*2.4. Data Analysis*

We followed all study participants until death or March 2021 (whichever occurred first). Patient demographics were expressed as frequencies (percentage), means (standard deviation), or medians (interquartile range). The primary endpoints were OS and progression-free survival (PFS). The OS was defined as the date of cancer diagnosis to the date of death or censored at the date of the last follow-up for surviving patients. PFS was calculated from the date of treatment initiation to the date of disease progression (e.g., growth of a residual tumor or development of new metastatic lesion), the date of disease recurrence after complete remission, the date of death, or censoring at the date of the last follow-up. Continuous variables were selected and the optimal cut-off values for each continuous variable were determined using receiver-operating characteristic (ROC) curve analyses. Only variables that were statistically significant predictors of death in the ROC curve analyses were selected for the survival analysis. Cut-off values with the highest

summation of sensitivity and specificity were selected as the optimal cut-offs for each continuous variable [28–30]. The variable selection and optimal cut-off determinations are summarized in the Supplementary Table S1. The continuous variables adjusted in the survival analysis were age, primary tumor SUVmax, primary tumor TLG, nodal SUVmax, nodal TLG, total TLG, and TNSUVproduct, and their optimal cut-off values were 75.5, 8.05, 42.5, 2.94, 18.3, 81, and 27, respectively. We used univariate and multivariate Cox regression analyses to study the association of the study variables with survival outcomes. First, we tested the effects of each variable on OS and PFS using univariate analyses. Then, the statistically significant variables from the univariate analysis were incorporated into the multivariate analysis to identify independent predictors of survival. We expressed the results of the survival analysis as hazard ratios (HR) and 95% confidence intervals. Statistical analyses were performed using SPSS software (version 20.0; SPSS Inc., Chicago, IL, USA).

The results of the multivariate Cox regression analysis were used to model the OS and PFS. The patient survival hazard was calculated by multiplying the HRs of the existing independent risk factors. For example, if a patient had three independent risk factors (HR1–HR3), the total hazard for this patient was calculated as HR1 × HR2 × HR3. If a patient had no risk factors, then the total hazard was the baseline hazard. If a patient had risk factors 1 and 3, then the risk of having shorter survival compared to no risk factor was HR1 × HR3.

#### *2.5. Survival Model Validation and Comparison*

The results of the multivariate Cox regression survival analysis were validated using the bootstrapping method. The validation process was performed with 1000 bootstrap samples. The results of bootstrapping validation were expressed as β, bias-corrected accelerated 95% confidence intervals, standard errors, and *p*-values. The validation process was executed using SPSS software (version 20.0; SPSS Inc., Chicago, IL, USA). The performance of our Cox regression models was assessed and compared with the AJCC staging system using Harrell's concordance index (c-index) and Kaplan–Meier curves (log-rank test) [19,31]. Different c-indices were compared using the "compareC" package installed on the R open-source statistical software version 3.4.2 (R Foundation, Vienna, Austria). A two-tailed *p*-value of <0.05 was considered statistically significant.
