Scalable Causal-Interpretable Machine Learning for Cancer Prescreening Using Electronic Health Records

Electronic health records provide valuable information for disease risk prediction, but many machine learning models learn hidden associations that are difficult to understand for clinical decision-making. A research team developed UPEBNL, a new framework based on causal Bayesian networks that enables scalable and interpretable learning from large-scale health data. The method combines adaptive data slicing, quality-aware structure aggregation, and global construction of directed acyclic graphs. In simulations with high-dimensional data and millions of samples, UPEBNL improved structural recovery accuracy by nearly 40 percent and achieved up to 221-fold speedup compared to conventional approaches. When applied to prescreening for esophageal and colorectal cancer, the model achieved validation AUC values of 0.8171 and 0.784 respectively. The results suggest that this method can support interpretable cancer prescreening from large-scale electronic health records.