Automatic literature screening using the PAJO deep-learning model for clinical practice guidelines.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Bibliometrics

Background:

• Clinical practice guidelines (CPGs) rely on high-quality research, but manual literature screening is inefficient.
• Existing artificial intelligence (AI) methods lack effective literature screening capabilities.
• Developing automated literature screening is crucial for advancing CPG development.

Purpose of the Study:

• To propose the Paper title, Abstract, and Journal (PAJO) model, an AI-based system for automatic literature screening.
• To address the limitations of manual and existing AI-based literature screening methods.
• To enhance the efficiency and accuracy of identifying relevant research for CPGs.

Main Methods:

• The PAJO model frames article screening as a classification task, using articles in CPGs as positive samples.
• It leverages features from article titles, abstracts, and journal characteristics via a pretrained bidirectional-encoder-representations-from-transformers (BERT) model.
• An attention mechanism integrates text and journal encoders for final classification.

Main Results:

• The PAJO model was trained and evaluated on a dataset of 89,940 neck pain articles from PubMed.
• It outperformed state-of-the-art baselines by 1.91% in F1 score and 2.25% in area under the ROC curve.
• Expert evaluation confirmed PAJO's ability to successfully screen high-quality articles.

Conclusions:

• The PAJO model offers an effective solution for automated literature screening.
• It can significantly improve the efficiency of developing CPGs, starting with neck pain research.
• The PAJO methodology is adaptable for literature screening in other disease areas.