Deep learning models for the analysis of high-dimensional survival data with time-varying covariates while handling missing data

Abstract Recent advances in deep learning have expanded the potential for predictive modeling in survival analysis, particularly in high-dimensional datasets with time-varying covariates. This paper applies deep learning approaches, DeepSurv, DeepHit, and Dynamic DeepHit, to model HIV incidence (time-to-event outcome) using high-dimensional longitudinal data, incorporating time-varying cytokine profiles alongside baseline covariates. We employ the time-dependent concordance index (C-index) and Brier scores to assess the models’ predictive accuracy. We also address missing data using missForest, evaluating model performance on imputed and complete-case datasets. Different strategies for integrating cytokine profiles were explored: DeepSurv and DeepHit utilized derived variables, mean, and difference between the first and last measurements, while Dynamic DeepHit preserved the original time-varying nature of the cytokine data. Our findings demonstrate that retaining the dynamic nature of cytokine covariates, rather than relying on derived summary measures, underscores the robustness and suitability of Dynamic DeepHit as a clinical prediction model, particularly in scenarios where key variables evolve over time.