying Spanish clinical data.
• The XLM-RoBERTa Transformer achieved high performance metrics (Precision: 0.946, Recall: 0.954, F1: 0.95) on the expanded corpus.
• A functional web application was developed to assist clinicians in EHR de-identification.

Conclusions:

• State-of-the-art Transformer models are practically applicable for precise de-identification of Spanish clinical notes.
• These models can enhance the utility of EHRs in real-world medical settings.
• Further performance improvements may be achievable with continual pre-training strategies.