Table 3:

Tenfold cross-validation results of customized dictionary with added features in the training set (feature contribution)

Features	No. of features	PPV	Recall	F1 score
Setting 1: Baseline 6 + section parsing to discover mentions in relevant sections + section of relevant medications	109	0.759	0.806	0.782
Setting 2: Setting 1 + MTX signature	169	0.738	0.829	0.781
Setting 3: Setting 2 + DocTimeRel	409	0.740	0.868	0.798
Setting 4: Setting 2 + nearby words	4806	0.781	0.884	0.829
Setting 5: Setting 2 + nearby verbs’ part-of-speech tags	875	0.762	0.891	0.821
Setting 6: Setting 2 + nearby words + nearby verbs part-of-speech tags	5512	0.780	0.907	0.839
Setting 7: Setting 2 + nearby words + nearby verbs’ part-of-speech tags + DocTimeRel	5752	0.800	0.899	0.847
Setting 8: same feature settings as Setting 7 but patient-level classification	5752	0.814	0.727	0.768

For examples with the features, see figure 3

MTX, methotrexate; PPV, positive predictive value.