Risk predictions are useful to inform patients and to guide medical decision making and the basis for precision medicine. The basis of a risk prediction model typically use variables such as age, gender, smoking status, disease status and history, blood tests and genetic markers. A suitable regression model combines these variables into a prediction. Prediction performance describes how well a model will do on future patients. Unfortunately, this is never knowable. The best we can do is to simulate the model being applied to future patients by repeatedly splitting the data set into training and validation part.
In this course we address the survival setting with competing risks. For a given time horizon there are three groups: subjects with event, subjects with a competing risk, and subjects event-free. There also is a possibility that subjects are not followed until the horizon (right censored). We discuss descriptive tools (calibration plots, retrospective risk re-classification) and inference based on the performance measures Brier score and time-dependent AUC adapted to the situation with right censored data and competing risks.
Attendees should have a basic understanding of regression models for survival analysis and experience using the R language. This short course is based on R.