Using Knowledge Sources to Improve Classification of Medical Text Reports

Domain knowledge has been shown to be an important component of machine learning. However, the cost of obtaining domain knowledge to improve classifier generation can exceed the cost of manually creating classifiers. An alternative approach is to use existing knowledge sources to collect relevant domain knowledge, and improve machine learning. We investigated the use of two existing knowledge sources (a natural language processor and controlled vocabulary metathesaurus) to improve machine learning algorithm performance in building classifiers for medical text reports. Both knowledge sources were found to significantly improve classifier performance. This demonstrates that existing knowledge sources can easily be used to improve machine learning performance.