Opening the Black Box: Explainable AI Tool for Transparent Citation Screening in Systematic Reviews

Background

Artificial intelligence (AI) tools for citation screening can reduce workload in systematic reviews, but limited transparency affects trust and adoption. This study develops an explainable AI screening system that classifies studies as relevant or irrelevant and provides interpretable explanations for each decision, improving trust, verification, and decision-making in evidence synthesis.

Methods

A transformer-based binary classifier was built using a fine-tuned BERT model trained on the CLEF TAR dataset. SHAP (SHapley Additive Explanations) was used to generate token-level attributions showing how words influence predictions. The model was integrated into an interactive interface enabling users to screen citations, view predictions, inspect explanations, and identify key terms driving decisions. The system supports iterative screening, feedback, and manual validation.

Results

The model achieved good performance (F1-score = 0.83, accuracy = 0.81, precision = 0.77, recall = 0.89). Users can explore influential features and understand how text drives classification. The interface supports efficient screening while maintaining transparency of model reasoning, helping detect errors and biases.

Conclusions

This work shows explainable AI can improve transparency and trust in citation screening workflows. Explanations support validation, accountability, and error detection. From a public health perspective, more efficient and transparent evidence synthesis may support faster translation of research into guidelines and policy decisions. Future work should assess user experience, reviewer confidence, and real-world efficiency. The tool may also support standardized and collaborative screening, offering a scalable framework for AI-enhanced evidence synthesis and improving auditability through user feedback.

Key messages

• Explainable AI tool using BERT + SHAP enables transparent citation screening with interpretable decisions.

• Improves trust, accuracy, and efficiency in systematic reviews, supporting faster, auditable evidence synthesis for public health and guideline development.