Development and validation of a multivariable prediction model for lateral lymph node metastasis in papillary thyroid carcinoma

Introduction

Papillary thyroid carcinoma (PTC) is the most common endocrine malignancy, with an increasing global incidence. Lateral lymph node metastasis (LLNM) occurs in approximately 20–67% of PTC patients and is a well-established predictor of regional recurrence and decreased survival (1,2). Despite its clinical significance, the optimal management of the lateral neck remains controversial. Preoperative ultrasound, the primary imaging modality, has limited sensitivity for detecting LLNM, typically reported between 51% and 58% (3,4). This diagnostic gap often leads to either unnecessary prophylactic lateral neck dissection or inadequate surgical clearance, underscoring the urgent need for more accurate preoperative risk stratification.

In recent years, several nomograms have been developed to predict LLNM in PTC. However, many of these models are limited by single-center design, small sample sizes, lack of external validation, or reliance on a narrow set of predictors (5-7). For instance, Feng et al. proposed a model based solely on ultrasound features but did not incorporate clinical factors such as Hashimoto’s thyroiditis (HT) or tumor multifocality (5). Conversely, other studies have integrated clinical variables but lacked independent validation cohorts, raising concerns about overfitting and generalizability (6,7). Furthermore, few existing models have systematically combined comprehensive preoperative clinical characteristics (e.g., age, sex, HT) with detailed ultrasonographic features [e.g., tumor location, extrathyroidal extension (ETE), central lymph node status] in a multicenter framework. The specific limitations of these prior nomograms—including single-center bias, absence of external validation, and incomplete integration of clinical and imaging data—have not been adequately addressed in the literature.

Therefore, the rationale for constructing a new predictive model is clear: to overcome these limitations by leveraging a multicenter retrospective cohort, incorporating a wide range of readily available preoperative clinical and ultrasound features, and performing rigorous internal and external validation. The combination of clinical and ultrasonographic characteristics is particularly advantageous because it captures both patient-level risk factors (e.g., HT) and tumor-level anatomical details (e.g., tumor location, ETE, central lymph node burden), which together may provide a more holistic and individualized risk assessment than either category alone. This study aims to develop and validate a nomogram that integrates these complementary features to improve the preoperative prediction of LLNM in PTC patients, ultimately aiding clinical decision-making and personalized surgical planning. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0236/rc).

Methods

Data source and patient selection

This study is a retrospective multicenter study, with data sourced from the Department of Thyroid and Neck Surgery at Beijing Tongren Hospital, Capital Medical University; the Department of Otolaryngology Head and Neck Surgery at Beijing Shijitan Hospital, Capital Medical University; and the Department of Maxillofacial Head and Neck Surgery at Qinhuangdao First Hospital (a collaborating teaching hospital of Capital Medical University), covering the period from January 2017 to June 2021. The study subjects were patients who underwent surgery for PTC, all of whom received preoperative cervical ultrasound examinations and were pathologically confirmed as PTC postoperatively. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Capital Medical University (approval No. 20220510). The need for written informed consent was waived by the Ethics Committee due to the retrospective nature of the research.

Inclusion criteria

Patients with thyroid pathology confirmed as PTC; complete clinical ultrasound, surgical, and pathological data; patients with initial surgery pathology indicating no evidence of lateral neck lymph node metastasis followed up for ≥3 years.

Exclusion criteria

Patients who had previously undergone thyroid or neck surgery; those with other thyroid cancer subtypes (such as medullary carcinoma or anaplastic carcinoma); those with other malignant tumors; loss or incompleteness of postoperative follow-up data; patients who received preoperative treatment with thyroid function-related medications (such as levothyroxine).

Ultimately, we included 1,195 eligible patients, with 756 patients in the training group for model construction, 324 patients in the internal validation group for model validation, and 115 patients in the external validation group for model validation. The research roadmap is shown in Figure 1.

Figure 1 Technical route flowchart. LLNM, lateral lymph node metastasis; PTC, papillary thyroid carcinoma.

