Imputation Free Deep Survival Prediction with Conditional Variational Autoencoders.

Area of Science:

• Healthcare Informatics
• Machine Learning in Medicine
• Biostatistics

Background:

• Electronic Health Records (EHRs) are valuable for clinical decision-making but often contain incomplete data due to selective recording.
• Missing data patterns in EHRs can be informative but pose challenges for predictive model reliability and transportability.
• Existing methods struggle with missingness that varies between model development and real-world deployment.

Purpose of the Study:

• To develop an imputation-free framework for risk prediction directly from incomplete EHR data.
• To improve the reliability and transportability of predictive models in healthcare settings.
• To address challenges posed by informative and shifting missing data patterns.

Main Methods:

• Jointly training Conditional Variational Autoencoders with deep survival models.
• Utilizing the deep survival model DeSurv for risk prediction.
• Evaluating performance through simulation studies and retrospective cohorts from the Clinical Practice Research Datalink database.

Main Results:

• The proposed framework consistently outperformed conventional missing data methods.
• Achieved superior performance on ground-truth metrics in simulations.
• Demonstrated improved calibration-based survival metrics and robustness to unseen missingness patterns in real-world cohorts.

Conclusions:

• The imputation-free framework offers a unified strategy for handling missing EHR data across the model lifecycle.
• Advances methodological robustness in healthcare informatics.
• Supports more reliable and robust clinical risk prediction in practice.