Neural negated entity recognition in Spanish electronic health records

Area of Science:

• Computational linguistics
• Medical informatics
• Artificial intelligence

Background:

• Negation detection is crucial for clinical decision support systems.
• Accurate identification of negated medical events (e.g., absence of disorders, allergies) is vital for patient care.
• Negation can alter the meaning of clinical text, impacting tasks like drug adjustment and disease detection.

Purpose of the Study:

• To develop a novel approach for negation detection in clinical texts, specifically focusing on negated medical entities.
• To address the challenge of lexical variability in Spanish Electronic Health Records (EHRs).
• To improve the performance of Named Entity Recognition (NER) for clinical text by specifically identifying negated entities.

Main Methods:

• A deep learning approach combining character embeddings, Long Short-Term Memory (LSTM) networks, and Conditional Random Fields (CRF).
• Exploration of word embeddings versus lemmatized embeddings to handle lexical variability and out-of-vocabulary (OOV) words.
• Application of the system to Spanish EHRs for negated medical entity recognition.

Main Results:

• The deep learning model effectively identified negated medical entities.
• Lemmatized embeddings significantly improved the model's performance compared to word embeddings.
• Achieved an f-measure of 65.1 for exact-match and 82.4 for partial-match.

Conclusions:

• The proposed deep learning method, particularly with lemmatized embeddings, is effective for negation detection in Spanish clinical texts.
• This approach enhances the ability of LSTMs to manage high lexical variability in EHRs.
• Accurate negation detection of medical entities can significantly benefit clinical decision support and AI applications in healthcare.