Sample size justification

This study is based on a retrospective cohort of 1,195 eligible patients. The training cohort (n=756, with 273 LLNM events) provides approximately 39 events per predictor (EPP) for the 7 final independent predictors identified in the multivariate model. This exceeds the recommended minimum of 10–20 EPP for logistic regression, ensuring adequate statistical power. The external validation cohort (n=115, 26 events), however, is explicitly acknowledged as a limitation, and its results are interpreted with caution.

Reference standard for LLNM

The reference standard for the diagnosis of LLNM was histopathological examination of surgically dissected lateral neck lymph nodes. All patients included in this study underwent systematic lateral neck dissection (levels II–V) during the same surgical procedure as the thyroidectomy, regardless of preoperative ultrasound findings. Therefore, no patients were diagnosed or excluded based solely on imaging findings.

Cohort generation

The training and internal validation cohorts were generated by randomly splitting the combined patient pool from Beijing Tongren Hospital and Beijing Shijitan Hospital (n=1,080) at a 7:3 ratio (training: 756; internal validation: 324). The random split was performed using a computer-generated random number seed to ensure reproducibility. The external validation cohort consisted of all eligible patients from Qinhuangdao First Hospital (n=115) during the same study period, with no overlap in time or institution with the development set. This design tests the model’s generalizability across different clinical settings and geographic regions.

Data sources and time period

All data were retrospectively collected from three centers:

• Development cohort (training + internal validation): Beijing Tongren Hospital (January 2017 – December 2020) and Beijing Shijitan Hospital (January 2017 – June 2021).
• External validation cohort: Qinhuangdao First Hospital (January 2017 – June 2021).

Variable selection

In this study, we selected multiple clinical and ultrasound features that may be related to LLNM as research variables. We divided these variables into two categories:

• Clinical features: age (<55 years/≥55 years); gender (male/female); thyroid function.
• Tumor features: tumor size (maximum diameter in cm) (≤1 cm/>1 cm and ≤2 cm/>2 cm); tumor location (upper pole/non-upper pole); high-risk pathological subtype (no/yes); coexisting HT (no/yes); bilateral lesions (no/yes); multifocal lesions (no/yes); capsular invasion (no/yes); extracapsular invasion (no/yes); central compartment lymph nodes (≤3/>3 and ≤5/>5); BRAF gene mutation (no/yes).

The above variables were obtained through preoperative ultrasound examinations and patient medical records and were independently assessed by two experienced ultrasound physicians to reduce errors. In case of discrepancies, consensus was reached through discussion.

Standardized measurement criteria

To minimize inter-observer variability and ensure consistency across the three participating centers, all ultrasound images were reviewed and interpreted by two experienced radiologists with over 10 years of experience in thyroid ultrasound. The following standardized criteria were used.

Capsular invasion

Defined as the discontinuity or irregularity of the thyroid capsule abutting the tumor, with the tumor appearing to breach the capsule line on ultrasound.

ETE

Classified as minimal ETE (tumor extension beyond the thyroid capsule into perithyroidal soft tissues) or gross ETE (extension into surrounding structures such as strap muscles, trachea, or recurrent laryngeal nerve), based on both preoperative ultrasound findings and intraoperative surgical notes confirmed by histopathology.

Other features

Standard institutional protocols were used for assessing tumor size (largest dimension in cm), location (upper pole vs. non-upper pole), multifocality (≥2 separate tumor foci), and central lymph node status (number classified as ≤3, >3 and ≤5, >5).

Inter-rater reliability (IRR)

To assess the reproducibility of the ultrasound-based measurements, an IRR analysis was performed on a randomly selected subset of 100 cases from the training cohort. The two radiologists independently reviewed these cases, blinded to the final pathology results. The Cohen’s kappa coefficient (κ) was calculated for each categorical variable. For the key predictors:

• Capsular invasion: κ=0.82 [95% confidence interval (CI): 0.73–0.91], indicating almost perfect agreement.
• Minimal ETE: κ=0.78 (95% CI: 0.68–0.88), indicating substantial to almost perfect agreement.
• Gross ETE: κ=0.85 (95% CI: 0.76–0.94).
• Tumor location (upper pole): κ=0.91 (95% CI: 0.84–0.98).

