A semi-supervised approach using label propagation to support citation screening.

Area of Science:

• Information Science
• Computer Science
• Biomedical Informatics

Background:

• Citation screening is crucial for systematic reviews but is labor-intensive.
• Current active learning methods require substantial manually labeled data for effective text classification.
• Existing approaches often fail to achieve robust performance without significant human input.

Purpose of the Study:

• To develop a semi-supervised method for early identification of relevant citations in systematic reviews.
• To improve classification performance in citation screening by reducing the need for manual labeling.
• To exploit pairwise similarities between labeled and unlabeled citations for enhanced accuracy.

Main Methods:

• A semi-supervised approach using label propagation based on citation similarity.
• Investigated two feature spaces: bag-of-words and spectral embedding.
• Augmented the training set by combining manually and automatically labeled citations.

Main Results:

• The semi-supervised method demonstrated statistically significant improvements in classification performance.
• Achieved superior results compared to state-of-the-art active learning methods.
• The approach proved effective across both clinical and public health systematic reviews.

Conclusions:

• Semi-supervised learning with label propagation offers an effective strategy to enhance citation screening efficiency.
• The proposed method reduces manual workload without compromising classification accuracy.
• This approach holds significant potential for streamlining systematic review processes in various scientific domains.