Recursive partitioning analysis improves risk stratification in node-negative differentiated thyroid cancer: a retrospective cohort study

Introduction

Thyroid cancer (TC) incidence has been rising globally, establishing it as one of the most common endocrine malignancies. Differentiated thyroid cancer (DTC) which includes papillary thyroid cancer (PTC) and follicular thyroid cancer (FTC), is the most common type of TC (1,2). Although the prognosis for majority of patients with DTC is excellent, substantial heterogeneity exists in long-term outcomes, particularly among patients classified as node-negative. The current American Joint Committee on Cancer (AJCC) Tumor-Node-Metastasis (TNM) staging system is the cornerstone for risk stratification, guiding treatment decisions such as the extent of surgery and the use of radioactive iodine therapy, and predicting patient outcomes (3). However, the TNM system, particularly for node-negative disease, has recognized limitations. It primarily relies on basic clinicopathological factors like age and tumor size, leading to significant heterogeneity within stages. Consequently, a substantial proportion of patients classified as early-stage may experience recurrence, while others may be over-treated, highlighting the need for a more precise and personalized staging approach (4,5).

Recursive partitioning analysis (RPA) is a robust statistical method that recursively splits a population into distinct prognostic subgroups based on covariates that maximize the difference in outcomes, such as overall or disease-specific survival (DSS) (6). Unlike traditional staging that uses fixed criteria, RPA creates a decision tree model that can integrate a wider array of variables, including patient age, tumor size, multifocality, histological type, etc., which can identify significant risk factors and their complex interactions (7). RPA-based models have demonstrated improved prognostic performance in several malignancies, suggesting potential value in refining risk stratification for node-negative DTC.

Previous studies have attempted to improve risk stratification for TC, but many are limited by small sample sizes, insufficient follow-up time, or inclusion of all histological subtypes, limiting their generalizability. Therefore, this study was conducted to develop and validate a novel staging system for node-negative DTC using RPA by leveraging a large sample data with long-term follow-up and to compare its prognostic performance with the AJCC TNM staging system. We present this article in accordance with the STROBE reporting checklist (8) (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0059/rc).

Methods

Study design and population

This retrospective cohort study analyzed the clinicopathological data of patients from the Surveillance, Epidemiology, and End Results (SEER) database whose pathological type were FTC or PTC. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

The patients who met the following inclusion criteria were enrolled between January 2004 and December 2019: (I) pathological type is FTC or PTC; (II) no distant metastasis was identified; and (III) no lymph node metastasis. Patients were excluded based on the following criteria: (I) have multiple primary cancers; (II) with tumor grade or marital status or survival data unknown; or (III) with no race. Finally, 8,816 eligible patients were included in this study and randomly divided into a training set (n=6,172) for model development and a validation set (n=2,644) for internal validation. Figure 1 displays the criteria for inclusion and exclusion. The T stage, N stage and final stage of all study patients were re-staged according to the 7th edition of the AJCC TNM staging system to avoid any conflict (9). Follow-up time was calculated from diagnosis to death or last follow-up.

Figure 1 Illustration of the study design for the inclusion and exclusion criteria for patient enrollment. M, metastasis; N, lymph node; T, tumor.

The primary outcome was overall survival (OS), defined as time from diagnosis to death from any cause. Covariates included age at diagnosis, gender, race, marital status, histological type, tumor grade, T stage, income, and rural-urban and group according to the SEER database grade categories. Age was treated as a continuous variable in regression analyses and categorized in RPA according to data-driven cutoffs.

RPA

To develop a prognostically distinct staging system, this study employed RPA, a multivariate statistical technique designed to categorize patients into homogeneous risk groups, which involved an iterative dichotomous partitioning process until no further subdivision was feasible, followed by a pruning of the initial tree to maximize predictive accuracy and prevent overfitting. But this is a complex programming and verification process; therefore, Xie et al. developed an autoRPA server, providing a user-friendly interface for RPA that enables users to manually prune and regroup tree nodes based on clinical expertise (10). The tool incorporates robust validation metrics—hazard consistency, hazard discrimination, percentage of variation explained, and sample size balance—and offers a standardized bootstrap method for objectively comparing the performance of newly developed models against established staging systems like the AJCC TNM classification.

