Development and validation of a prognostic nomogram to predict overall survival and cancer-specific survival for patients with anaplastic thyroid carcinoma

Study population

Cases diagnosed as anaplastic thyroid cancer during 1983–2013 were extracted from the Surveillance, Epidemiology and End Results (SEER-18) registries (National Cancer Institute, 2015). It includes data from 18 high-quality registries, which represents about 26% of the US population. Information about demographic, tumor characteristics and treatment for each case were obtained from the medical record. We identified anaplastic thyroid cancer patients using the International Classification of Diseases for Oncology, Third Edition (ICD-O-3). Consistent with previous research (Lim et al., 2017), only patients with ICD-O-3 histologic codes of 8020–8035 and topography code of C73 were included. Missing data including patients’ unknown race, radiation treatment and survival time were imputed with multiple imputations. We carried out multiple imputations based on R version 3.5.1 using the R package of mice. Then patients were excluded if they were not first malignant primary and had undefined surgery. We divided our data into internal validation and external validation (Fig. 1).

Covariates

The demographics of patients (age, sex, race and year of diagnosis), tumor characteristics (SEER historic stage A and tumor size) and treatment (surgery and radiotherapy) were gathered from the database. Patients were divided into <65 and ≥65 years old groups according to age at diagnosis. We have used splines to explore non-linear relationships between age at diagnosis and survival months. The result was shown as Fig. S1. As shown in the figure, survival months of patients exhibited a steep decline between 60 and 80 years old. It’s obvious that the age in this range is related to the survival months which was coherent with numerous studies. Some studies reported that age ≥70 years old was an indicator of poor survival for ATC patients (Onoda et al., 2020; Sugitani et al., 2012). Some studies reported that people younger than 60 years old had superior prognosis (Araque, Gubbi & Klubo-Gwiezdzinska, 2020; Roche et al., 2018). Taken these into consideration, we divided patients into <65 and ≥65 years old groups. SEER historic stage A was an indicator to distinguish tumor stage: localized (confined to thyroid), regional (tumor extend to adjacent tissues, blood vessels and lymph nodes), distant (distant metastasis) and unstaged. Only after 1983, SEER database began to record tumor size. So far, there were 3 codes to identify tumor size: extent of disease-4 codes (EOD-4) for 1983–1987, extent of disease-10 codes (EOD-10) for 1988–2003, and collaborative staging codes (CS) for 2004–2013. We combined these codes to define tumor size in patients with anaplastic thyroid cancer. Furthermore, patients were divided into lobectomy, total/near total thyroidectomy and no cancer-direct surgery/unknown groups according to the surgical procedure.

Outcomes

In this study, the outcomes we were interested in were overall survival (OS) and cancer-specific survival (CSS), which were defined as the interval from diagnosis until death for all causes and the interval from diagnosis until death only due to this tumor, respectively. SEER database provides specific codes to define both outcomes.

Data analysis

Only the first matching record was selected into analysis. The Kaplan–Meier method was used to compare OS and CSS between each covariate, and their differences were assessed by log-rank test. Multivariate Cox regression analysis was used to determine the independent risk factors of OS and CSS for anaplastic thyroid cancer patients through the SPSS statistical software (version 24). Besides, as various cause-specific deaths were competing risk events, we also used competing risk analyses as additional analyses to make our results more credible. We carried out competing risk analyses using the R package of cmprsk (Scrucca, Santucci & Aversa, 2007). All independent risk factors analyzed by multivariate Cox regression were included to construct the nomograms.

Nomograms were built and validated by R version 3.5.1, using the R package of rms and survival (Frank, 2019). We used bootstrap self-sample verification while establishing nomogram prediction model based on the data calculated by multivariate logistic regression analysis. Internal calibration and external calibration curves were used to assess the accuracy of the nomograms. To create a calibration diagram, we used Cox regression predictive model. X-axis means predictive survival time and Y-axis means actual survival time. In a perfectly predictive model, predictive rates would fall on a 45° diagonal line. We used concordance index (C-index) to describe the discrimination between predicted probability and actual outcome. The value of the C-index should fall between 0.5 and 1.0, with 0.5 presenting random chance and 1.0 presenting a perfect accordance. Generally, the value greater than 0.7 indicated a strong predictive ability. To evaluate the accuracy and practicability of our nomogram, we carried out receiver operating characteristic (ROC) analysis and Decision Curve Analysis (DCA) based on the R package of survival ROC and rmda respectively (Rousson & Zumbrunn, 2011). ROC represented sensitivity and specificity of nomogram while couldn’t ignore false positive rate and false negative rate. DCA represented net benefit of a clinical decision. To compare the accuracy of our nomogram with that of traditional TNM staging system, the net reclassification index (NRI) and integrated discrimination index (IDI) were analyzed based on the R package of nricens and predictABEL as reported (Zhou et al., 2019).

A two-sided P value of 0.05 was considered statistically significant. The Kaplan–Meier curves were built by GraphPad Prism 6.0.

Patient characteristics

A total of 1,404 patients diagnosed as ATC during 1983–2013 was included in the analysis. We divided our patients into two parts: internal validation and external validation. Internal validation contained patients diagnosed from 2000 to 2013, and external validation contained patients diagnosed from 1983 to 1999. Demographic, tumor characteristics and treatment information of patients in internal validation were presented in Table 1 (Year of diagnosis of patients in external validation ranged from 1983–1999). Elder (654 (64.5%)), female (630 (62.1%)) and white patients (815 (80.4%)) constituted the main part of the cases. Median age at diagnosis of cases was 70 years old and median follow-up time was 2 months. Most patients (567 (56.1%)) underwent tumor metastasis, while only 61 (6.0%) patients had localized tumors. When stratifying by tumor size, more than half of the patients had a tumor size greater than 4 cm. In addition, of the patients who underwent surgery, 177 (17.5%) cases underwent lobectomy and 287 (28.3%) cases underwent total or near total thyroidectomy. And 571 (56.3%) patients received radiotherapy.