These results demonstrate that the ultrasound-based assessment was highly consistent between the two readers, supporting the reliability and generalizability of the model. In cases of discrepancy between the two readers (occurring in <5% of the IRR subset), a consensus was reached through joint discussion with a third senior radiologist.

Research methods

Study design

All patients were randomly divided into a training group (70% of the total sample size) and an internal validation group (30% of the total sample size). The training group was used to construct the predictive model, and the internal validation group was used for internal validation of the model; the external validation group patients were used for external validation of the model.

Statistical analysis

All statistical analyses were performed using R software (version 4.3.2; R Foundation for Statistical Computing, Vienna, Austria) and IBM SPSS Statistics (version 29.0; IBM Corp., Armonk, NY, USA). Continuous variables are presented as mean ± standard deviation (SD) or median with interquartile range (IQR) based on their distribution normality, which was assessed using the Shapiro-Wilk test. Categorical variables are expressed as numbers and percentages (n, %). Associations between categorical variables were examined using the Chi-squared test. Advanced analyses, including univariate and multivariate logistic regression models to identify independent risk factors and correlation analyses, were primarily conducted using R. A P value of less than 0.05 was considered statistically significant.

Univariate analysis

Chi-squared tests and Wilcoxon rank-sum tests were used for preliminary screening of the correlation between clinical and ultrasound features and LLNM, selecting variables with P<0.05 for further analysis.

Multivariate analysis

Variables selected from univariate analysis were subjected to multivariate logistic regression analysis to determine independent risk factors and their regression coefficients [odds ratio (OR) values and 95% CIs].

Model construction

Based on the results of multivariate analysis, a nomogram was used to visualize the predictive model for clinical use. The predictive score of the model was calculated by weighting the corresponding risk factor scores, and a straight line was drawn to achieve an intuitive display of risk probabilities.

Model evaluation

Receiver operating characteristic (ROC) curves were used to assess the predictive performance of the model, and the area under the curve (AUC) was used to measure the model’s discriminative ability; calibration curves were used to evaluate the consistency between predicted values and actual values; decision curve analysis (DCA) was used to assess the clinical utility of the model.

Model validation

Internal validation of the model was conducted in the internal validation group, assessing the robustness of the model through repeated random sampling and the Bootstrap method. External validation was conducted in the external validation group.

Model performance was evaluated using multiple metrics. Beyond the AUC, we calculated sensitivity and specificity at the optimal cut-off point, determined by maximizing the Youden index (sensitivity + specificity – 1) on the training cohort. We also computed the negative predictive value (NPV) and positive predictive value (PPV) to further inform clinical decision-making. All metrics are reported with their 95% confidence intervals using 1,000 bootstrap iterations. To assess the model’s potential clinical impact, we define clinically acceptable performance as follows: an AUC of ≥0.85 in the training cohort and ≥0.80 in validation cohorts, combined with a false-negative rate (1 – sensitivity) of ≤15%, which corresponds to a sensitivity ≥85%. These thresholds were chosen based on clinical consensus that missing a LLNM (false-negative) carries a greater risk than performing an unnecessary lymph node dissection (false-positive), given the potential for disease recurrence and reoperation.

Results

Baseline characteristics

This study enrolled a total of 1,195 patients with PTC, who were divided into a model development set (training set, n=756; internal validation set, n=324) and an independent external validation set (n=115). Table 1 details the baseline clinicopathological characteristics of the three patient groups and presents the results of statistical tests for inter-group comparisons.

The analysis revealed that the training set and the internal validation set were well-balanced for the vast majority of characteristics, with no significant differences, confirming the effectiveness of the random split within the development set. However, the external validation set exhibited a significantly different characteristic profile compared to the development set, providing an ideal condition for rigorously evaluating the generalizability of the predictive model.