Statistical analysis

Baseline characteristics were summarized using descriptive statistics. Used Cox regression analysis for univariate and multivariate analyses to screen for independent influencing factors, and RPA divided patients into different groups based on the result of MVA. The Kaplan-Meier survival curves was used to estimate the OS rate of two sets. The assessment of the relative discriminatory abilities among the staging systems was based on the time-dependent receiver operating characteristic (tROC) curves and the Harrell’s concordance index (C-index). The tROC is evaluated by calculating the area under the curve (AUC). The larger AUC, the more accurate and reliable the model’s prediction is. A higher C-index means a better discriminatory ability. R 4.2.1 software (The R Foundation for Statistical Computing, Vienna, Austria) was used for statistical analysis. For all statistical comparisons, significance was determined using a two-sided test with a threshold of P<0.05.

Results

Patient characteristics

A total of 8,816 patients with node-negative DTC patients were included in this study. Of the total cohort, 6,986 (79.2%) patients were female and 1,830 (20.8%) were male. The mean age of the entire study population was 47.4±14.6 years. FTC and PTC accounted for 747 (8.52%) and 8,069 (91.5%) cases, respectively. Most tumors (81.7%) were classified as Grade I, with the number of cases decreasing sharply as the grade increased. So, Grade II and Grade III tumors accounted for 14.4% and 3.19% of cases, with only a very small percentage (0.65%) being the most severe, Grade IV. The entire cohort was randomly divided into a training set (n=6,172, 70%) for model development and a validation set (n=2,644, 30%) for internal validation. The baseline characteristics of the overall cohort and the two subsets are summarized in Table 1. The median follow-up time was 85 months [interquartile range (IQR), 59–117 months]. There were no significant differences in age, gender, Grade, or histologic type between the training and validation sets (all P>0.05), which indicate a successful randomization.

Univariate and multivariate Cox regression analysis for OS

Univariate and multivariate Cox regression analyses were conducted. Univariate analysis showed that age, histological type, gender, marital status, and T stage (all P<0.05) are associated with OS (Table 2). There was no significant correlation between race, income, and radiation. Further MVA showed that age, histological type, gender, marital status were independent influencing factors (all P<0.05). Older age, FTC, higher grade, and being unpaired indicated a poor prognosis.

RPA staging system

To achieve this goal, patients in the training set were reclassified into prognostically homogeneous groups by applying RPA, which categorized them based on comparable 15-year OS rates. Age, tumor grade, and T stage were included for analysis. The training set was successfully divided into three groups (RPA stages I–III) according to autoRPA’s prioritization of independent variables. The three groups were as follows: RPA stage I (Grade I–IV, age <72 years) included 5,883 patients (95.3%); RPA stage II (T1–3, age ≥72 years) included 258 patients (4.2%); and RPA stage III (T4, age ≥72 years) included 31 patients (0.5%) (Figure 2). The 15-year OS rates of training set for RPA I–III stages were 93.6%, 45.6%, and 0% (Figure 3).

Figure 2 Proposed stage grouping derived by RPA. RPA I: Grade I–IV, age <72 years; RPA II: T1–T3, age ≥72 years; RPA III: T4, age ≥72 years. 15y-OS, 15-year overall survival; RPA, recursive partitioning analysis; T, tumor.

Figure 3 Kaplan-Meier analysis of OS by AJCC TNM and RPA staging systems in training and validation sets. Training set: (A) Kaplan-Meier estimates of overall survival stratified by AJCC TNM staging system; (B) Kaplan-Meier estimates of overall survival stratified by RPA staging system. Validation set: (C) Kaplan-Meier estimates of overall survival stratified by AJCC TNM staging system; (D) Kaplan-Meier estimates of overall survival stratified by RPA staging system. AJCC TNM, American Joint Committee on Cancer Tumor-Node-Metastasis; C-index, concordance index; OS, overall survival; RPA, recursive partitioning analysis; T, tumor.

