Identifying key elements for evidence-base medicine using pretrained model and graph convolution network

Abstract

Objective: Evidence-based medicine (EBM) provides a framework to support clinicians' decision-making processes using the best evidence currently available in the field. The key elements of clinical research can be defined by a framework called PICO, which identifies the sentences in a medical literature text that belong to the four key elements reported in clinical trials: Participants/Problem (P), Intervention (I), Comparison (C) and Outcome (O). This study aims to establish an effective detection method for key elements of evidence-based medical research.Methods: Based on the text features of key elements framework, we propose a deep learning model BERTGCN which combined large-scale pretraining model and graph convolution network (GCN) for detecting key elements of evidence-based medicine. In this model, the sentences which are initialized with pre-trained BERT representations and the words in the EBM evidence were recognized as the sentence nodes and word of the graph which were used to train GCN model. At the same time, the sentences were used to fine tune the pretraining model.Results: We tested our proposed approach over PubMed-PICO dataset is a data set containing tens of thousands of EBM key elements extracted from PubMed. The F1-score of P/I/O in the model we proposed reached 91.3%, 85.8% and 90.0% respectively. Experimental results show that our model outperforms the current optimal model.Conclusions: BERTGCN is able to leverage the advantages of both worlds: large-scale pretraining and transudative learning to improve the efficiency of detecting evidence-based medical evidence from research publications.