Enhancing Abstract Screening Classification in Evidence-Based Medicine: Incorporating Domain Knowledge into Pre-trained Models

Evidence-based medicine (EBM) represents a cornerstone in medical research, guiding policy and decision-making. However, the robust steps involved in EBM, particularly in the abstract screening stage, present significant challenges to researchers. Numerous attempts to automate this stage with pre-trained language models (PLMs) are often hindered by domain-specificity, particularly in EBMs involving animals and humans. Thus, this research introduces a state-of-the-art (SOTA) transfer learning approach to enhance abstract screening by incorporating domain knowledge into PLMs without altering their base weights. This is achieved by integrating small neural networks, referred to as knowledge layers, within the PLM architecture. These knowledge layers are trained on key domain knowledge pertinent to EBM, PICO entities, PubmedQA, and the BioASQ 7B biomedical Q &A benchmark datasets. Furthermore, the study explores a fusion method to combine these trained knowledge layers, thereby leveraging multiple domain knowledge sources. Evaluation of the proposed method on four highly imbalanced EBM abstract screening datasets demonstrates its effectiveness in accelerating the screening process and surpassing the performance of strong baseline PLMs.