**2. Results**

We first evaluated the impact of the new TNM-8 on patients' risk stratification compared to the previous TNM-7. In this regard, 1113 patients for whom all clinical data required were available, were staged according to both TNM systems. Since in the TNM-8 the patients' age cutoff has been shifted from 45 to 55 years, the percentage of younger patients (age < 55 years) in our case study increased from 46.4% (516/1113) in TNM-7 to 68.6% (764/1113) in TNM-8. Due to the lack of distant metastases, all these patients were classified as stage I regardless of TNM edition. As expected, among older patients (age > 55 years) we observed a considerable reduction in the relative frequencies of individuals at stage III and IV, with a concomitant increase in those at stage I and II, moving from TNM-7 to TNM-8 (Figure 1). As noted in Figure 1, classification of PTC according to TNM-8 led to a considerable downstaging of patients.

**Figure 1.** Distribution of 1113 papillary thyroid cancer (PTC) patients according to the 7th or the 8th Tumor-Node-Metastasis (TNM) staging system. (**A**) displays the staging of patients younger than 45 years (516 out of 1113 patients) according to the TNM-7 or younger than 55 years (764 out of 1113 patients) according to the TNM-8. (**B**) displays the distribution of stages for older patients as defined in the 7th or the 8th TNM edition.

Univariate analysis was performed to evaluate the association of several clinicopathological parameters, including age at diagnosis, gender, autoimmune thyroid disease (AITD), tumor histology, size (T), lymph node metastases (N), stage, multifocality, capsular, muscle, and vascular invasion, with PTC recurrences. Since some categories were poorly represented, especially among tumor sizes and stages, they were combined with other categories to avoid the inclusion of too small groups in the statistics (see Section 4). As shown in Table 1, all the parameters analyzed with the exclusion of age and multifocality were significantly associated with PTC recurrences. In particular, the PTC sclerosing and tall cell variants and lymph node metastases had a very strong correlation with recurrences (Cramer's V index > 0.25).


**Table 1.** Univariate analysis of the association between clinicopathological parameters and PTC recurrences. AITD, Autoimmune Thyroid Disease. *p*-values < 0.05 are evidenced in bold.

These observations were confirmed by the Kaplan–Meier analysis, reported in Figure 2. All clinicopathological parameters, with the exception of age and multifocality, were found to impact significantly on disease-free interval (DFI). Patients classified as T1a and T1b had the same DFI.

We finally created Cox regression models to predict the probability of DFI as a function of different sets of independent variables. Clinical parameters were categorized as in the univariate analysis (see Table 1). The first series of covariates included gender, histological variants (classical, follicular, sclerosing, and tall cell), AITD, multifocality, stage, capsular, muscle, and vascular invasion. Tall cell

variant and stage turned out to be significant predictors of DFI (see Table 2), and the stage was significant either if calculated with the 7th or with the 8th edition. Of note is that using TNM-7 instead of TNM-8, the Akaike information criterion (AIC) decreased by about 12 units, indicating a significant improvement of the model.

**Figure 2.** *Cont.*

**Figure 2.** Kaplan–Meier analysis. Kaplan–Meier curves with Mantel–Cox log-rank statistical text were made to estimate the impact on disease-free interval (DFI) at age <55 years or ≥55 year (**A**), age <45 years or ≥45 years (**B**), PTC histological variants (**C**), AITD (**D**), multifocality (**E**), lymph node metastases (**F**), tumor size (7th edition) (**G**), tumor size (8th edition) (**H**), stage (7th edition) (**I**), stage (8th edition) (**J**), capsular invasion (**K**), vascular invasion (**L**), muscle invasion (**M**), or gender (**N**). *p* values < 0.05 are evidenced in bold.



We then replaced stage with stage-related parameters, so that the new set of covariates comprised lymph node metastases, tumor size, dichotomized age at diagnosis, gender, histological variants, AITD, multifocality, muscle and vascular invasion (see Table 3). Capsular invasion was excluded due to its collinearity with tumor size (variance inflation factor > 5). In this setting, the presence of lymph node metastases beyond the central compartment (N1b) was the independent variable most strongly associated with DFI (Hazard Ratio = 37.55, *p* < 0.001). Although in the univariate analysis dichotomous age was not significantly associated with recurrences or DFI, in Cox regression it became a significant predictor both applying the 55-year and 45-year threshold. When the hazard function was generated with T and age categorized according to the TNM-7 edition, the sclerosing and tall cell variants also displayed significant hazard ratios. Moreover, this model had lower AIC and BIC than that resulting from categories based on TNM-8, similarly to what was observed for the Cox regression including stage.

After that, we sought to verify if the DFI prediction could be strengthened by combining lymph node metastases and age with further histological and molecular parameters that were not available for our patients. These data were acquired from a previous study by the Cancer Genome Atlas Research Network for Cancer Genomics [10–12]. Specifically, we considered lymph node metastases, age, number of total non-silent mutations, number of CpGT mutations, BRAF-RAF score, ERK score, miRNA cluster, RPPA cluster, ploidy, differentiation score, and follicular component. Univariate analysis showed that BRAF-RAF score (*p* = 0.014), differentiation score (*p* = 0.045), and lymph node metastases (*p* = 0.005) associated significantly with recurrences. Nonetheless, adding all the above variables with *p* < 0.25 in the univariate analysis to lymph node metastases and age in the Cox regression did not improve significantly the prediction of DFI. Considering a subgroup of patients (*n* = 310) for which lymph node metastases were categorized as N0, N1a, and N1b in the database, N1b displayed the highest hazard ratio and significance, similar to what emerged from the analysis of our patient cohort (Table 3). In particular, N1a showed a HR of 6.53 (95% CI 1.14–37.37, *p* < 0.05) and N1b had a HR of 14.46 (95% CI 2.29–91.07, *p* < 0.01).

**Table 3.** Cox regression for the prediction of DFI performed with stage-determining factors and other clinicopathological parameters. The table shows predictors retained in the model after backward selection. The variables entered in the model were: gender, histological variants, AITD, multifocality, dichotomized age, T, N, muscle and vascular invasion. HR: hazard ratio. CI: confidence interval. AIC: Akaike information criterion. BIC: Bayesian information criterion. *p*-values < 0.05 are evidenced in bold.