Prognostic performance compared: RPA staging system against the AJCC TNM staging system

According to the Kaplan-Meier survival curve analysis, the 15-year OS rates of training set for each stage based on two staging system showed that, in the AJCC TNM system, the T1–T3 curves are relatively close (T1–T4 were 94.4%, 89.9%, 85.4% and 54.1%, Figure 3B) and partially overlap, and the RPA staging system shows better discrimination (all P<0.0001). Through the comparison between T4 stage and RPA III stage (54.1% vs. 0%) in OS rate, the RPA staging system also has a distinct advantage in identifying extremely high-risk groups. Further comparison of two staging system showed that RPA staging system has higher Harrell’s C‐index (0.947 vs. 0.750, P<0.001) and larger AUC (Figure 4A). A direct comparison confirmed the superior predictive performance of the RPA staging system over the AJCC TNM system.

Figure 4 The tROC curves for predicting patient survival of training set (A) and validation set (B). AUC, area under the curve; RPA, recursive partitioning analysis; TNM, Tumor-Node-Metastasis; tROC, time-dependent receiver operating characteristic.

Internal validation

Then we used internal validation set to verify the RPA staging system. Kaplan-Meier survival curves for OS showed the same performance as the training set (Figure 3C,3D). The RPA staging system demonstrated clearer survival distinctions between adjacent stages than the AJCC TNM system. The RPA staging system had a higher Harrell’s C index (0.944 vs. 0.802, P<0.005) and larger AUC (Figure 4B). The superior prognostic accuracy of the RPA staging system compared to the AJCC TNM staging system was firmly established in the internal validation set.

Discussion

The primary objective of this study was to develop a novel staging system for patients with node-negative DTC. Analysis of OS demonstrated that the RPA-based staging system exhibited superior risk stratification and predictive accuracy compared to the traditional TNM staging system. Furthermore, the RPA system shows significant potential for more accurately identifying ultra-high-risk patients with poor prognosis, thereby offering a more reliable foundation for guiding personalized clinical management strategies.

AJCC established TNM staging system which was accepted internationally to provide a standardized framework for prognostic prediction and treatment guidance in oncology. It has evolved into the globally recognized standard in clinical practice from its first edition to the present day. Following the 2018 recommendation to adopt the AJCC TNM staging system for TC, its prognostic performance has been evaluated in subsequent studies. van Velsen et al. analyzed OS and DSS of 792 patients whose histological type were DTC (11). When compared to the 8th edition, they also found that the 7th edition was inferior in terms of its efficacy and prognostic validity. This trend held true for all subclasses of DTC (12). However, this superiority is not universally observed in every component of the staging system. Contrary to the findings of Xiang et al. reported that the eighth edition showed better predictive ability only in the T stage (13). This is consistent with our finding through multivariate analysis (MVA) that T stage is one of the independent influencing factors. However, relying solely on T stage prediction would lack discrimination and accuracy, so RPA is a more comprehensive prediction model that includes other factors, addressing a key gap left by the singular reliance on TNM staging components.

Wang et al. analyzed 1,175 medullary TC patients’ data and found that introducing the number of metastatic lymph nodes into the staging system can more effectively distinguish the prognosis of patients (14). But some people believe that the location of lymph nodes especially cervical lymph nodes has a more important prognosis impact (15). Fei et al. found that regardless of T-staging, the cN0 PTC patients who without examined lymph nodes had OS similar with the pN1a patients but poorer than two or more examined lymph nodes (16). However, this view is not universally held, as other studies have found no significant association between the number of examined lymph nodes and survival outcomes (13). These discrepancies may stem from variations in sample sizes, surgical practices, and study populations. Our study addresses a critical gap identified in the beforementioned literature. While previous works debated the prognostic value of nodal examination, none successfully developed a comprehensive and practical staging system tailored for node-negative TC patients. Therefore, this study included cases with other factors in order to develop a new staging system through RPA for patients with node‐negative DTC based on large sample data with long‐term follow‐up. Future external validation in multi-institutional cohorts is warranted to confirm its generalizability before widespread clinical adoption.

