Performance of classifier on hold-out test set using different feature sets.

We performed feature ablation experiments to assess the contribution of different feature sets to the performance of the classifier for detecting used-to-treat relationships. The first column indicates the features used to train and test the classifiers. Classifier performance was evaluated in a hold out test set of 1,749 positive and 7,035 negative examples of drug usage after training in a set of 7,112 positive and 27,938 negative examples. The first row shows performance using STRIDE derived features in which co-mentions are counted without regard to present known indications in the clinical record.