Evaluation of cervical lymph node metastasis after total thyroidectomy for differentiated thyroid carcinoma: construction of a multimodal non-invasive predictive model based on serum thyroglobulin and ultrasonic features

Introduction

Differentiated thyroid carcinoma (DTC) comprises over 90% of thyroid malignancies. Although generally having a favorable prognosis, cervical lymph node metastasis (LNM) occurs in 30–80% of cases (1,2). Despite standard primary treatment with thyroidectomy and lymph node dissection (LND) (3,4), 10–30% of patients experience disease persistence or recurrence, predominantly (74%) as cervical LNM (5). These recurrent lesions are often more aggressive, potentially invading vital structures and necessitating reoperation (6).

Among patients with persistent/recurrent disease, 5–43% undergo secondary surgery (3,7,8), with multiple operations being linked to worse prognosis (9). Therefore, accurate preoperative assessment of metastatic burden is critical for surgical success and patient outcomes. Secondary surgery, however, is more challenging and risky due to altered anatomy and scarring (10,11). Inadequate dissection risks residual disease, whereas overly extensive surgery increases complications such as nerve injury (4,12). Thus, preoperative evaluation is key to balancing efficacy and safety (9).

Post-thyroidectomy surveillance relies on serum thyroglobulin (Tg) and neck ultrasound (US) (3,13,14). Tg elevation suggests residual or recurrent tumor. Bachelot et al. demonstrated a correlation between serum Tg and metastatic lymph node volume (15), supporting its use for burden assessment. However, no quantitative clinical threshold was established. In practice, under thyroid stimulating hormone (TSH) suppression, Tg’s predictive threshold remains unstandardized and is confounded by assay variability and anti-Tg antibodies (3). Some patients have confirmed metastasis even with Tg in the “indeterminate response” range (e.g., non-stimulated Tg 0.2–1 ng/mL or stimulated Tg 1–10 ng/mL) (3,16). Moreover, Tg becomes undetectable in only 30–50% after compartment-oriented dissection for recurrence (17), highlighting the need for better preoperative tools.

Neck US is the primary imaging modality but its accuracy is operator-dependent and limited for atypical metastases (18). Critically, most predictive research focuses on initial DTC, leaving a gap for the recurrent population (19,20). No study has integrated serum Tg with US morphology into a multimodal tool to quantify nodal burden in this group. Decisions often rely on experience rather than objective data.

Therefore, a multimodal model integrating Tg and US features holds significant clinical value. This study aimed to develop and validate such a model to identify patients with high-volume metastasis (LNM ≥5) after total thyroidectomy for DTC. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0082/rc).

Methods

Study population

This single-center retrospective study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Ethics Review Board of China-Japan Friendship Hospital (No. 2019-103-K71). The requirement for informed consent was waived due to the retrospective nature of the study. We included 97 consecutive patients (May 2020 to December 2024) with pathologically confirmed recurrent DTC after total thyroidectomy who underwent secondary LND. The inclusion criteria were as follows: (I) prior total thyroidectomy ± neck dissection for DTC; (II) secondary dissection with pathological confirmation of DTC metastasis; (III) a TSH-suppressed (0.1–0.5 mU/L) serum Tg measurement within 6 months pre-surgery; and (IV) preoperative high-frequency neck US with available images. The exclusion criteria were as follows: positivity for thyroglobulin antibody (TgAb), thyroid peroxidase antibody (TPOAb), or thyrotropin receptor antibody (TRAb); or missing key data. In addition, telephone follow-up was conducted to confirm the absence of any history of autoimmune thyroiditis in all included patients. Therefore, this is a complete-case analysis with no missing data for any variables used in the primary analysis.

Data collection

Clinical/pathological data included age, sex, and number/regions of metastatic nodes from secondary surgery pathology. The primary outcome was high-volume cervical LNM, defined pathologically as the presence of ≥5 metastatic lymph nodes during the secondary surgery. This threshold was selected based on established guidelines: the 2015 American Thyroid Association guidelines use >5 nodes to define intermediate-risk for recurrence, and the 2023 Chinese guidelines recommend total thyroidectomy for patients with ≥5 metastatic nodes (17,21). Serum Tg was measured using a consistent, high-sensitivity electrochemiluminescence immunoassay (ECLIA, Roche Diagnostics) with a functional sensitivity of <0.9 ng/mL. All samples were obtained 30–45 days after the last radioactive iodine therapy (RAI) under TSH suppression.