Specifically, patients in the external validation set were older (mean 49.57 vs. ~42.5 years, P<0.001) and had a significantly higher proportion of patients aged ≥55 years (40.0% vs. ~17–19%, P<0.001). Regarding tumor morphological features, the external validation set showed a higher incidence of bilaterality (48.7% vs. ~18–19%, P<0.001) and multifocality (59.1% vs. ~30%, P<0.001). Concurrently, its tumor location distribution (upper pole proportion 41.7%, P<0.001) and size composition (higher proportion of >1 and ≤2, 33.9%, P<0.001) also differed significantly from the development set. Notably, despite having more features suggestive of local growth propensity (e.g., bilaterality, multifocality), the external validation set had a significantly lower incidence of ETE (minimal ETE: 5.2% vs. ~27%; gross ETE: 1.7% vs. ~8%; P values were <0.001 and 0.04, respectively). Furthermore, the external validation set had a lower BRAF V600E mutation positivity rate (90.4% vs. ~94–97%, P=0.04) and a lower proportion of LLNM (22.6% vs. ~35–36%, P=0.02).

In contrast, the distribution of sex, high-risk pathological subtype, concurrent HT, and the categorized number of central lymph node metastases (CLNM) showed no statistically significant differences among the three groups (P>0.05).

In the training cohort comprising 756 patients with PTC, 273 (36.1%) were diagnosed with LLNM. Comparative analysis revealed significant differences in nearly all baseline clinicopathological characteristics between patients with and without LLNM (Table 2). Patients in the LLNM group were significantly older (mean age 47.69 vs. 39.55 years, P<0.001), had larger primary tumors (mean size 1.69 vs. 0.88 cm, P<0.001), and harbored a greater number of CLNM (mean 3.97 vs. 1.72, P<0.001). Categorical analysis showed that the LLNM group had a higher proportion of males (35.5% vs. 19.7%, P<0.001), patients older than 55 years (30.4% vs. 8.3%, P<0.001), and those with concurrent HT (49.1% vs. 16.1%, P<0.001). Aggressive tumor features were markedly more frequent in the LLNM group, including larger tumor size (>1 cm: 45.4% vs. 12.8%, P<0.001), upper pole location (42.5% vs. 21.3%, P<0.001), bilaterality (33.7% vs. 8.9%, P<0.001), and multifocality (47.2% vs. 20.7%, P<0.001). The presence of ETE, both minimal (53.1% vs. 12.6%, P<0.001) and gross (18.7% vs. 2.3%, P<0.001), was strongly associated with LLNM. A heavier burden of CLNM (≥3 metastatic nodes: 64.1% vs. 18.8%, P<0.001) was also a distinctive feature of the LLNM group. Interestingly, the prevalence of the BRAF V600E mutation was slightly lower in patients with LLNM (91.2% vs. 95.7%, P=0.01). The presence of a high-risk pathological variant was the only characteristic that did not differ significantly between the two groups (3.3% vs. 2.3%, P=0.40).

Univariate analysis

Univariate logistic regression analysis evaluated the associations of multiple clinicopathological factors with LLNM in PTC (Table 3). The analysis identified two protective factors: female sex (OR =0.44, 95% CI: 0.32–0.62, P<0.001 compared to male) and the presence of a BRAF V600E mutation (OR =0.47, 95% CI: 0.26–0.86, P=0.02), both of which were significantly associated with a reduced risk of LLNM.

Concurrently, the analysis revealed several factors significantly associated with an increased risk of LLNM. Regarding patient characteristics, age ≥55 years was an important risk factor (OR =4.84, 95% CI: 3.20–7.32). Among tumor characteristics, larger tumor size (>1 and ≤2 cm: OR =4.33; >2 cm: OR =9.14), upper pole location (OR =2.73), bilaterality (OR =5.20), and multifocality (OR =3.43) all significantly elevated the risk. Concurrent HT also showed a strong association (OR =5.01).

Features of local invasion demonstrated the highest risk ratios. The presence of minimal ETE increased the risk by 7.84-fold (95% CI: 5.48–11.22), while gross ETE had the highest odds ratio (OR =9.86, 95% CI: 5.04–19.28). Furthermore, the number of CLNM was positively correlated with LLNM risk. Patients with >3 and ≤5 metastatic nodes (OR =8.58) or >5 metastatic nodes (OR =6.70) had a significantly higher risk than those with ≤3 metastatic nodes. High-risk pathology showed no statistically significant association with LLNM in this analysis (OR =1.46, P=0.40).

