Design characteristics of external validation studies influencing the performance of risk prediction models

Article type
Year
Authors
Damen JAAG1, Debray TPA1, Heus P2, Hooft L3, Moons KGM4, Pajouheshnia R5, Reitsma JB1, Scholten RPJM3
1Julius Center for Health Sciences and Primary Care, and Cochrane Netherlands, University Medical Center Utrecht, Utrecht, The Netherlands
2Cochrane Netherlands, Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, Netherlands
3Cochrane Netherlands, Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, The Netherlands
4Julius Center for Health Sciences and Primary Care, and Cochrane Netherlands, University Medical Center Utrecht, Utrecht, Netherlands
5Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, The Netherlands
Abstract
Background: Meta-epidemiological studies have shown that study results are directly influenced by study design characteristics. The results of a randomized trial may for example be biased by inadequate allocation concealment and diagnostic test accuracy may be overestimated in case-control studies. The influence of design features on the results of prognostic research remains unclear.

Objectives: To determine which study characteristics influence performance of a prognostic model upon external validation, taking the validations of three established prediction models for cardiovascular disease (CVD) as an example.

Methods: In December 2015, MEDLINE, Embase, Web of Science, and Scopus were searched for articles investigating the external validation of three CVD risk equations (Framingham Wilson 1998, Framingham ATP III 2002 and Pooled Cohort Equations (PCE) 2013). Studies published before June 2013 were identified from a previous review. Studies were eligible if they validated the original prediction model in a general population setting. Data were extracted on key study characteristics. Random-effects meta-regression will be used to determine which study characteristics influence model performance (c-statistic and observed/expected ratio).

Results: The search identified 10,687 references, of which 1501 were screened in full text and 47 met our eligibility criteria. These articles described the external validation of Framingham Wilson (27 articles), Framingham ATP III (16 articles) and the PCE (10 articles). The c-statistic varied between 0.56 and 0.92. We will investigate sources of heterogeneity and present the range of performance for different design characteristics, including study design (e.g. cohort), sample size, assessment of outcomes, and handling of missing data.

Conclusions: This study will identify design characteristics influencing the performance of CVD risk prediction models in external validation studies, and thereby facilitate risk of bias assessment in systematic reviews of prognostic studies.