Margriet Kwint

Chapter 5 90 Statistical analysis Baseline characteristics Tumor and patient characteristics at baseline are presented as proportions, mean (+standard deviation (SD)) or median (+ interquartile range (IQR)) in case of a not normally distributed variable. Quality of Deformable Registration To test the reliability of the DIR, we performed a sensitivity analysis in 66 patients for whom we manually delineated the GTV of the primary tumor on the CBCTs of the first and last fraction and compared this with the deformable GTV. (These 66 patients were used in a pilot study, in which we delineated manually the GTVs on the CBCTs). The quality of the deformable GTVs was subsequently assessed based on the correlation between the manual and the deformable relative GTV using the Pearson correlation coefficient. Further, a Bland Altman plot was used to further investigate the agreement between the manual and the deformable relative GTV. Latent Class Mixed Modelling Latent Class Mixed Modelling (LCMM) was used to identify subgroups with distinct treatment responses (16-19), specifically, relative GTV-change during CCRT. LCMM assumes a heterogeneous population of subjects with each subgroup having a specific average profile of the longitudinal marker. The GTV on the MidP-CT was regarded as baseline volume (100%) and the relative volume compared to baseline was calculated at each CBCT. To best reflect the shape of trajectories observed within the data and the distribution of GTV-change, a class-specific quadratic trajectory over time was assumed and a normalization of GTV-change was simultaneously done by splines (quadratic I-splines with 3 internal knots placed at the quantiles). Intra-patient correlation was captured by correlated individual Gaussian random effects on the intercept, time and time squared. To determine the optimal number of subgroups, a stepwise forward approach was used. First, a model with one subgroup was constructed (one average trajectory applying for all patients), and step by step, an additional subgroup, with a conceivably distinct trajectory from the already identified subgroups was added. For each model, a grid of initial values (20 departures) was also considered to ensure the convergence toward the global maximum. The optimal number of subgroups was chosen, based on a combination of (i) goodness-of-fit