Leveraging two-phase data for improved prediction of survival outcomes with application to nasopharyngeal cancer

Accurate survival predicting models are essential for improving targeted cancer therapies and clinical care among cancer patients. In this article, we investigate and develop a method to improve predictions of survival in cancer by leveraging two-phase data with expert knowledge and prognostic index. Our work is motivated by two-phase data in nasopharyngeal cancer (NPC), where traditional covariates are readily available for all subjects, but the primary viral factor, human papillomavirus (HPV), is substantially missing. To address this challenge, we propose an expert-guided method that incorporates prognostic index based on the observed covariates and clinical importance of key factors. The proposed method makes efficient use of available data, not simply discarding patients with unknown HPV status. We apply the proposed method and evaluate it against other existing approaches through a series of simulation studies and real data example of NPC patients. Under various settings, the proposed method consistently outperforms competing methods in terms of c-index, calibration slope, and integrated Brier score. By efficiently leveraging two-phase data, the model provides a more accurate and reliable predictive ability of survival models.