Related Experiment Videos
Multiple hierarchical classification of free-text clinical guidelines.
Robert Moskovitch1, Shiva Cohen-Kashi, Uzi Dror
1Medical Informatics Research Center, Department of Information Systems Engineering, Ben Gurion University, P.O. Box 653, Beer Sheva 84105, Israel. robertmo@bgu.ac.il
Artificial Intelligence in Medicine
|May 30, 2006
Summary
Automating clinical guideline classification using supervised learning significantly reduces manual effort. This approach effectively categorizes documents within hierarchical structures, demonstrating feasibility even with limited training data.
Area of Science:
- Computer Science
- Medical Informatics
Background:
- Manual classification of medical documents is time-consuming.
- Automating this process is crucial for efficient medical information management.
Purpose of the Study:
- To develop a supervised learning approach for automating the classification of clinical guidelines into hierarchical categories.
- To leverage the hierarchical structure of concepts to overcome limited training examples.
Main Methods:
- Decomposed classification into a continuous decision process, utilizing similarity functions at each concept.
- Exploited hierarchical concept structure to enable multiple path selections, unlike traditional decision trees.
- Formulated conservative and aggressive stop-criterion strategies for navigating the concept hierarchy.
Main Results:
- Achieved variable precision ranging from 44% to 60% based on training method settings.
- Evaluated on a test collection of 1038 clinical practice guidelines (CPGs) across two hierarchies with approximately 5000 concepts.
- Each CPG was classified by an average of 10 concepts.
Conclusions:
- The supervised learning approach demonstrates feasibility for automated clinical guideline classification.
- Effectiveness is notable given the low ratio of guidelines to classification concepts in the evaluated dataset.