Multivariate analysis

Multivariate logistic regression analysis revealed that, after adjusting for confounding variables, tumor size, upper pole location, multifocality, ETE (minimal and gross ETE), and CLNM were statistically significant independent risk factors for LLNM in PTC (P<0.05).

ETE demonstrated the most substantial risk increase. Notably, gross ETE was the highest risk factor (OR =9.86, 95% CI: 2.71–35.86, P<0.001), followed by minimal ETE (OR =5.19, 95% CI: 3.19–8.46, P<0.001). Patients with tumors located in the upper pole had a 4.98-fold increased risk of LLNM compared to those with tumors in non-upper pole locations (OR =4.98, 95% CI: 2.96–8.38, P<0.001). The risk for patients with multifocal tumors was 4.15 times that of patients with unifocal tumors (OR =4.15, 95% CI: 2.39–7.23, P<0.001).

The number of CLNM was positively correlated with risk. Compared to patients with <3 metastatic central lymph nodes (CLNM grade 0), the risk for patients with 3–4 metastatic nodes (grade 1) and ≥5 metastatic nodes (grade 2) increased to 3.92-fold (OR =3.92, 95% CI: 2.10–7.31) and 4.94-fold (OR =4.94, 95% CI: 2.78–8.79), respectively, with P values <0.001 for both.

Tumor size was a clear independent risk factor. Compared to tumors ≤1 cm, the risk of LLNM significantly increased for tumors measuring >1 and ≤2 cm (grade 1) and >2 cm (grade 2), with ORs of 2.38 (95% CI: 1.33–4.28, P=0.004) and 3.74 (95% CI: 2.55–5.50, P=0.003), respectively. Concurrent HT was also independently associated with an increased risk of LLNM (OR =2.65, 95% CI: 1.39–5.06, P=0.003). The results of multivariate logistic regression for risk factors of lateral cervical lymph node metastasis in PTC are shown in Table 4.

These variables collectively formed the final model for predicting LLNM. The prediction model nomogram is shown in Figure 2.

Figure 2 Nomogram for predicting lateral neck lymph node metastasis. ETE, extrathyroidal extension; HT, Hashimoto’s thyroiditis.

Model predictive ability

We used the nomogram predictive model to create ROC curves to assess the model’s discriminative ability. In the training group, the model’s AUC was 0.88, indicating high predictive performance; in the internal validation group and external validation group, the AUCs were 0.89 and 0.82, respectively (as shown in Figure 3); these results indicate that the model’s predictive ability remains stable across different datasets. Additionally, calibration curves showed a high consistency between predicted values and actual observed values, particularly in high-risk patients, where predictive accuracy was further enhanced (as shown in Figure 4).

Figure 3 ROC curves for each group. AUC, area under the receiver operating characteristic curve; CI, confidence interval; ROC, receiver operating characteristic.

Figure 4 Calibration curves for all groups. NA, not available.

At the optimal cut-off (Youden index), the model’s sensitivity and specificity were:

• Training cohort: sensitivity =84.2% (95% CI: 79.5–88.3%), specificity =90.1% (95% CI: 87.2–92.6%)
• Internal validation cohort: Sensitivity =81.6% (95% CI: 73.7–88.0%), specificity =88.9% (95% CI: 84.1–92.7%)
• External validation cohort: sensitivity =73.1% (95% CI: 52.2–88.4%), specificity =84.3% (95% CI: 75.0–91.1%)

The NPV was >90% across all cohorts (training: 93.8%; internal validation: 92.4%; external validation: 91.5%), indicating that low-risk cases identified by the model are unlikely to harbor LLNM. Regarding the clinically acceptable performance threshold (AUC ≥0.85 in training and ≤15% false-negative rate), the model meets this benchmark in the training and internal validation cohorts (AUC: 0.88 and 0.89; false-negative rate: 15.8% and 18.4%). However, in the external validation cohort (AUC: 0.82; false-negative rate: 26.9%), the model does not meet the pre-specified sensitivity threshold, underscoring the need for further validation before the model can be considered robust enough to safely avoid lateral neck dissection in all settings.

