A framework for meta-analysis of prediction models for binary and time-to-event outcomes

Citation

Debray T, Damen J, Reitsma H, Collins G, Hooft L, Riley R, Moons K. A framework for meta-analysis of prediction models for binary and time-to-event outcomes. Abstracts of the Global Evidence Summit, Cape Town, South Africa. Cochrane Database of Systematic Reviews 2017;(9 Suppl 1). https://doi.org/10.1002/14651858.CD201702.

Tags: Oral

2017 Cape Town [Global Evidence Summit]

Debray T¹, Damen J¹, Reitsma H¹, Collins G², Hooft L¹, Riley R³, Moons K¹

¹Julius Center for Health Sciences and Primary Care, ²Oxford University, ³Keele University

Background: It is widely recommended that any developed - diagnostic or prognostic - prediction model should be externally validated across different settings and populations. When multiple validations have been performed, a systematic review followed by a formal meta analysis may help to understand whether and under what circumstances the model performs accurately or requires further improvements.

Objectives: To discuss methods for summarising the performance of prediction models with both binary and time-to-event outcomes.

Methods: We present statistical methods for dealing with incomplete reporting (of performance and precision estimates), and to obtain time-specific summary estimates of the c-statistic, the calibration-in-the-large and the calibration slope. In addition, we provide guidance on the implementation of a Bayesian estimation framework, and discuss several empirically based prior distributions. All methods are illustrated in two example reviews where we evaluate the predictive performance of EuroSCORE II and Framingham Wilson.