metaScreener: A Human-AI Collaborative Desktop Tool for Systematic Literature Screening

In fields such as academic research and pharmacovigilance, Systematic Literature Review (SLR) is a core method for obtaining high-quality evidence. However, manual screening has issues like being time-consuming, labor-intensive, and subject to subjective differences. metaScreener is an open-source desktop application developed by the Laval University team. By combining deterministic heuristic rules with large language model reasoning, it builds an auditable pipeline to achieve human-AI collaborative literature deduplication and relevance assessment, aiming to address the efficiency and accuracy challenges in SLR screening.

Systematic literature reviews require reviewing each article according to inclusion/exclusion criteria, divided into initial screening of titles and abstracts and full-text detailed screening. Large-scale projects may handle tens of thousands of articles in the initial screening phase, resulting in extremely high labor costs. metaScreener is positioned as a human-AI collaborative workbench; it does not pursue full automation but instead amplifies researchers' decision-making capabilities through technology while retaining the final adjudication right of human experts.

metaScreener adopts a plug-in architecture, with its core screening logic being a pipeline composed of configurable processing nodes. Node types include:

• Deterministic Filter: Quickly eliminate obviously irrelevant literature based on rules (keywords, publication year, literature type, etc.);
• LLM Reasoning Node: Call large language models to handle complex judgments (e.g., whether the research design meets requirements). The hybrid architecture balances efficiency (fast filtering via rules) and accuracy (semantic understanding via LLM).

metaScreener values academic transparency and reproducibility:

• Decision Log: Records the processing trajectory of each article (filter pass/exclusion status, LLM judgment reasons);
• SHA-256 Verification: Ensures the integrity of the software version to avoid result deviations;
• Configuration Export: Screening strategies can be exported as configuration files for easy collaboration and reproducibility.

Typical usage workflow of metaScreener:

1. Import Literature: Import literature to be screened from EndNote, Zotero, or database search results;
2. Configure Pipeline: Set up screening nodes according to inclusion/exclusion criteria (e.g., keyword filtering first, then LLM semantic judgment);
3. Automatic Initial Screening: The system categorizes literature into three types: definitely included, definitely excluded, and needing manual review;
4. Manual Adjudication: Researchers handle literature needing review; the interface displays metadata, abstracts, and LLM reasons to assist decision-making;
5. Export Results: Export the included list and decision logs for writing the review methodology.

metaScreener is open-source under the MIT license and supports cross-platform use on Windows, macOS, and Linux. Its plug-in architecture facilitates community contributions (e.g., discipline-specific filters, LLM integration plug-ins). Applicable groups include:

• Systematic review researchers (evidence-based fields such as medicine and psychology);
• Pharmacovigilance teams (monitoring drug safety signals);
• Policy analysts (grasping the current state of thematic research);
• Bibliometric scholars (large-scale literature preprocessing).

metaScreener embodies a pragmatic AI application concept: it does not replace humans but enhances human capabilities. In fields like literature screening that require large-scale data processing and complex judgments, human-AI collaborative solutions are more reliable and practical. As LLM capabilities improve, such tools will play an increasingly important role in academic research.