Prognostic modeling and dynamic prediction for competing risks and multi-state models

A key question in clinical practice is prediction of the prognosis of a patient. These predictions are usually based on a prognostic model for the outcome of interest. There is great interest in the scientific community in the development of prognostic models in the context of survival data. One example in the field of breast cancer is the resounding success of Adjuvant! online, originally a computer program, later web-based, to assist in making decisions about adjuvant therapy for women with early breast cancer (Ravdin et al. 2001). The majority of prognostic models for survival data are based on multivariate Cox proportional hazards regression models (Cox 1972). The regression coefficients can be used as a prognostic score, with higher values indicating higher risk of the outcome. Other approaches like neural nets (Biganzoli et al. 2003) and classification and regression trees (CART, Breiman et al. 1984) are also in use, but considerably less frequently. Often when dealing with failure time data, more than one type of outcome can be distinguished. In breast cancer for instance, a distinction is often made between local and distant recurrence of the tumor, the development of new primary tumors, contralateral breast cancer, and death. Focusing on one outcome may divert attention from other outcomes, which are often equally important. Often, composite outcomes like disease-free survival (an event is defined as the first occurrence of local recurrence, distant metastasis, new primary or death) are then considered. This approach still ignores important information. This project is devoted to the development and implementation of prognostic models for prediction in the context of survival data with multiple outcomes, which are either mutually exclusive (competing risks) or can occur sequentially (multi-state model). An often ignored aspect of prediction is the fact that clinicians not only see patients immediately after primary treatment, but also on follow-up visits. Over time, a patient may have experienced intermediate events, and the clinician would need to update the prognosis to reflect this new information. (Importantly, not having experienced an intermediate event is also information.) The aspect of updating prognosis on the basis of accumulating information is called dynamic prediction and plays an important role in this project. The basis of (dynamic) prediction is formed by statistical models. These models may also be used directly to gain insight into the effects of covariates and intermediate events in the multi-state model. The standard techniques of model building in the context of competing risks and multi-state models fall short exactly in this respect. The alternative approaches that we propose in this project are designed to address the interpretational difficulties that come with the standard approaches to competing risks and multi-state models.