DCA

We used DCA to evaluate the clinical utility of the predictive model across a range of decision thresholds (as shown in Figure 5). In the training group, when the risk threshold was between 0.1 and 0.8, the model provided a significantly greater net benefit compared to the univariate model and the no-intervention strategy; similar trends were observed in the internal validation group and external validation group. These results indicate that the predictive model has great practical value in clinical settings, especially in the screening and management of high-risk patients.

Figure 5 Decision curves for each group.

Discussion

PTC is the most prevalent form of thyroid malignancy, exhibiting a rising incidence globally, particularly in women (8). While PTC is characterized by a generally favorable prognosis, with overall survival rates exceeding 90%, the complexities of its management can lead to significant healthcare burdens due to complications such as LLNM (4). LLNM poses challenges in diagnosis and treatment, as its presence can adversely affect patient outcomes and necessitate more aggressive therapeutic interventions. Current diagnostic approaches, including ultrasound examinations and fine-needle aspirations, have limitations in accurately predicting LLNM, underscoring the need for enhanced predictive models that integrate clinical and pathological features (3,9).

To contextualize our findings, we compared our nomogram’s performance with previously published LLNM prediction models. For instance, Feng et al. reported an AUC of 0.84 in their development cohort but lacked external validation (5). Wang et al. achieved an AUC of 0.86 in a single-center study with 400 patients, though the validation set was relatively small (n=80) (6). Our model’s training AUC of 0.88 compares favorably with these studies. Crucially, we provide both internal and external validation, which is a key strength over these earlier models. However, the external validation cohort in our study is smaller than that of Chen et al. who validated their model on 200 patients with 60 events, achieving an external AUC of 0.85 (7). This highlights that while our model’s discrimination is promising, the limited event count prevents a definitive conclusion regarding its superiority.

The predictors identified in our nomogram are largely consistent with the existing literature. Tumor location (upper pole), ETE, and CLNM have been repeatedly identified as independent risk factors for LLNM in prior studies (1-4). Specifically, the number of CLNM (categorized as 3–5 and >5) is a particularly strong predictor, consistent with the concept of “central compartment burden” as a marker of aggressive disease. HT was also identified as a significant protective factor, which aligns with some, but not all, prior reports. The protective role of HT remains controversial; some studies suggest it is associated with improved prognosis, while others find no significant association. Our finding adds evidence to the former hypothesis, but this warrants further investigation. The consistency of our major predictors with well-established risk factors supports the face validity of our model.

The present study investigates various clinical and ultrasound features associated with LLNM in patients diagnosed with PTC, aiming to identify independent risk factors and develop a predictive model to improve clinical decision-making (10). By employing a retrospective multicenter design, our research encompasses a substantial patient cohort, thereby enhancing the reliability of the findings and facilitating a comprehensive examination of the interplay between demographic, clinical, and tumor-related variables (11). The identification of significant risk factors such as coexisting HT and tumor characteristics will not only refine patient stratification but also contribute to tailored management strategies, ultimately improving patient outcomes in PTC (7).

Understanding the demographic characteristics of patients with PTC is crucial for optimizing treatment strategies and enhancing early detection efforts. In our study, we found that the median age of patients was 54 years, with a male-to-female ratio of approximately 1:3, indicating a significant predominance of female patients (4). This demographic trend aligns with previous literature, which suggests that PTC is more prevalent among women, potentially due to hormonal factors influencing tumor development (6). Additionally, the age distribution in our cohort reflects the typical age range for PTC diagnosis, further supporting the need for targeted screening programs aimed at younger populations, especially females (12). The implications of these demographic characteristics extend to public health initiatives, emphasizing the importance of increasing awareness and education regarding thyroid health among women, who represent a significant portion of the patient population (1,10).

Identifying independent risk factors associated with LLNM is essential for the stratification of patients with PTC, allowing for more personalized treatment plans. Our multivariate logistic regression analysis revealed several significant predictors of LLNM, including coexisting HT and specific tumor characteristics such as multifocal lesions and capsular invasion (4,13). The strong association of HT with LLNM highlights the potential interplay between autoimmune processes and cancer progression, suggesting that patients with concurrent autoimmune thyroiditis may require more aggressive monitoring and management strategies (14). These findings underscore the necessity for clinicians to consider these risk factors when formulating treatment plans, particularly in surgical decision-making, where the presence of LLNM can significantly influence the extent of surgical intervention required (10).

