Two-Stage Modeling for Dynamic Survival Prediction from Longitudinal Data

In dynamic survival prediction, landmarking predicts risk at a fixed future horizon from data observed at a single landmark time. Accuracy typically worsens as the prediction horizon increases. We propose a simple yet novel two-stage extension: first forecast near-future \emph{laboratory measurements}, then predict the outcome over the resulting shorter window. This \textit{naive} two-stage modeling already improves performance; an \textit{extended} version that also passes distributional summaries from the measurement forecast (e.g., predictive mean and variance) to the outcome model yields further gains. In experiments on a hospital cohort with routine laboratory measurements and the MIMIC-IV dataset, the two-stage approach consistently outperforms one-stage landmarking across horizons, with the extended variant best overall. In aggregate, our method improves AUC by about 3 percentage points at most compared with the one-stage baseline.