Sentence Extraction using Outcome Prediction Model Trained from Clinical Data

Abstract

This study aims to extract clinically important sentences from accumulated medical documents to assist medical workers to search documents. Unsupervised document summarization methods such as LexRank are commonly used in situations where it is difficult to prepare training data. However, these methods are based on a hypothesis that important topics are frequently referred which does not match the medical document. Many previous studies have predicted the length of hospital stay and mortality using clinical data, and we propose these outcomes can be distant labels of clinical importance. Namely, an output from the outcome prediction model becomes high when an input sentence is clinically important. Therefore, in this study, we propose a model to extract clinically important sentences using an outcome prediction model. Experimental results show our text extraction model with an outcome prediction model can summarize more accurately than the conventional models.