Comparing Different Methods for Named Entity Recognition in Portuguese Neurology Text.

Area of Science:

• Natural Language Processing (NLP)
• Computational Linguistics
• Medical Informatics

Background:

• Electronic Medical Records (EMRs) contain unstructured, natural language data.
• Information Extraction (IE) and Named Entity Recognition (NER) are crucial for knowledge acquisition from EMRs.
• Limited research exists on NER in Portuguese medical texts.

Purpose of the Study:

• To evaluate and compare different Named Entity Recognition (NER) models for Portuguese neurological texts.
• To assess the impact of word embeddings (WEs) trained on clinical versus general language data.
• To determine the feasibility of extracting information from hospital EMRs using models trained on publicly available medical journal data.

Main Methods:

• Comparison of Conditional Random Fields (CRF), BiLSTM-CRF, and residual BiLSTM-CRF models.
• Utilized Portuguese medical journal texts and Coimbra Hospital and Universitary Centre (CHUC) Neurology Service EMRs.
• Evaluated performance using word embeddings trained on in-domain (clinical) and out-of-domain (general) text corpora.

Main Results:

• Deep learning models (BiLSTM-CRF variants) outperformed shallow learning models (CRF).
• Models achieved F1-Scores up to 83% (relaxed) and 75% (strict) on medical journal texts.
• Models achieved F1-Scores up to 71% (relaxed) and 62% (strict) on hospital EMRs.
• In-domain WEs yielded superior results compared to general-domain WEs.

Conclusions:

• Deep learning approaches are effective for NER in Portuguese neurological EMRs.
• Training word embeddings on domain-specific clinical text significantly improves NER performance.
• Information extraction from hospital EMRs is feasible using models trained on journal data, but requires expert validation.