Abstract
The automatic detection of temporal relations between events in electronic medical records has the potential to greatly augment the value of such records for understanding disease progression and patients' responses to treatments. We present a three-step methodology for labeling temporal relations using machine learning and deterministic rules over an annotated corpus provided by the 2012 i2b2 Shared Challenge. We first create an expanded training network of relations by computing the transitive closure over the annotated data; we then apply hand-written rules and machine learning with a feature set that casts a wide net across potentially relevant lexical and syntactic information; finally, we employ a voting mechanism to resolve global contradictions between the local predictions made by the learned classifier. Results over the testing data illustrate the contributions of initial prediction and conflict resolution.