Development and Validation of a Literature Screening Tool: Few-Shot Learning Approach in Systematic Reviews

Systematic reviews (SRs) are considered the highest level of evidence, but their rigorous literature screening process can be time-consuming and resource-intensive. This is particularly challenging given the rapid pace of medical advancements, which can quickly make SRs outdated. Few-shot learning (FSL), a machine learning approach that learns effectively from limited data, offers a potential solution to streamline this process. Sentence-bidirectional encoder representations from transformers (S-BERT) are particularly promising for identifying relevant studies with fewer examples.

This study aimed to develop a model framework using FSL to efficiently screen and select relevant studies for inclusion in SRs, aiming to reduce workload while maintaining high recall rates.

We developed and validated the FSL model framework using 9 previously published SR projects (2016-2018). The framework used S-BERT with titles and abstracts as input data. Key evaluation metrics, including workload reduction, cosine similarity score, and the number needed to screen at 100% recall, were estimated to determine the optimal number of eligible studies for model training. A prospective evaluation phase involving 4 ongoing SRs was then conducted. Study selection by FSL and a secondary reviewer were compared with the principal reviewer (considered the gold standard) to estimate the false negative rate.

Model development suggested an optimal range of 4-12 eligible studies for FSL training. Using 4-6 eligible studies during model development resulted in similarity thresholds for 100% recall, ranging from 0.432 to 0.636, corresponding to a workload reduction of 51.11% (95% CI 46.36-55.86) to 97.67% (95% CI 96.76-98.58). The prospective evaluation of 4 SRs aimed for a 50% workload reduction, yielding numbers needed to screen 497 to 1035 out of 995 to 2070 studies. The false negative rate ranged from 1.87% to 12.20% for the FSL model and from 5% to 56.48% for the second reviewer compared with the principal reviewer.

Our FSL framework demonstrates the potential for reducing workload in SR screening by over 50%. However, the model did not achieve 100% recall at this threshold, highlighting the potential for omitting eligible studies. Future work should focus on developing a web application to implement the FSL framework, making it accessible to researchers.