Page 49 - 18F-FDG PET as biomarker in aggressive lymphoma; technical and clinical validation

Predictive value of interim PET in DLBCL
Statistical analyses
Two approaches of meta-analysis
For the meta-analyses of the HRs, individual log hazard ratios (HRs) and standard errors (SE) were pooled using a random effects model (REML, restricted maximum likelihood). Together with the individual study results, the pooled effect estimate—expressed as HR and 95% confidence interval— was visualized in a Forest plot. Between-study heterogeneity was assessed by using Cochran’s Q and I2 statistics [29]. A 95% prediction interval around the HR was calculated to predict the expected range of the HR of a new (future) study [30]. A funnel plot was presented to visually assess if publication bias was likely [31]. For the diagnostic meta-analysis, the pooled sensitivity and specificity was obtained by Hierarchical Summary ROC curve (HSROC) models and ROC curves constructed in RevMan [32] using the input parameters of the HSROC models.
Influence of covariates
Several prespecified subgroup analyses—which included both clinical and methodological issues—were performed using univariate meta-regression models for the HRs and as covariate interaction term in the HSROC models. The following subgroup analyses were performed: study design (retrospective or prospective studies; blinded review or not reported; PFS or EFS), characteristics of patients (100% DLBCL or between 80 and 100%), treatments (ASCT upfront or not, preplanned or consolidative radiotherapy used or unknown), properties of scans (PET/CT or a combination of PET/CT and PET standalone systems, availability of a baseline PET or CT), and scoring issues (DS -, IHP -, or custom criteria, central review or local review).
Software
Statistical analysis was performed in R (version 3.2.5) [33] using the Metafor package and SAS Proc Nlmixed was used for the HSROC models. A P value of less than 0.05 was considered statistically significant.
47
3