US assessment

Preoperative US was conducted using Philips EPIQ systems (Philips, Amsterdam, Netherlands) with linear transducers by sonographers with >20 years of experience. We defined three novel aggregate metrics: TotalLength (sum of maximum diameters of the largest suspicious node in each involved level II–VII), SolidLD (sum of solid component long diameters), and SolidSD (sum of solid component short diameters). Qualitative features of the largest node per level were recorded: microcalcifications, cystic change, hyperechogenicity, abnormal blood flow (peripheral/mixed), and aspect ratio (≤2 considered suspicious).

A data-driven ultrasound score (USscore) was constructed. Given that individual sonographic features showed limited independent predictive value [with least absolute shrinkage and selection operator (LASSO) excluding all features at optimal lambda in exploratory analysis], we constructed an exploratory composite score using a multivariable logistic regression model (LNM ≥5 as outcome) with the five qualitative features. Weights were assigned based on standardized regression coefficients, calculated as (coefficient × feature SD), then divided by the smallest absolute value (0.13) and rounded. Final weights: microcalcifications [2], cystic change [2], hyperechogenicity [1], abnormal blood flow [3], aspect ratio ≤2 [−2]. Thus: USscore = (microcalcifications × 2) + (cystic change × 2) + (hyperechogenicity × 1) + (abnormal blood flow × 3) + [aspect ratio ≤2 × (−2)]. The US feature extraction and scoring were performed without knowledge of the surgical pathology outcome (LNM ≥5). Conversely, the pathologists who determined the surgical outcome were blinded to the US reports.

Statistical analysis

Continuous variables were presented as mean ± standard deviation (SD) or median [interquartile range (IQR)], compared by Student’s t-test or Mann-Whitney U test. Categorical variables were presented as numbers (percentages), compared by chi-square test or Fisher’s exact test.

Univariate logistic regression was conducted to screen for factors (P<0.05) to be inclusion in the multivariable model, presented as odds ratios (ORs) with 95% confidence intervals (CIs). Internal validation used 2,000 bootstrap samples for optimism-corrected area under the receiver operating characteristic (ROC) curve (AUC). Calibration used the Hosmer-Lemeshow test and Brier score. Decision tree analysis determined optimal cut-offs for Tg and TotalLength. Decision curve analysis (DCA) was performed to assess clinical utility. Analyses used R software (version 4.2.1; R Foundation for Statistical Computing, Vienna, Austria). A P value <0.05 was considered statistically significant.

The final predictive models were developed using multivariable logistic regression. The following variable coding rules were applied: serum Tg level was dichotomized as ≥2.57 ng/mL (coded as 1) or <2.57 ng/mL (coded as 0); the ultrasound metric TotalLength was dichotomized as ≥2.35 cm (coded as 1) or <2.35 cm (coded as 0); SolidLD (total solid component long diameter, in cm) and RegionCount (number of involved neck levels) were treated as continuous variables. The USscore, calculated as described in the ‘US assessment’ section, was also entered as a continuous variable.

The general form of the logistic regression model is: logit(P) = β₀ + β₁X₁ + β₂X₂ + … + βₖXₖ, where P is the probability of high-volume LNM (≥5). For clinical application, the predicted probability for an individual patient is calculated using the logistic function: P=11+e−z, where z (the linear predictor) is the sum of the intercept and the product of each predictor value and its corresponding regression coefficient (i.e., z = β₀ + β₁X₁ + β₂X₂ + … + β_kX_k).

The data-driven USscore was categorized into three groups for clinical interpretation: low [≤1], intermediate [2–4], and high [≥5]. No risk groups were created based on the multivariable model’s predicted probabilities. During internal validation, the probability calculation (P=11+e−z) was applied within each bootstrap sample using the coefficients derived from that sample. The final reported predictions and performance metrics were based on applying the coefficients from the original final model to the entire cohort.

