Two-Stage Modeling for Dynamic Survival Prediction from Longitudinal Data

Abstract

Predicting a patient’s future health risk is essential for medical care, especially when deciding treatments or identifying patients who may need urgent attention. However, many patients’ conditions change over time, meaning that predictions made at one point may become outdated. Dynamic prediction methods address this challenge by updating risk estimates as new laboratory measurements become available. A widely used method called landmarking predicts the risk of an event (such as relapse or death) based on clinical data observed at a specific time point. However, its accuracy often becomes worse as the prediction window becomes longer---that is, the farther into the future we try to predict, the less reliable the prediction becomes. To address this limitation, we propose a simple but effective extension of landmarking called a two-stage modeling approach. Instead of predicting the outcome far into the future all at once, our method first forecasts near-future laboratory measurements and then uses those predicted values to make a shorter-range survival prediction. Even this straightforward naive two-stage method improves accuracy compared with standard landmarking. We further extend this approach by also using summary information about the uncertainty of the laboratory forecast (such as predicted mean and variance). This extended two-stage model achieves better. Across multiple clinical datasets, the proposed two-stage approach consistently improves prediction accuracy over conventional landmarking. These results show that a simple forecasting step can meaningfully improve dynamic risk prediction, offering a practical direction for clinical machine learning models that use longitudinal laboratory values.

Files

• …