Patients with terminal cancers commonly receive aggressive and sub-optimal treatment near the end of life, which may not be beneficial in terms of duration or quality of life. To improve end-of-life care, it is essential to develop methods that can accurately predict the short-term risk of death. However, most prediction models for patients with cancer are static in the sense that they only use patient features at a fixed time. We proposed a dynamic prediction model (DPM) that can incorporate time-dependent predictors. We apply this method to patients with advanced non-small-cell lung cancer from a real-world database. Inverse probability of censoring weighted AUC with bootstrap inference was used to compare predictions among models. We found that increasing ECOG performance status and decreasing albumin had negative prognostic associations with overall survival (OS). Moreover, the negative prognostic implications strengthened over the patient disease course. DPMs using both time-independent and time-dependent predictors substantially improved short-term prediction accuracy compared to Cox models using only predictors at a fixed time. The proposed model can be broadly applied for prediction based on longitudinal data, including an estimation of the dynamic effects of time-dependent features on OS and updating predictions at any follow-up time.
Keywords: IPCW AUC; dynamic prediction; landmarking; model-based calibration; overall survival; time-dependent features.