MetaGD-CAN: A Hybrid Generative–Discriminative Method for Cancer Detection in EHR Data

Electronic Health Records (EHR) present a challenge for cancer detection due to the inherent noise caused by the clinical differential diagnosis process. Existing methods are prone to shortcut learning from spurious correlations in noisy EHR graphs, hindering the effective distillation of critical clinical signals. To address this, we propose MetaGD-CAN, a simple yet effective hybrid approach that integrates generative and discriminative models at the representation level. MetaGD-CAN deploys a cancer-conditional generative method to capture core dependency structures of EHR graphs by generating meta-graphs. Based on these meta-graphs, cleaner and more informative representations are formulated for the discriminative classifier, improving cancer detection. Experiments on the MIMIC-IV dataset demonstrate MetaGD-CAN’s superior performance in cancer detection, securing at least 4% increase in accuracy compared to baselines. Further analysis also validates MetaGD-CAN’s ability to distill cleaner information from EHR data, highlighting the strong potential of our approach.