Results

Patient characteristics and univariate analysis

This study included 97 DTC patients after total thyroidectomy, all with local-regional recurrence confirmed by secondary surgery pathology. Baseline characteristics stratified by lymph node burden are summarized in Table 1. The median age of the cohort was 36 (IQR, 32–42) years, with a female predominance (67.0%). Age and gender did not differ between high-volume (LNM ≥5, n=38) and low-burden (LNM <5, n=59) groups (P>0.05).

Postoperative Tg was significantly higher in the high-volume group [1.95 (IQR, 0.90–4.00) ng/mL] compared to the low-burden group [1.15 (IQR, 0.38–1.90) ng/mL, P<0.001]. Decision tree analysis identified 2.57 ng/mL as the optimal Tg cut-off [sensitivity 85.3% (95% CI: 75.0–91.8%), specificity 72.4% (95% CI: 54.3–85.3%)], a significantly higher proportion of patients in the high-volume group had Tg ≥2.57 ng/mL (75.0% vs. 25.0%, P<0.001).

All quantitative US indicators differed significantly between groups. TotalLength was significantly larger in the high-volume group [2.25 (IQR, 1.6–2.9) vs. 1.4 (IQR, 1.05–2.1) cm, P<0.001], with 2.35 cm as the optimal cut-off (sensitivity 74%, specificity 73%). TotalLength ≥2.35 cm was significantly associated with LNM ≥5 (73.91% vs. 27.03%, P<0.001). SolidLD (median 2.0 vs. 1.3 cm, P<0.001) and SolidSD (median 1.0 vs. 0.6 cm, P<0.001) also differed significantly. The number of involved neck levels was significantly higher in the high-volume group [2.5 (IQR, 2–3) vs. 1.0 (IQR, 1–2), P<0.001].

Individual sonographic features showed no significant association with LNM ≥5 (P>0.05), but the USscore was higher in the high-volume group [3 (IQR, 2–5) vs. 3 (IQR, 1–3.25), P=0.01].

Model development and multivariate analysis

The results of the multivariate logistic regression analysis are presented in Table 2. In Model 1 (the baseline model), Tg ≥2.57 ng/mL (OR =5.79, 95% CI: 1.33–28.88; P=0.02) and the number of involved neck levels (OR =4.76, 95% CI: 2.06–13.02; P<0.001) were strong independent predictors of high-volume LNM. TotalLength ≥2.35 cm and SolidLD were not significant (P>0.05).

In Model 2 (with USscore), Tg ≥2.57 ng/mL (OR =5.60, 95% CI: 1.31–27.70; P=0.03) and the number of involved neck levels (OR =3.38, 95% CI: 1.54–8.49; P=0.003) remained robust predictors. The USscore itself showed a trend toward significance but did not reach statistical independence (OR per point increase =1.21, 95% CI: 0.94–1.58; P=0.15).

The derived regression coefficients were used to construct the following complete equations for the linear predictor (z) of each model:

z1=−3.00+1.76×(Tg≥2.57ng/mL)+1.01×(TotalLength≥2.35cm)+(−0.92)×SolidLD+1.56×RegionCount[1]

z2=−3.50+1.72×(Tg≥2.57ng/mL)+0.44×(TotalLength≥2.35cm)+(−0.35)×SolidLD+1.22×RegionCount+0.19×USscore[2]

In these models, Tg level was dichotomized as ≥2.57 ng/mL (coded as 1) or <2.57 ng/mL (coded as 0). The variable TotalLength, representing the sum of maximum diameters, was dichotomized using its decision tree-derived optimal cut-off of 2.35 cm (≥2.35 cm coded as 1). SolidLD, the total solid component long diameter in centimeters, and RegionCount, the number of involved neck levels, were both entered as continuous variables. The USscore was calculated as described in the Methods.

To assess multicollinearity among predictors, we calculated variance inflation factors (VIFs) for both models. All VIF values were below 3 (Model 1: range 1.42–2.92; Model 2: range 1.05–2.51), indicating no severe multicollinearity that would compromise model stability.

These equations, combined with the logistic function P=11+e−z, form the precise mathematical basis for the clinical nomograms presented in Figures 1,2, enabling individualized risk estimation.

