Fig. 1: Overall workflow of automatic evidence triangulation using LLM. | Nature Communications

Fig. 1: Overall workflow of automatic evidence triangulation using LLM.

From: Evidence triangulator: using large language models to extract and synthesize causal evidence across study designs

Fig. 1

a The pipeline of using LLM to extract study designs, entities, and relations from textual titles and abstracts. b The framework of evidence triangulation; The Convergency of Evidence represents the integration algorithm of the supporting and opposing evidence behind the relation. The LoC score represents the reliability of causal relationship with scaled classification in three levels: weak (one star*), moderate (two stars**), and strong (three stars***). Excitatory(\({{\mathcal{E}}}\)) Occurs when the outcome changes in the same direction as the exposure or intervention. For instance, if an increased salt intake increases blood pressure or a decreased salt intake decreases blood pressure, we label the effect as excitatory. No Change (\({{\mathcal{N}}}\)) Refers to findings that indicate the exposure or intervention does not lead to a statistically significant change in the outcome (e.g., salt intake has no measurable impact on blood pressure levels). Inhibitory (\({{\mathcal{J}}}\)) Occurs when the outcome changes in the opposite direction of the exposure or intervention. For example, if an increased salt intake decreases blood pressure, we label this effect as inhibitory.

Back to article page