Exploiting External Data for Training a Cancer Clause Classifier

Automatically detecting cancer-revealing clauses in a medical report can be helpful for medical experts in various tasks such as cancer staging. While the problem of detecting such clauses can be formulated as classifying individual clauses into cancer and non-cancer categories, standard classification algorithms suffer from the fact that the training data contains much more non-cancer clauses than cancer clauses in radiology reports. In order to alleviate the data imbalance and sparseness problems related to the radiology reports, we attempt to use cancer-related external data in term weighting at the training stage. Our experiment shows that this approach indeed changes term feature statistics and improve effectiveness of the classifier.