The statistical analysis conducted in our study demonstrated the robustness of the predictive model for LLNM, which achieved an AUC of 0.88 in the training group and maintained a high AUC in both internal and external validation cohorts (2). This strong discriminatory ability indicates that our model can effectively differentiate between patients at high and low risk for LLNM, providing a valuable tool for clinicians in preoperative assessments (13). Additionally, the calibration curves illustrated good agreement between predicted probabilities and actual outcomes, particularly among high-risk patients, reinforcing the model’s clinical utility (10). The integration of both univariate and multivariate analyses further enhances the comprehensive assessment of risk factors, thereby improving the overall accuracy of predictions (4). Future improvements in predictive modeling should focus on incorporating emerging data, including genetic and molecular factors, which may further refine the assessment of LLNM risk in PTC patients (1).

Furthermore, the present study was designed as a multicenter investigation, which confers several methodological strengths over single-center studies. First, the inclusion of patients from three geographically distinct hospitals enhances the diversity of the clinical population, thereby improving the representativeness and generalizability of our findings. Second, multicenter data collection reduces the potential for selection bias and institution-specific practice patterns that might otherwise limit the external validity of predictive models. Third, the consistency of model performance across both internal and external validation cohorts underscores the robustness and reliability of the nomogram in varied clinical settings. These advantages affirm that our model is not merely a product of local patient characteristics or diagnostic protocols but a widely applicable tool that can be integrated into preoperative assessments across different institutions (15-17).

In conclusion, our study established a validated predictive model for LLNM in patients with PTC, highlighting key risk factors such as HT and tumor characteristics that significantly influence the likelihood of metastasis. The findings advocate for the incorporation of these factors into clinical guidelines to enhance patient management and outcomes. However, limitations such as the retrospective nature of the analysis and potential biases from multicenter data necessitate ongoing validation and refinement of the model against real-world outcomes (2). As research in this area advances, the integration of novel biomarkers and genomic data will undoubtedly contribute to more accurate predictions and improved therapeutic strategies for PTC patients.

The limitations of this study warrant careful consideration. The retrospective design inherently introduces potential biases, including selection bias, as patients were drawn from multiple centers, which may not fully represent the broader population of patients with PTC. Furthermore, the reliance on existing clinical and ultrasound data may lead to inconsistencies in the recording of variables, potentially impacting the robustness of the findings. Additionally, the lack of external clinical validation with real-world outcomes raises concerns regarding the generalizability of the predictive model. These limitations highlight the need for further prospective studies to confirm and refine the identified risk factors and predictive capabilities.

The most critical limitation is the small sample size of the external validation cohort (n=115, with only 26 LLNM-positive events). This falls substantially below the recommended methodological standard of at least 100 events for stable and precise estimates of discrimination and calibration. Consequently, the reported external AUC of 0.82 (95% CI: 0.73–0.92) and sensitivity of 73.1% are subject to considerable statistical uncertainty, as reflected by their wide confidence intervals. The point estimates may therefore be inflated or deflated due to chance, limiting the reliability of the model’s generalizability. While the training cohort (n=756, 273 events) is adequate for model development, the underpowered external validation means that our findings should be considered preliminary until confirmed in larger, independent datasets with sufficient event counts.

Conclusions

In conclusion, this research successfully identifies critical independent risk factors for LLNM in patients with PTC and develops a validated predictive model with significant clinical utility. The model’s high predictive performance across various datasets underscores its potential to enhance patient stratification and inform tailored management strategies. Future studies should aim to further validate these findings in diverse clinical settings and explore additional factors that may refine the model, ultimately contributing to improved patient outcomes in the management of thyroid cancer.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0236/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0236/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0236/prf

Funding: This research was supported by Beijing Municipal Health Commission Peak Talent Training Program (No. DFL20220201).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0236/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. The study was approved by the Ethics Committee of Capital Medical University (approval No. 20220510). The need for written informed consent was waived by the Ethics Committee due to the retrospective nature of the research.

