Clinical prediction models are an important asset in contemporary decision making. These models typically estimate the risk or probability that a certain condition or disease is present (diagnostic models) or will occur in the future (prognostic models) by combining information from multiple variables (predictors) from an individual, e.g. predictors from patient history, physical examination or medical testing (including e.g. imaging, electrophysiology and biomarker tests). Prediction models may, for instance, be used to inform the referral of patients for further testing, to plan lifestyle or therapeutic decisions or to risk-stratify participants in therapeutic clinical trials. Unfortunately, the vast majority of clinical prediction models are developed from small datasets, leading to so-called particularistic prediction models that poorly discriminate between individuals from new populations, settings or time periods and systematically over- or underestimate their risk. As a result, most prediction models require extensive validation and adaptation, and should not simply be implemented in daily practice or in clinical guidelines. To evaluate and improve the generalizability of developed prediction models, it has been recommended to use data from different studies, populations, settings and time periods. This strategy is now increasingly implemented by conducting large scale multi-center studies or by combining randomized and observational (registry) study data with different patient in- and exclusion criteria.

In this seminar, I will describe statistical methods that make use of multiple Individual Participant Data sets to develop, validate and/or update prediction models. Furthermore, I will illustrate their implementation and potential advantages using several examples, and highlight important issues for which further research is warranted.