Enriching the Semantics of Temporal Relations for Temporal Pattern Mining.

This paper proposes a method to enrich the semantics of before/after relations based on the closeness between two events and contexts surrounding two events identified by a key event in a period. The proposed method captures four types of before/after relations: continuous relation, discrete relation, same contextual relation and different contextual relation. We derive five embedding methods from the combination of the four relation types, and apply them to clinical data for the prediction of hospital length of stay where the events are treatments, the key event is surgery and the period is seven days after hospital admission. The experimental results showed that on the whole the embedding method employed all of the four relation types has higher scores of precision, recall, F1 score and accuracy than other embedding methods. This suggests that the potential for the elucidation of the candidates of medically meaningful temporal patterns increases by exploring the temporal patterns generated from the embedding method. A paired t-test indicated that significant differences are partially confirmed for discrete relation and same contextual relation but not confirmed for different contextual relation. We will apply the proposed method to a number of hospitals for performing the further analysis of the four relation types and elucidating medically meaningful temporal patterns.