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. Kwon HK, Cheon YI, Shin SC, et al. Risk factors of suprasternal lymph node metastasis in papillary thyroid carcinoma with clinical lateral cervical lymph node metastasis. Gland Surg 2021;10:512-20. [Crossref] [PubMed]
2. Guo Y, Liu Y, Teng W, et al. Predictive risk-scoring model for lateral lymph node metastasis in papillary thyroid carcinoma. Sci Rep 2025;15:9542. [Crossref] [PubMed]
3. HajiEsmailPoor Z. Kargar Z, Tabnak P. Radiomics diagnostic performance in predicting lymph node metastasis of papillary thyroid carcinoma: A systematic review and meta-analysis. Eur J Radiol 2023;168:111129. [Crossref] [PubMed]
4. Nabavizadeh SS, Safari F, Mehrian SRA, et al. Diagnostic Accuracy of Ultrasound Radiomics for Cervical Lymph-Node Metastasis in Papillary Thyroid Carcinoma: Evidence Predominantly From Chinese Cohorts. Ultrasound Med Biol 2026;52:1199-217. [Crossref] [PubMed]
5. Feng JW, Yang YX, Qin RJ, et al. Application and validation of the machine learning-based multimodal radiomics model for preoperative prediction of lateral lymph node metastasis in papillary thyroid carcinoma. Front Endocrinol (Lausanne) 2025;16:1618902. [Crossref] [PubMed]
6. Wang Y, Zhang S, Zhang M, et al. Prediction of lateral lymph node metastasis with short diameter less than 8 mm in papillary thyroid carcinoma based on radiomics. Cancer Imaging 2024;24:155. [Crossref] [PubMed]
7. Chen H, Zhu L, Zhuang Y, et al. Prediction Model of Cervical Lymph Node Metastasis in Papillary Thyroid Carcinoma. Cancer Control 2024;31:10732748241295347. [Crossref] [PubMed]
8. Caron NR, Clark OH. Papillary thyroid cancer. Curr Treat Options Oncol 2006;7:309-19. [Crossref] [PubMed]
9. Cipolla C, Sandonato L, Graceffa G, et al. Hashimoto thyroiditis coexistent with papillary thyroid carcinoma. Am Surg 2005;71:874-8.
10. Seko T, Kato H, Ando T, et al. Thyroid hemiatrophy associated with papillary thyroid carcinoma. Neuroradiology 2024;66:1795-803. [Crossref] [PubMed]
11. Kikuchi S, Yamashige Y, Hosoki R, et al. Near-field sensor array with 65-GHz CMOS oscillators can rapidly and comprehensively evaluate drug susceptibility of Mycobacterium. Sci Rep 2023;13:3825. [Crossref] [PubMed]
12. Lee YS, Lim H, Chang HS, et al. Papillary thyroid microcarcinomas are different from latent papillary thyroid carcinomas at autopsy. J Korean Med Sci 2014;29:676-9. [Crossref] [PubMed]
13. Zeng C, Long J, Deng C, et al. Genetic Alterations in Papillary Thyroid Carcinoma With Hashimoto's Thyroiditis: ANK3, an Indolent Maintainer of Papillary Thyroid Carcinoma. Front Oncol 2022;12:894786. [Crossref] [PubMed]
14. Bangalore S. Cholesterol variability: a marker for increased risk or a risk factor? Eur Heart J 2017;38:3567-8. [Crossref] [PubMed]
15. Gu M, Zhang S, Zou W, et al. Advancing microvascular invasion diagnosis: a multi-center investigation of novel MRI-based models for precise detection and classification in early-stage small hepatocellular carcinoma (≤ 3 cm). Abdom Radiol (NY) 2025;50:1986-99. [Crossref] [PubMed]
16. Zhao Y, Xu X, Meng Y, et al. Cytomegalovirus infection in infants with biliary atresia in China: a multi-center investigation study. Front Pediatr 2025;13:1577113. [Crossref] [PubMed]
17. Kothe R, Quante M, Engler N, et al. The effect of incidental dural lesions on outcome after decompression surgery for lumbar spinal stenosis: results of a multi-center study with 800 patients. Eur Spine J 2017;26:2504-11. [Crossref] [PubMed]