Negation recognition in clinical natural language processing using a combination of the NegEx algorithm and a...

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Electronic Health Records (EHRs) contain crucial patient information in unstructured free text.
• Recognizing negation modifiers in clinical text is a significant challenge for clinical Natural Language Processing (cNLP).
• Effective negation recognition solutions for languages other than English, particularly Spanish, are scarce.

Purpose of the Study:

• To develop a robust negation recognition solution for Spanish EHRs.
• To combine a customized rule-based NegEx layer with a convolutional neural network (CNN) for improved accuracy.
• To enhance the reliability of clinical Named Entity (cNE) extraction from EHRs.

Main Methods:

• A binary classification approach ('affirmative' vs. 'non-affirmative') was adopted for negation recognition.
• A rule-based NegEx layer was customized using Spanish corpus rules and custom additions.
• A CNN binary classifier was trained on EHRs annotated by medical doctors for cNEs and negation markers.

Main Results:

• The pipeline achieved high performance metrics: 0.93 precision, 0.94 recall, and 0.94 F1-score for the 'affirmative' class.
• For the 'non-affirmative' class, the pipeline obtained 0.86 precision, 0.84 recall, and 0.85 F1-score.
• Consistent performance was observed on a separate data source and state-of-the-art results were achieved on public Spanish negation corpora.

Conclusions:

• The combined rule-based NegEx layer and CNN approach effectively addresses negation recognition challenges in Spanish EHRs.
• This methodology significantly improves the precision of cNE retrieval from clinical free-text.
• Accurate negation recognition is vital for reducing false positives and increasing the credibility of cNLP systems in real-world evidence studies.