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

Health strategies increasingly emphasize both behavioural and biomedical interventions, yet the complex and often contradictory guidance on diet, behavior, and health outcomes complicates evidence-based decision-making. Evidence triangulation across diverse study designs is essential for balancing biases and establishing causality, but scalable, automated methods for achieving this are lacking. In this study, we assess the performance of large language models in extracting both ontological and methodological information from scientific literature to automate evidence triangulation. A two-step extraction approach—focusing on exposure-outcome concepts first, followed by relation extraction—outperforms a one-step method, particularly in identifying the direction of effect (F1 = 0.86) and statistical significance (F1 = 0.96). Using salt intake and blood pressure as a case study, we calculate the Convergency of Evidence and Level of Convergency, finding a strong excitatory effect of salt on blood pressure (942 studies), and weak excitatory effect on cardiovascular diseases and deaths (124 studies). This approach complements traditional meta-analyses by integrating evidence across study designs, and enabling rapid, dynamic assessment of scientific controversies.