Mixture models in diagnostic meta-analyses--clustering summary receiver operating characteristic curves accounted for heterogeneity and correlation.

OBJECTIVES

Bivariate linear and generalized linear random effects are frequently used to perform a diagnostic meta-analysis. The objective of this article was to apply a finite mixture model of bivariate normal distributions that can be used for the construction of componentwise summary receiver operating characteristic (sROC) curves.

STUDY DESIGN AND SETTING

Bivariate linear random effects and a bivariate finite mixture model are used. The latter model is developed as an extension of a univariate finite mixture model. Two examples, computed tomography (CT) angiography for ruling out coronary artery disease and procalcitonin as a diagnostic marker for sepsis, are used to estimate mean sensitivity and mean specificity and to construct sROC curves.

RESULTS

The suggested approach of a bivariate finite mixture model identifies two latent classes of diagnostic accuracy for the CT angiography example. Both classes show high sensitivity but mainly two different levels of specificity. For the procalcitonin example, this approach identifies three latent classes of diagnostic accuracy. Here, sensitivities and specificities are quite different as such that sensitivity increases with decreasing specificity. Additionally, the model is used to construct componentwise sROC curves and to classify individual studies.

CONCLUSION

The proposed method offers an alternative approach to model between-study heterogeneity in a diagnostic meta-analysis. Furthermore, it is possible to construct sROC curves even if a positive correlation between sensitivity and specificity is present.