Objectives: This workshop will introduce participants to meta-analysis of risk prediction models. We will discuss opportunities and challenges of combining previously published prediction models. Subsequently, we will introduce statistical methodology and describe common software packages. Finally, we will illustrate the derivation of a meta-analytical prediction model using real example data.
Description: Risk prediction models are commonly developed for predicting the presence (diagnostic models) or future occurrence (prognostic models) of a particular outcome. They aim to provide absolute outcome risks for distinct individuals based on multiple predictors such as subject characteristics, clinical history and physical examination items, or more complex clinical measures such as medical imaging and biomarker results. Unfortunately, many prediction models are developed from relatively small datasets and perform more poorly than anticipated when applied in routine care. For this reason, researchers often reject previously published prediction models and develop a new one from their own data. This practice has lead to a cycle of underpowered model development, poor generalizability, and redevelopment, and generated a plethora of prediction models with similarities in populations and intended usage. Recently, it has been demonstrated tat meta-analysis of previously published prediction models may help to improve model performance across different patient populations, and to prevent users having to choose between a cacophony of existing models.
In this workshop we will discuss why meta-analysis offers unique opportunities to risk prediction research, and describe how previously published prediction models can appropriately be combined for this aim. We will demonstrate how researchers can accommodate for heterogeneity across study populations, and develop a meta-analytic model that yields accurate predictions in specific study populations. Finally, we will illustrate how researchers should interpret the achieved meta-analytical model performance and provide strategies for improving upon its generalizability.