Abstract

Modern AI systems excel at pattern recognition but fail at understanding causal relationships – the core of rational decision-making. At the same time, an enormous wealth of causal knowledge already exists, fragmented across millions of scientific publications, clinical trials, and expert reports. CausalMosaic addresses this gap by systematically synthesizing this fragmented evidence into unified, mathematically valid Causal Bayesian Networks (CBNs) that can be used for evidence-based decision support.

The framework combines LLM-based information extraction with classical causal structure discovery (CSD) algorithms. LLMs are used exclusively for structured extraction of causal and associative relationships from unstructured scientific texts – not for causal reasoning itself. The resulting observations are then processed by deterministic causal discovery algorithms, which construct a probabilistic graph model. A key innovation is the systematic propagation of uncertainty from the source material itself – through the learned graph structure and edge weights – all the way to the final probability distributions used for inference and intervention modelling.

The system is designed to be trustworthy by construction: each causal claim is traceable back to the underlying peer-reviewed studies, relationships are weighted by methodological quality (inspired by the GRADE framework), and conflicting evidence is resolved through Bayesian belief aggregation rather than subjective judgment. Users can ask interventional "what-if" questions (e.g., "If this patient improves their diet, how much does heart disease risk decrease?") and receive probability distributions with confidence intervals, full causal pathway explanations, and source citations – not opaque point estimates.

A primary use case is personalized medicine: domain experts (physicians, researchers, policymakers) define the variables they care about, and CausalMosaic automatically discovers the causal structure from the literature, enabling them to plan targeted interventions. The system is designed for domain experts who work with data but are not data scientists. CausalMosaic will be published as an open-source framework with a novel expert-annotated benchmark dataset, contributing to the democratization of causal AI methods.

Research Questions

RQ1 (Extraction). How can LLM-based extraction identify causal and associational relationships (directed effects, correlations, conditional independencies) in scientific text, and weight each one by the evidential strength of its source study (design, sample size, risk of bias) rather than by the model's own certainty?

RQ2 (Synthesis). How can these individually extracted, quality-weighted relationships be aggregated into one consistent causal graph — resolving contradictions across sources and representing each edge as an uncertainty-aware distribution rather than a binary decision?

RQ3 (Parameterization). How can the graph's conditional probability distributions be learned from the summary statistics reported in publications, without raw data, when different studies measure the same variable in different ways (e.g., smoking as yes/no, pack-years, or years smoked)?

RQ4 (Validation against the gold standard). How can the structural and probabilistic validity of automatically synthesized causal models be assessed against a purpose-built benchmark grounded in secondary-research outputs (systematic reviews, meta-analyses, guidelines)?