Kim et al. also used RPA to refine the TNM staging system for DTC, they mainly through added age, extrathyroidal extension, tumor size to analysis and group up. However, the grouping is complex and too emphasizes the influence of age, and for example, the 10-year DSS between stage IA and IB (99.6% vs. 98.1%), stage IIA and IIB (93.0% vs. 92.4%) are too close, resulting in a lack of discrimination, at the same time, there was an overlap in the survival rates of stage III and stage IV, indicating insufficient predictive ability for critically ill patients (17). Wang et al. believe that young age (<55 years) cannot completely offset the risks brought by advanced T and N stages (18,19). So, we broke the conventional age grouping and to some extent weakened the influence of age on risk stratification. Among untreated PTC patients, who in the advanced-stage also have poorer prognosis, and the progression rate of the disease increases significantly with the progression of tumor staging (20). This also applies to both staging systems. Therefore, high-risk patients should be taken seriously, detected and intervened in a timely manner. The differentiation between different staging is higher in our analysis, especially at RPA stage III, the 15-year OS rate is 0%, which is significantly lower than that of stage II (45.6%, P<0.0001), hence, it has good predictive power for extremely high-risk patients.

This research contributes to the growing of evidence advocating for a fundamental revision of the TNM staging system for DTC, particularly for young patients. Patients were classified as RPA stage III, with a profoundly poor prognosis; therefore, this indicates the urgent need for these patients to be managed with the highest level of vigilance.

The reliability of the study’s findings is enhanced by its large sample size and extended follow-up period. Nevertheless, there are several limitations to this study. Owing to the potential selection bias introduced by the retrospective study design, confirmation of these results is required via prospective, multicentric studies. Furthermore, the primary limitation of internal validation lies in its inherent risk of overfitting and its consequent inability to reliably assess a model’s performance on external, independent datasets, which limits the generalizability of the findings. Besides, the lack of data on adjuvant chemotherapy regimens in the SEER database would potentially influence the prognostic performance observed for the RPA staging system. Finally, the most commonly treatment used for DTC is surgery, followed by radiotherapy and chemotherapy, may have a certain impact on the prognosis prediction of patients. In future research, the decision-making value of different staging systems for patients undergoing concurrent chemotherapy after surgery can be explored.

Conclusions

A novel staging scheme for predicting long-term OS in patients with DTC is presented in this large cohort study. This work fills a gap in node-negative DTC management, where the TNM system’s utility is limited. Developed through RPA, the system enhances prognostic accuracy by integrating new age cutoffs and tumor grade and demonstrates superior discriminatory power especially in extremely high-risk patients.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STROBE reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0059/rc

