Researchers from Harvard Medical School and Massachusetts General Hospital have developed EviDAG, a browser-based tool that automates the construction of causal directed acyclic graphs (DAGs) from biomedical literature using large language models.

The system addresses a core challenge in medical research where analysts must manually connect study variables to prior literature and evaluate uncertain causal claims while preserving sufficient evidence for expert review.

How EviDAG Works

Given free-text descriptions of study concepts, EviDAG creates a reproducible literature snapshot and uses an LLM-based reasoning module to generate structured pairwise causal judgments. Each judgment links directly to verbatim evidence excerpts from the source literature.

The tool assembles these judgments into a constraint-checked graph where each proposed edge includes confidence estimates, provenance trails, and reviewable rationales. The interface supports study specification, progress monitoring, evidence review, graph comparison, and adjustment-set computation.

Researchers can export the completed DAGs for use in downstream causal analysis workflows.

Evaluation Results

In testing against both compact benchmark DAGs and reference DAGs derived from published literature, EviDAG achieved high edge recall on literature-based cohorts. The system retained verifiable evidence trails that were absent from LLM-only baseline approaches.

The evaluation demonstrated that EviDAG could successfully identify causal relationships supported by existing biomedical literature while maintaining full auditability of its reasoning process.

Research Applications

Causal DAGs serve as foundational tools in biomedical research for identifying confounding variables, planning study designs, and interpreting statistical results. The manual construction process typically requires extensive domain expertise and literature review.

EviDAG aims to reduce this curation burden while making the underlying causal assumptions transparent and auditable for peer review.

The research team published their findings on arXiv, with the paper authored by Yi-han Sheu, Michael R. Steigman, Yu Zhou, Bo Wang, Fan-Yu Yen, and Jordan W. Smoller. The tool supports the design, analysis, and interpretation of biomedical studies by automating evidence synthesis while preserving scientific rigor.