Abstract for: Systems Literature Analysis Engine: Automating the connection of scientific evidence to inform causal loop diagrams
Understanding the complex interactions that drive population-level health behaviors demands robust system mapping tools. However, synthesizing causal relationships from large bodies of scientific literature is time-intensive. This study introduces the Systems Literature Analysis Engine (SLAE), an automated pipeline that extracts cause-and-effect relationships using natural language processing, facilitating the creation of stakeholder-informed causal loop diagrams. Four research assistants independently coded 146 articles, yielding 1,165 causal relationships. The relationships were used to create a dataset split 80/20 for training and validation. To use SLAE, users can input proposed causal relationships and documents. From documents, SLAE’s Analysis Pipeline searches for a candidate causal relationship and classifies it as positive, negative, independent, or N/A. The output is used to generate a CLD in Kumu (kumu.io). SLAE classified the polarity of relationships with an average accuracy of 52–60%. It drastically reduced manual effort, completing the processing of 146 documents in two hours. In contrast, human coders spent 231 work hours performing similar tasks. Preliminary findings revealed that prompt engineering techniques, such as confidence scoring, reduced false attributions by about 17%. The approach demonstrates the potential for scaling systematic, evidence-based causal mapping across public health domains. By automating the extraction of cause-and-effect linkages from scientific literature, SLAE accelerates the process of building empirically grounded causal loop diagrams, offering considerable time savings. This methodological innovation can mitigate confirmation bias by systematically surfacing negative or null findings. The pipeline’s scalability also paves the way for broader model-based policy analyses in public health and other fields. Future work will refine accuracy and explore additional strategies to reduce hallucinations. NLP, longformer model, GPT4o