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0059/prf

Funding: This work was supported by grants from the National Natural Science Foundation of China (Nos. 82172671 and 81972970 to S.S.D.), and the Supporting Research Fund for High-level Full-time Talents Introduction of Guangdong Provincial People’s Hospital (No. KY0120220119 to S.S.D.).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0059/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Li M, Dal Maso L, Vaccarella S. Global trends in thyroid cancer incidence and the impact of overdiagnosis. Lancet Diabetes Endocrinol 2020;8:468-70. [Crossref] [PubMed]
2. Wang W, Yang Z, Ouyang Q. A nomogram to predict skip metastasis in papillary thyroid cancer. World J Surg Oncol 2020;18:167. [Crossref] [PubMed]
3. Haugen BR. 2015 American Thyroid Association Management Guidelines for Adult Patients with Thyroid Nodules and Differentiated Thyroid Cancer: What is new and what has changed? Cancer 2017;123:372-81. [Crossref] [PubMed]
4. Yu XM, Wan Y, Sippel RS, et al. S Should All Papillary Thyroid Microcarcinomas Be Aggressively Treated? An Analysis of 18,445 Cases. Ann Surg 2011;254:653-60.
5. Huang H, Liu Y, Ni S, et al. A prediction model for identifying high-risk lymph node metastasis in clinical low-risk papillary thyroid microcarcinoma. BMC Endocr Disord 2023;23:260. [Crossref] [PubMed]
6. Fors MM, Viada CE, Gonzalez P. Use of Recursive Partitioning Analysis in Clinical Trials and Meta-Analysis of Randomized Clinical Trials, 1990-2016. Rev Recent Clin Trials 2017;12:3-7. [Crossref] [PubMed]
7. Husain ZA, Chen T, Corso CD, et al. A Comparison of Prognostic Ability of Staging Systems for Human Papillomavirus-Related Oropharyngeal Squamous Cell Carcinoma. JAMA Oncol 2017;3:358-65. [Crossref] [PubMed]
8. von Elm E, Altman DG, Egger M, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. J Clin Epidemiol 2008;61:344-9. [Crossref] [PubMed]
9. Perrier ND, Brierley JD, Tuttle RM. Differentiated and anaplastic thyroid carcinoma: Major changes in the American Joint Committee on Cancer eighth edition cancer staging manual. CA Cancer J Clin 2018;68:55-63.
10. Xie Y, Luo X, Li H, et al. autoRPA: A web server for constructing cancer staging models by recursive partitioning analysis. Comput Struct Biotechnol J 2020;18:3361-7. [Crossref] [PubMed]
11. van Velsen EFS, Stegenga MT, van Kemenade FJ, et al. Comparing the Prognostic Value of the Eighth Edition of the American Joint Committee on Cancer/Tumor Node Metastasis Staging System Between Papillary and Follicular Thyroid Cancer. Thyroid 2018;28:976-81.
12. Kim M, Kim WG, Oh HS, et al. Comparison of the Seventh and Eighth Editions of the American Joint Committee on Cancer/Union for International Cancer Control Tumor-Node-Metastasis Staging System for Differentiated Thyroid Cancer. Thyroid 2017;27:1149-55.
13. Xiang J, Wang Z, Sun W, et al. A relook at the 8th edition of the AJCC TNM staging system of anaplastic thyroid carcinoma: A SEER-based study. Clin Endocrinol (Oxf) 2021;94:700-10.
14. Wang Z, Tang C, Wang Y, et al. Inclusion of the Number of Metastatic Lymph Nodes in the Staging System for Medullary Thyroid Cancer: Validating a Modified American Joint Committee on Cancer Tumor-Node-Metastasis Staging System. Thyroid 2022;32:536-43. [Crossref] [PubMed]
15. Zhang Y, Ji X, Zhang X, et al. Risk factors for cervical lymph node metastasis at different lateral levels in papillary thyroid cancer: level III as the central hub. Gland Surg 2024;13:1921-30. [Crossref] [PubMed]
16. Fei Y, Wang B, Yao X, et al. Factors associated with occult lateral lymph node metastases in patients with clinically lymph node negative papillary thyroid carcinoma: a systematic review and meta-analysis. Front Endocrinol (Lausanne) 2024;15:1353923. [Crossref] [PubMed]
17. Kim YN, Kim M, Ahn HS, et al. Refining the tumor-node-metastasis staging system for individualized treatment of differentiated thyroid carcinoma. Oral Oncol 2019;89:8-13. [Crossref] [PubMed]
18. Wang W, Bai N, Li X. A critical analysis of the current TNM classification for differentiated thyroid carcinoma in young patients: Time for a change? Front Endocrinol (Lausanne) 2022;13:939131. [Crossref] [PubMed]
19. Tuttle RM, Haugen B, Perrier ND. Updated American Joint Committee on Cancer/Tumor-Node-Metastasis Staging System for Differentiated and Anaplastic Thyroid Cancer (Eighth Edition): What Changed and Why? Thyroid 2017;27:751-6.
20. Wang D, Yang Y, He Y, et al. Natural History and Prognostic Model of Untreated Papillary Thyroid Cancer: A SEER Database Analysis. Cancer Control 2024;31:10732748241253956. [Crossref] [PubMed]