Automatic negation detection in narrative pathology reports.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Pathology reports contain crucial negations that impact clinical decision-making.
• Accurate detection of medical negations is vital for automated analysis of clinical text.

Purpose of the Study:

• To evaluate and compare the performance of lexicon-based, syntax-based, and machine-learning-based approaches for medical entity negation detection.
• To identify the most effective method for identifying negations in free-text pathology reports.

Main Methods:

• Three negation detection methods were applied: lexicon-based (trigger terms, termination clues), syntax-based (Stanford parser dependency), and machine-learning-based (Support Vector Machine).
• A dataset of 284 English lymphoma pathology reports was used for training and testing.

Main Results:

• The machine-learning-based approach achieved the highest performance with a micro-averaged F-score of 82.56%.
• The syntax-based approach yielded the lowest F-score (78.62%), primarily due to parsing errors.
• The lexicon-based approach showed strong precision (89.74%) and recall (76.09%), outperforming a similar existing tool.

Conclusions:

• Machine learning offers significant advantages for medical negation detection in pathology reports.
• Tailoring methods, such as customizing lexicon-based approaches to specific corpora, enhances performance.
• Future improvements may involve applying different detection strategies to distinct sections of pathology reports.