Figure 1 Nomogram for LNM ≥5 in recurrent DTC patients after total thyroidectomy (Model 1, without USscore). DTC, differentiated thyroid carcinoma; LNM, lymph node metastasis; SolidLD, total solid component length of suspicious lymph nodes; Tg, thyroglobulin; TotalLength, total maximum diameter of suspicious lymph nodes; TSH, thyroid-stimulating hormone; USscore, ultrasound score.

Figure 2 Nomogram for LNM ≥5 in recurrent DTC patients after total thyroidectomy (Model 2). This nomogram is based on Model 2, which adds a data-driven USscore (range: −2 to 8) to the four predictors in Model 1. The USscore is calculated from five sonographic features as detailed in the Methods. Model 1, without USscore; Model 2, with USscore. DTC, differentiated thyroid carcinoma; LNM, lymph node metastasis; SolidLD, total solid component length of suspicious lymph nodes; Tg, thyroglobulin; TotalLength, total maximum diameter of suspicious lymph nodes; USscore, ultrasound score.

Model performance and clinical utility

The performance metrics of the two predictive models are detailed in Table 3. Model 1 had an AUC of 0.84, sensitivity of 62%, and specificity of 90%. Model 2 (with USscore) showed slightly improved discrimination (AUC =0.865), sensitivity of 63%, and specificity of 85% (Figure 3). Hosmer-Lemeshow tests showed χ²=7.58 (P=0.48) for Model 1 and χ²=7.31 (P=0.50) for Model 2, indicating good fit for both models.

Figure 3 ROC curves for predicting high-volume cervical lymph node metastasis in recurrent DTC patients after total thyroidectomy. Model 1, without USscore; Model 2, with USscore. AUC, area under the curve; DTC, differentiated thyroid carcinoma; LNM, lymph node metastasis; ROC, receiver operating characteristic; USscore, ultrasound score.

Bootstrap validation (2,000 samples) showed optimism-corrected AUCs of 0.851 (95% CI: 0.758–0.927) for Model 1 and 0.865 (95% CI: 0.783–0.934) for Model 2. Minimal differences between original and corrected AUCs (0.008 for Model 1, 0.011 for Model 2) and narrow 95% CIs indicated robust performance and low overfitting risk.

After categorizing the USscore into low (≤1), medium [2–4], and high (≥5) groups, a statistically significant difference in the proportion of LNM ≥5 was observed among the groups (χ²=10.32, P=0.006). The proportion was 23.3% (7/30) in the low-score group (n=30), 35.4% (17/48) in the medium-score group (n=48), and 68.4% (13/19) in the high-score group (n=19), showing a clear increasing trend and confirming its risk stratification value.

Calibration analysis showed good agreement between predicted and observed outcomes for both models. For Model 1, the calibration slope was 1 (95% CI: 0.625 to 1.463) and the intercept was 0 (95% CI: −0.533 to 0.564), with a mean absolute error (MAE) of 0.112 and Brier score of 0.154 (Figure 4A). For Model 2, the calibration slope was 1 (95% CI: 0.630 to 1.461) and the intercept was 0 (95% CI: −0.538 to 0.566), with an MAE of 0.076 and Brier score of 0.150 (Figure 4B), indicating smaller deviations between predicted and actual risks.

Figure 4 Calibration curves for LNM ≥5 in recurrent DTC patients after total thyroidectomy. The red dashed line represents the ideal calibration line (perfect agreement between predicted and observed values). DTC, differentiated thyroid carcinoma; LNM, lymph node metastasis; MAE, mean absolute error (lower values indicate marginally improved calibration); USscore, ultrasound score.

DCA (Figure 5) showed that both Model 1 (blue curve) and Model 2 (red curve) demonstrated positive net benefit across threshold probabilities ranging from 0% to approximately 80%, exceeding the net benefit of both the “treat all” and “treat none” strategies. The two model curves showed substantial overlap across most of this range. Within the 60–80% threshold interval, the decline in net benefit was more gradual for Model 2, which maintained marginally higher net benefit values than Model 1. The integrated net benefit (NBci) was −0.298 for Model 2 and −0.362 for Model 1.

