<p>Electronic Health Records (EHRs) provide rich opportunities for developing risk prediction tools to support clinical decision-making, yet they are inherently incomplete because data are recorded selectively during routine care. Such missingness may be informative, reflecting clinical judgment and patient status, and missing data patterns can shift between model development and real-world deployment. These challenges limit the reliability and transportability of predictive models in healthcare settings. We propose an imputation-free framework that jointly trains Conditional Variational Autoencoders with deep survival models to enable risk prediction directly from incomplete EHR data. We demonstrate the approach using the deep survival model DeSurv and evaluate its performance through simulation studies and two retrospective cohorts from the Clinical Practice Research Datalink primary care database. The proposed framework consistently outperforms conventional missing data methods, achieving superior performance on ground-truth metrics in simulations and improved calibration-based survival metrics in real-world cohorts. It also demonstrates increased robustness to unseen missingness patterns and distributional shifts. By providing a unified strategy for handling missing data across development, validation, and deployment, this work advances methodological robustness in healthcare informatics and supports more reliable clinical risk prediction in practice.</p>