Exemplar-Based Medical Text Classification

The goal of this research is to develop improved techniques for fully automated andcomputer-assisted classification of medical text. A primary focus will be mechanisms that supporteasy-to-construct classifiers, thus enabling research on existing collections of free-text that are nowdifficult to analyze. This approach is exemplar-based: i.e., compare new text with a training set ofpreviously classified texts, and use the classifications of the closest retrieved texts to generatesuggested codes for the new text. Natural language extraction techniques will pre-process texts toassist the retrieval machinery. Phase I will conduct a series of experiments to test the approach, utilizingcoded radiology and pathology reports from Brigham and Women's Hospital and HCHP in Boston. PhaseII will develop a full software prototype and deploy it for on-site evaluation. Advanced retrievaltechniques and expert system back ends for alarming will also be explored in Phase II. The majortechnical innovation is a novel combination of document retrieval and natural language extractiontechnologies to permit easy construction of automated medical text classifiers. The major health-relatedcontribution is an enhanced ability to classify existing free-text records to permit statistical analysis forresearch and clinical quality-measurement initiatives.