DySurv: dynamic deep learning model for survival analysis with conditional variational inference.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Statistics

Background:

• Traditional machine learning models predict events at fixed time points.
• Survival analysis offers dynamic risk prediction by estimating time-to-event distributions.
• Electronic health records (EHR) contain valuable longitudinal data for patient risk assessment.

Purpose of the Study:

• To introduce DySurv, a novel conditional variational autoencoder-based method for dynamic risk prediction.
• To utilize both static and longitudinal EHR data for estimating individual risk of death.
• To develop a non-parametric approach for time-to-event analysis without underlying stochastic process assumptions.

Main Methods:

• DySurv employs a conditional variational autoencoder framework.
• It directly estimates the cumulative risk incidence function.
• The method was evaluated on 6 benchmark time-to-event datasets and 2 real-world EHR datasets (eICU, MIMIC-IV).

Main Results:

• DySurv demonstrated superior performance compared to existing statistical and deep learning methods.
• Achieved over 60% time-dependent concordance on the eICU dataset.
• Outperformed clinical scores (APACHE, SOFA) by over 12% in accuracy and 22% in sensitivity.

Conclusions:

• DySurv represents an interdisciplinary framework integrating deep learning, survival analysis, and intensive care for robust time-to-event prediction.
• The method shows consistent predictive capacity and disentangled survival estimates across diverse datasets.
• Further exploration of deep learning paradigms for survival analysis is warranted.