Article type
Year
Abstract
Background: Individual participant data (IPD) meta-analyses have several advantages when examining the diagnostic accuracy of a continuous marker. These include the direct estimation of the summary ROC curve and greater flexibility, validity and power to investigate differences in accuracy between clinical subgroups. However, no preferred statistical approach exists.
Objectives: To illustrate our approach based on ROC regression modelling using standardised marker values.
Methods: Our IPD meta-analysis approach consisted of the following steps: (1) determine the cumulative marker distribution among controls stratified by study; (2) use this distribution to standardize marker values for corresponding cases, known as placement values; (3) the cumulative distribution of 1 minus the placement values in cases will produce the summary ROC curve; (4) the mean value of the placement values in cases is equal to the AUC; (5) covariates can be added to the binary regression model to examine whether they affect discrimination.
Results: The IPD data of 16 studies were available, including a total of 1026 patients with mesothelioma (cases) and 4491 without (controls). Most studies applied a case-control design, often with multiple control groups per study. At a common threshold of 2 nmol L−1, the sensitivities and specificities of mesothelin across studies ranged widely from 19 to 68% and 88 to 100%, respectively. Heterogeneity was largely attributable to differences in study population, because type of control group, mesothelioma stage, and histologic subtype significantly affected the diagnostic accuracy. For mesothelioma patients with stage 1 and 2 compared to symptomatic controls, the area under the ROC curve was 0.77 (95% CI: 0.73–0.81). At 95% specificity, the corresponding sensitivity of mesothelin was 32% (95% CI: 26–40%).
Conclusions: IPD meta-analysis enables more insightful examination of the accuracy of a continuous marker, in particular to investigate differences in accuracy between patient subgroups.
Objectives: To illustrate our approach based on ROC regression modelling using standardised marker values.
Methods: Our IPD meta-analysis approach consisted of the following steps: (1) determine the cumulative marker distribution among controls stratified by study; (2) use this distribution to standardize marker values for corresponding cases, known as placement values; (3) the cumulative distribution of 1 minus the placement values in cases will produce the summary ROC curve; (4) the mean value of the placement values in cases is equal to the AUC; (5) covariates can be added to the binary regression model to examine whether they affect discrimination.
Results: The IPD data of 16 studies were available, including a total of 1026 patients with mesothelioma (cases) and 4491 without (controls). Most studies applied a case-control design, often with multiple control groups per study. At a common threshold of 2 nmol L−1, the sensitivities and specificities of mesothelin across studies ranged widely from 19 to 68% and 88 to 100%, respectively. Heterogeneity was largely attributable to differences in study population, because type of control group, mesothelioma stage, and histologic subtype significantly affected the diagnostic accuracy. For mesothelioma patients with stage 1 and 2 compared to symptomatic controls, the area under the ROC curve was 0.77 (95% CI: 0.73–0.81). At 95% specificity, the corresponding sensitivity of mesothelin was 32% (95% CI: 26–40%).
Conclusions: IPD meta-analysis enables more insightful examination of the accuracy of a continuous marker, in particular to investigate differences in accuracy between patient subgroups.