Factors Affecting the Diagnostic Performance of Time-Dependent ROC Curves for Longitudinal Data

Abstract

In medicine, ROC Curve Analysis is frequently used to determine the diagnostic performances of biomarkers. However, time-dependent ROC Curve is utilized in assessing the diagnostic accuracies of longitudinal biomarkers. One of the objectives of this thesis is to evaluate and to compare the diagnostic values of serial biomarker measurements taken from adults in predicting death in Intensive Care Units (ICU) at the end of follow-up period. Time-dependent Area Under Curve (AUC) values, which are calculated by performing joint modeling approach are used for this aim. The other objective is to compare the diagnostic performances of single measurement taken at baseline (t=0) and serial biomarker measurements taken within the follow-up period to determine whether a single value is sufficient to predict the event of interest. Furthermore, time-dependent diagnostic accuracies of these biomarkers are evaluated throughout the follow-up to identify which biomarker should be used at which time-point. Moreover, for each biomarker, cut-off values are determined with the help of Monte-Carlo simulation procedure. Also time-dependent cut-off values are obtained for discriminating subjects at risk and without risk of death on the first three days after the last biomarker measurement for each gender group. Besides, different joint model combinations are constructed for each biomarker to find out the best combination that provides the optimal diagnostic accuracy. In application part, diagnostic performances of serial C-Reactive Protein (CRP) and serial Procalcitonin (PCT) values in predicting death at ICU are investigated and determined that serial CRP values have higher diagnostic accuracy than serial PCT values in predicting death at the end of follow-up. Furthermore, the highest diagnostic accuracy is observed when single measurement of PCT is taken. PCT values are found to have higher diagnostic accuracy than CRP at especially later time-points within the follow-up period. Cut-off value of CRP is proposed to distinguish the groups since it has smaller Coefficient of Quartile Variation and smaller Robust Coefficient of Variation values compared to PCT. The first three days after the last biomarker measurement, cut-off values for PCT are found to be in decreasing trend for men and women, while constant cut-off values in the first two days; then decreasing trend for CRP are observed for both genders. Standard joint model gives the optimal diagnostic accuracy for both CRP and PCT. In conclusion, a comprehensive study has been carried out to assess the factors affecting the diagnostic performance of longitudinal biomarkers via a real-life data application. Coefficient of Quartile Variation measure and Robust Coefficient of Variation are suggested in the decision of choosing the relevant cut-off value. Taking serial biomarker values are suggested to better evaluate the longitudinal profiles of the subjects when needed.