Artificial Intelligence-Based Copilots to Generate Causal Evidence

AbstractWhile there is growing consensus that real-world data should play a larger role in generating causal evidence for health care, it is less clear whether and how AI can help. Current approaches to AI-driven analysis of health data are ill-equipped to account for the many threats to causal validity. However, the current human-reliant pipeline for causal analysis also falls short: analyses are complex, require multidisciplinary expertise, and are slow, labor-intensive and error-prone. Here, we speculate how a "human-in-the-loop" AI-based system could help relieve bottlenecks to high-quality causal analyses. We describe how an AI-based causal copilot, leveraging the formal inferential structure of the causal road map, could guide and support researchers through a structured process of translating a causal question into a hypothetical experiment; translating contextual knowledge into transparent and well-justified assumptions; designing, testing, and benchmarking a corresponding statistical analysis plan and code (including integration of machine learning on multimodal data); and supporting causal interpretation of results. Such a system could augment the speed and quality with which researchers conduct causal analyses with real-world data, improve transparency and verification of analyses and assumptions, and ultimately serve as a basis for point-of-care personalized decision support.