Survival-supervised variational autoencoders for robust clinical prognosis under missing not at random conditions

Abstract Clinical survival analysis is often affected by missing not at random (MNAR) data, where missingness is due to underlying patient health states and clinical decision-making. Existing imputation methods struggle with threshold-based clinical missingness mechanisms. We propose MNAR-SVAE, a variational autoencoder integrating Cox proportional hazards loss, and introduce, for evaluation, a threshold-interaction MNAR mechanism capturing non-linear decision boundaries. We evaluate MNAR-SVAE against five baseline methods (standard VAE, MICE, missForest, GAIN, and MIDA) on METABRIC (breast cancer, $$N = 1904$$ ) and MIMIC-IV (sepsis, $$N = 25{,}439$$ ) under light, moderate, and severe missingness scenarios. MNAR-SVAE achieves the highest C-index on MIMIC-IV under severe missingness [C-index 0.693 (RSF), 0.693 (XGBoost)] and remains competitive under light [0.821 (RSF)] and moderate [0.778 (RSF)] conditions, with the performance advantage most pronounced under severe missingness, where $$\mathscr {L}_{\textrm{Cox}}$$ optimization preserves prognostic structure that reconstruction-focused methods fail to recover. On METABRIC, minimal performance variance is observed across all methods, indicating that the low-dimensional feature space in this dataset limits the marginal benefit of survival-supervised imputation. Overall, our results highlight the value of explicit survival-supervised generative modeling for handling informative missingness in clinical prognosis.