Figure 5 DCA for LNM ≥5 in recurrent DTC patients after total thyroidectomy. Both models yield higher net benefit than the default strategies within the clinically relevant threshold range (0–80%). Model 1, without USscore; Model 2, with USscore. DCA, decision curve analysis; DTC, differentiated thyroid carcinoma; LNM, lymph node metastasis; USscore, ultrasound score.

Visual nomograms were constructed for clinical application. Figure 1 (Model 1, no USscore) includes Tg group, TotalLength, SolidLD, and number of involved neck levels; Figure 2 (Model 2) adds the USscore. These tools allow rapid calculation of individual LNM ≥5 risk via summing predictor-specific points.

Discussion

Key findings

This study developed and internally validated a non-invasive multimodal integrating serum Tg and standardized sonographic features for predicting high-volume metastasis (LNM ≥5) in recurrent DTC post-total thyroidectomy. The core findings were that Tg ≥2.57 ng/mL and the number of involved neck levels are robust independent predictors, and the data-driven USscore suggests potential value high-threshold clinical decision-making.

This study reaffirms serum Tg’s predictive value under TSH suppression. We found that serum Tg measured under TSH-suppressed conditions (0.1–0.5 mU/L) retained strong predictive value for high-volume LNM, with an optimal cut-off of 2.57 ng/mL (sensitivity 85%, specificity 72%). This aligns with the biological rationale of Tg as a DTC-specific secretory protein of which the serum level reflects the total mass of viable thyroid tissue, including metastases (3,22). Notably, reported cut-off values for predicting recurrence vary widely (e.g., suppressed Tg 0.5–3.5 ng/mL) across studies due to differences in assay sensitivity and patient populations, creating a lack of unified clinical standards (3,16,23,24). By focusing on patients undergoing secondary surgery for recurrent DTC and using a decision tree model (prioritizing sensitivity to reduce missed high-risk cases), our determined cut-off helps establish a reliable Tg threshold under suppression (22). The optimal Tg cut-off of 2.57 ng/mL was derived from our development cohort using decision tree analysis. While this threshold demonstrated good sensitivity (85%) and specificity (72%) within our dataset, it should be considered exploratory and requires external validation in independent cohorts before being adopted as a definitive clinical standard. The results showed that patients with Tg ≥2.57 ng/mL had a 5.79-fold increased risk (95% CI: 1.33–28.88) of LNM ≥5, providing a quantitative basis for selecting high-risk patients requiring more extensive dissection.

The number of involved neck levels was the strongest predictor. In the baseline model, each additional involved level increased the risk of LNM ≥5 by nearly 5-fold (OR =4.76, 95% CI: 2.06–13.02); after including the USscore, the risk increase per level was 3.38-fold (95% CI: 1.54–8.49). This aligns with the clinical understanding that multi-level involvement portends higher tumor burden and worse prognosis, exemplified by the direct association between a large number of metastatic nodes and decreased disease-specific survival (3,5,9,25). In clinical practice, this indicator is rapidly obtainable via preoperative US and may offer clinical utility once validated. For patients with Tg ≥2.57 ng/mL and ≥2 involved levels, extended dissection may be considered based on this preliminary model.

The data-driven USscore (integrating five qualitative features) suggested potential value in high-threshold decisions. An innovation of this study is the construction of the USscore based on five qualitative sonographic features (microcalcifications, cystic change, hyperechogenicity, abnormal blood flow, aspect ratio ≤2) using standardized regression coefficients. Although individual features were non-significant, the USscore showed a significant univariate association and dose-dependent relationship with LNM ≥5 (incidence 68.4% in USscore ≥5 group vs. 23.3% in ≤1 group, P=0.006), highlighting the utility of composite imaging metrics.

Notably, “aspect ratio ≤2” received a negative weight (−2), contradicting traditional views that a round shape (aspect ratio ≤2) indicates malignancy in initial DTC (3). This unconventional finding reflects the distinct biology of recurrent disease, where high-volume metastases often exhibit irregular shapes due to nodal confluence or extracapsular extension (ENE), whereas isolated small metastatic nodes more common in low-burden disease may retain a round morphology (26). This observation is supported by our data-driven analysis, which confirmed a negative association between “aspect ratio ≤2” and high-volume LNM (standardized coefficient: −0.309). The difference from initial diagnosis standards highlights the unique morphological features of recurrent DTC and underscores the need for recurrence-specific prediction tools rather than applying initial diagnostic criteria. However, this observation should be interpreted cautiously and requires validation in larger, independent cohorts before definitive conclusions can be drawn.

