Feature engineering combined with machine learning and rule-based methods for structured information extraction from...

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Automated analysis of clinical narratives is crucial for extracting valuable information.
• Translating unstructured medical text into structured data remains a significant challenge.

Purpose of the Study:

• To develop and evaluate a system for translating narrative medical text into structured representations.
• The system addresses three key sub-tasks: concept extraction, assertion classification, and relation identification.

Main Methods:

• A five-step process including sentence pre-processing, noun and adjective phrase marking, concept extraction using dynamic Conditional Random Fields (CRF) models, assertion classification via ensemble voting, and relation identification using normalized sentences and discriminating features.
• Utilized a dosage-unit dictionary to enhance concept extraction accuracy.
• Employed a two-staged architecture for relation identification, incorporating pairwise classifiers.

Main Results:

• Achieved competitive performance against state-of-the-art systems.
• Demonstrated a micro-averaged F-measure of 0.8489 for concept extraction.
• Reported a micro-averaged F-measure of 0.9392 for assertion classification and 0.7326 for relation identification.

Conclusions:

• The developed system leverages common features and achieves state-of-the-art performance through effective feature engineering.
• Model switching in concept extraction significantly improved treatment concept identification, particularly for telegraphic sentences.
• Rule-based classifier features enhanced assertion classification for challenging classes like conditional and possible, mitigating data scarcity issues.