268 curve (AuROCs) as overall measure of fit. ROC curves and associated AuROCs were established using the non-parametric, tie-corrected trapezoidal approximation method. Two correlated areas under the ROC curve were compared with each other using a non-parametric approach based on properties from generalized U-statistics to estimate a covariance matrix.28 Optimal thresholds for the most promising serum inflammatory biomarkers (serum amyloid A (SAA), Eotaxin-1, IL6, IL-8, IL-17A and TNF-α) were determined by equally maximizing sensitivity and specificity to compute the Youden’s index (J-statistic). Optimal thresholds or cut-off points (c) were established by selecting the highest Youden’s index, defined as: J = maxc {sensitivity (c) + specificity(c) - 1} Combinations of classifiers were empirically tested for their predictive performance using a nonparametric ROC estimation of combined predicted probabilities (derived from multivariable logistic regression) with bootstrap inference. Data were analyzed using SPSS Statistics 23.0 software package (SPSS Inc., Chicago, ILL, USA) and STATA software (version 15.0, Stata Corp, College Station, Texas, USA; commands used: ‘roctab’, ‘roccomp’ and ‘rocreg’) and visualized using GraphPad Prism version 6.0 (La Jolla, CA, USA). In case of multiple testing, Bonferroni corrections were applied. Two-tailed P-values ≤ 0.05 were considered as statistically significant. Internal validation Because all biomarker performances were tested on the same dataset, AuROCs and Youden’s indices as overall measures of predictive performance could potentially be overestimated due to the correlated nature of the data. To adjust for this potential bias, a bootstrap resampling procedure using 20,000 replicates was performed as internal validation and to obtain standard errors (SE) and confidence intervals (CI) for the AuROCs of best biomarker combinations. Chapter 8
RkJQdWJsaXNoZXIy MjY0ODMw