Disease Prediction and Topic Phrase Extraction from Clinical Reports by Attention-based LSTM model

Abstract

none

In this thesis, we focus on how to predict a certain disease from a given pathology report without the pathologist's diagnosis paragraph. Moreover, we aim to identify relevant diagnostic features within reports' paragraphs and get the determined clinical phrases that serve as clinical interpretations for the prediction model. We use the attention-based LSTM model for binary prediction of a given disease. Next, the attention weights learned from the model are extracted to generate attention terms. These attention terms are grouped under different MeSH terms defined by the United States National Library of Medicine. Moreover, the topic phrases are generated by using the frequency pattern method as representations of each group. The extracted topic phrases could provide as the determined clinical interpretation for the prediction.