Evaluating Methods for Identifying Cancer in Free-Text Pathology Reports Using Various Machine Learning and Data Preprocessing Approaches

Abstract

Automated detection methods can address delays and incompleteness in cancer case reporting. Existing automated efforts are largely dependent on complex dictionaries and coded data. Using a gold standard of manually reviewed pathology reports, we evaluated the performance of alternative input formats and decision models on a convenience sample of free-text pathology reports. Results showed that the input format significantly impacted performance, and specific algorithms yielded better results for presicion, recall and accuracy. We conclude that our approach is sufficiently accurate for practical purposes and represents a generalized process.