Adverse-Effect Relations Extraction from Massive Clinical Records

The rapid spread of electronic health records raised an interest to large-scale information extraction from clinical texts. Considering such a background, we are developing a method that can extract adverse drug event and effect (adverse–effect) relations from massive clinical records. Adverse–effect relations share some features with relations proposed in previous relation extraction studies, but they also have unique characteristics. Adverse–effect relations are usually uncertain. Not even medical experts can usually determine whether a symptom that arises after a medication represents an adverse– effect relation or not. We propose a method to extract adverse–effect relations using a machine-learning technique with dependency features. We performed experiments to extract adverse–effect relations from 2,577 clinical texts, and obtained F1-score of 37.54 with an optimal parameters and F1-score of 34.90 with automatically tuned parameters. The results also show that dependency features increase the extraction F1-score by 3.59.