Enrico Martin

116 Chapter 6 characteristics included in at least three independent studies using the package ‘meta4diag’ in R. 26 In case of overlapping data between studies, data from the largest and most appropriate study was chosen for inclusion in quantitative synthesis. Penalized complexity priors were used for prior distributions. 27 Bayesian bivariate meta-analyses allow between study heterogeneity and differences in threshold parameters even for smaller sized samples of studies. Summary data were presented using summary receiver-operating characteristic (SROC) plots. The models generate an SROC curve with summary operating points, including 95% confidence regions and 95% prediction regions. Precision of the summary operating point can be assessed by the 95% confidence region which shows the pooled variability of sensitivity and specificity. Heterogeneity was assessed visually. Sources of heterogeneity were searched through subgroup analyses categorizing both FDG-PET and MRI studies in: large number of lesions (≥50 lesions), large proportion of MPNST (>33%), symptomatic lesions included only, and histologically proven lesions included only (either by biopsy or resection). MRI studies were additionally categorized for inclusion of NF1 patients only or mixed cases. Heterogeneity in sensitivity and specificity were assessed separately. Using the individual patient data of SUVmax values Bayesian bivariate meta-analyses of diagnostic accuracy were performed for thresholds at 3.0, 3.5, 4.0, and 4.5. The best threshold was obtained by evaluating significant differences in sensitivity first, after which lowest sensitivity thresholds were excluded and highest specificity was evaluated. For all comparisons made, significant differences were concluded whenever the lower bound of the 95% credibility interval (CI) of the highest accuracy did not include the mean of the lower accuracy. For application purposes the likelihood ratios may be interpreted as follows. A positive likelihood ratio (pLR) of ≥10 or a negative likelihood ratio of <0.1 correspond to a strong certainty to rule an MPNST in or out respectively. 28 A pLR of 5-10 or an nLR of 0.1-0.2 correspond to a moderate certainty to rule an MPNST in our out. We anticipated only a few studies on liquid biopsies and functional MRI sequences which would exclude them from meta-analyses, thus characteristics found in these studies would be assessed qualitatively. All statistical analyses were performed using R version 3.6.0 (R Core Team, 2019). Quality assessment The quality of included studies was appraised by two independent authors (R.G. and E.M.) using the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool ( Supplementary File 1 ). Disagreements were solved through discussion. For patient selection, case control studies, exclusion of patients with difficult diagnosis, or inclusion of histologically proven lesions only were deemed at high risk of bias. For index testing studies were assessed at high risk of bias when radiologists and nuclear medicine physicians were not blinded for pathology results or when new thresholds were used in results which were previously not determined in their method section. The reference standard was at high risk of bias when the pathologist was not blinded for results of the index test or if the lesion was found a BPNST without histological confirmation