Survival analysis

The Kaplan–Meier analysis (Figs. 2 and 3) showed age ≥65 years old rather than age <65 years old was significantly associated with worse OS and CSS for patients with ATC. It also revealed that sex and race was not significantly associated with OS and CSS. When considering year of diagnosis, we found that patients diagnosed from 2010 to 2013 had worst survival. When stratifying by tumor stage, ATC with distant metastasis showed severely poorer OS and CSS than a localized or regional tumor. The results also indicated ATC with larger tumor size had significantly worse survival. OS and CSS were significantly better for patients who underwent total or near total thyroidectomy than those who underwent lobectomy, while patients with no cancer-direct surgery or unknown surgical procedure had the worst survival. As for radiotherapy, survival was better for patients receiving radiotherapy than patients who did not receive radiotherapy.

Independent factors associated with OS and CSS

The Cox proportional hazards regression analysis was used to determine independent risk factors for OS (Table 2), and Cox proportional hazards and competing risk analyses were used to determine independent risk factors for CSS (Table 3). In our study, age ≥65 years old was a risk factor with worse OS (HR, 1.525; (95% CI [1.326–1.752]); P < 0.001). The results also indicated that tumor stage (distant vs localized: HR, 2.641; (95% CI [1.923–3.628]); P < 0.001); regional vs localized: HR, 1.427; (95% CI [1.031–1.974]); P = 0.032, unstaged vs localized: HR, 2.138; (95% CI [1.393–3.282]; P = 0.001), tumor size (unknown vs ≤2.0 cm: HR, 1.647; (95% CI [1.115–2.433]); P = 0.012), surgery (total/near total thyroidectomy vs lobectomy: HR, 0.715; (95% CI [0.583–0.878]); P = 0.001; no cancer-direct surgery/unknown vs lobectomy: HR, 1.702; (95% CI [1.420–2.040]); P < 0.001) and radiotherapy (yes vs no: HR, 0.508; (95% CI [0.445–0.580]); P < 0.001). Besides, sex, race and year of diagnosis were not significantly associated with survival of patients. And after adjusting for other available variables, age ≥65 years old was independently associated with worse OS (HR, 1.471; (95% CI [1.276–1.696]); P < 0.001) and tumor size, surgery and radiotherapy were also independent risk factors. As for CSS (Table 3), we found that age, stage, surgery and radiotherapy were risk factors according to univariate survival analysis. After adjusting for other available variables, we found that age ≥65 years old was an independent risk factor with worse CSS (HR, 1.408; (95% CI [1.216–1.631]); P < 0.001), which was consistent with the results of competing risk analysis (HR, 1.205; (95% CI [1.053–1.379]); P < 0.01). Similar to age, stage, surgery and radiotherapy were also independent risk factors according to both multivariate survival analysis and competing risk analysis. While sex, race, year of diagnosis and tumor size were not associated with prognosis of ATC patients.

Nomogram construction and validation

All independent risk factors derived from Cox proportional hazards regression analysis were used to construct the nomograms, which were effective tools to predict 6-month, 1-year and 2-year OS or CSS (Fig. 4). Therefore, age group, tumor stage, tumor size, surgery and radiotherapy were included in the nomogram for OS and CSS. Both for OS and CSS, tumor stage contributed most for survival. Using the nomograms to assess the prognosis of a patient requires adding the scores corresponding to each predictor to get a total score. The total score corresponds to 6-month, 1-year, 2-year overall or CSS rate.

The nomograms were validated internally and externally. The C-index values of internal validation were 0.765 for OS and 0.773 for CSS, respectively, indicating acceptable discriminations. In addition, the calibration curves for OS and CSS demonstrated great accordance between predicted survival by nomogram and actual observed survival. Internal calibration curves (Fig. 5) and external calibration (Fig. 6) curves suggested that our nomograms had proper calibrations.

To evaluate accuracy of our nomograms, we analyzed ROC curves. The ROC curves indicated sensitivity and specificity of a clinical evaluating model. The Y-axis meant sensitivity and the X-axis meant false positive rate. Because TNM was an acknowledged clinical prognostic factor and age was a risk factor related to ATC, we compared our nomograms with the accuracy of TNM and age evaluations. The ROC curves of 6-month, 1-year, 2-year nomograms for OS and CSS were all superior to TNM and age evaluations (Fig. 7). To assess the practicability of our nomograms, we analyzed DCA (Fig. 8). DCA represented net benefits of clinical decisions. The Y-axis meant net benefit and the X-axis meant high risk threshold. The horizontal gray line meant all true negative rate and diagonal gray line meant all true positive rate. DCA suggested that nomogram for OS was superior to TNM and age evaluations, while nomogram for CSS was similar to TNM evaluation but superior to age evaluations. Furthermore, our model exhibited better discrimination compared based on NRI and IDI. The NRI for the 6-, 12- and 24-month follow-up examinations were 0.593 (95% CI [0.393–0.781]), 0.589 (95% CI [0.353–0.786]) and 0.473 (95% CI [0.103–0.705]), respectively. Similarly, the IDI for the 6-, 12- and 24-month follow-up examinations were 0.105 (P < 0.001), 0.092 (P < 0.001) and 0.061 (P < 0.001), respectively. Both NRI and IDI indicated a superior predictive ability of our model compared to TNM staging system.