BLINK_LSTM: BioLinkBERT and LSTM based approach for extraction of PICO frame from Clinical Trial Text

Rapid growth in publication of clinical trial reports has made it extremely challenging to conduct systematic reviews. Automatic extraction of Population, Intervention, Comparator, and Outcome (PICO) from clinical trial reports can alleviate the traditionally adopted time-consuming process of systematic reviews. In this paper, we propose a novel approach for automatically detecting the PICO-related terms from clinical trial reports. Our techniques use BioLinkBERT as our base model with two Bi-LSTM layers. Our proposed model, when trained on either coarse-grained datasets like EBM-NLP, EBM-COMET, and RedHOT, or fine-grained datasets such as EBM-NLPrev and EBM-NLPh, irrespective of the nature of the dataset, achieved state-of-the-art (SOTA) results. We have also shown with our framework that ensembling technique can be highly effective for the token span prediction task and empirical findings demonstrate that our ensemble approach, when applied on EBM-NLP dataset, achieves new state-of-the-art.