The effect of feature representation on MEDLINE document classification.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Accurate medical text classification is crucial for information retrieval and knowledge discovery.
• Traditional methods often rely on basic word representations, potentially missing nuanced medical concepts.
• Advanced text representation techniques are needed to capture the complexity of medical language.

Purpose of the Study:

• To evaluate the impact of different text representation techniques on medical text classification performance.
• To compare the effectiveness of bag-of-words, bag-of-phrases, and a hybrid approach.
• To develop a system capable of extracting medical phrases using knowledge bases and NLP.

Main Methods:

• Developed a text classification system supporting bag-of-words, bag-of-phrases, and hybrid representations.
• Utilized a medical knowledge base and natural language processing to extract medical phrases.
• Conducted experiments on the OHSUMED dataset (MEDLINE documents) to assess classification performance.
• Measured performance using information retrieval metrics: precision, recall, and F1-score.

Main Results:

• The hybrid approach achieved the highest classification performance (F1-score=0.60).
• Hybrid representation yielded better results (p=0.87, r=0.46, F1=0.60) compared to bag-of-words (p=0.85, r=0.44, F1=0.58).
• Bag-of-phrases alone showed comparable precision but lower recall (p=0.87, r=0.42, F1=0.57) than the hybrid method.

Conclusions:

• A hybrid text representation combining word and phrase information offers superior performance for medical text classification.
• Integrating medical phrase extraction enhances classification accuracy over basic word-level analysis.
• The findings support the use of sophisticated text representation for improving medical information systems.