DCA demonstrated that both models provided positive net benefit across a broad range of clinically plausible thresholds (0–80%), outperforming the “treat all” and “treat none” strategies. The substantial overlap between the two model curves indicates that the addition of the USscore does not dramatically alter overall model performance—a finding consistent with the modest improvement in AUC (0.851 to 0.865) and the non-significant P value for the USscore in multivariable analysis. Within this framework, we observed that Model 2 maintained a marginally higher net benefit in the 60–80% threshold range—a region corresponding to the high clinical decision uncertainty in our cohort. The marginal differences in net benefit suggest that the USscore offers incremental, supportive value for risk refinement in clinically ambiguous cases, without fundamentally transforming the model’s utility. This supports the potential role of ultrasound features as supplementary tools for surgical planning, particularly when clinical uncertainty is high.

The comparative analysis (Table 3) reveals a notable trade-off between the two models. Model 2 demonstrates superior discrimination (optimism-corrected AUC: 0.865 vs. 0.851) and improved clinical utility (integrated net benefit: −0.298 vs. −0.362), but shows a slight decline in overall accuracy (0.76 vs. 0.79) and specificity (0.85 vs. 0.90) compared to Model 1. This trade-off was intentionally accepted based on clinical priorities in recurrent DTC. First, our primary goal is improving identification of high-risk patients who benefit most from extended dissection. Second, the modest specificity decline is acceptable because false negatives (missing high-volume disease) carry greater consequences—disease progression and re-recurrence—than false positives (potentially manageable surgical procedures). Thus, prioritizing improved discrimination aligns with optimizing outcomes in this high-risk population.

Comparison with literature

Previous predictive studies for DTC LNM have predominantly focused on treatment-naïve patients, with limited applicability to recurrent disease. A systematic review and meta-analysis by Wu et al. of Chinese real-world studies, integrating 57 predictive nomograms from 2,835 studies, reported C-indices of 0.703–0.960 for models predicting central compartment LNM and 0.702–0.956 for lateral compartment LNM in initial DTC (19). Despite showing good discriminative ability, most models lacked external validation and exhibited significant heterogeneity. More critically, due to differences in tumor biology (e.g., common ENE in recurrences) and anatomical changes (e.g., postoperative scarring interfering with imaging assessment) between initial and recurrent disease, these models cannot be directly applied to secondary surgery decisions (10,26). Bachelot et al. linked serum Tg to the total surface area and volume of metastatic lymph nodes (P=0.002 and P<0.0001), indicating Tg’s potential for quantitative tumor burden assessment, but did not establish a high-volume threshold or integrate imaging features (15).

AI-based studies have achieved high diagnostic accuracy for per-node metastasis detection using deep learning (AUC 0.93–0.95) or radiomics (AUC 0.92–0.95) in multicenter cohorts (27,28). However, these models address a different clinical question—distinguishing metastatic from benign nodes at the individual node level—and require specialized computational infrastructure. In contrast, our model predicts patient-level high-volume disease burden (LNM ≥5) specifically in recurrent DTC using readily interpretable semantic features, offering complementary value for surgical planning. The strengths of our study lie in: (I) focusing on recurrent DTC to address the specific challenge of determining secondary surgery extent; (II) combining Tg (biochemical burden) with US (anatomical extent/morphology) to overcome single-modality limitations; (III) implementing data-driven USscore standardization, combining qualitative and quantitative features to reduce operator subjectivity.

Despite a relatively small sample size (n=97), the model demonstrated good performance (optimism-corrected AUC=0.865), benefiting from a focused endpoint and clinically relevant predictors. Future work could explore hybrid approaches combining semantic features with deep learning-derived scores.

Internal validation and model stability

The internal validation via 2,000 bootstrap resamples provided a robust estimate of model performance while correcting for optimism. The optimism-corrected AUC was 0.851 for Model 1 and 0.865 for Model 2. The minimal optimism (difference of 0.008 and 0.011, respectively) and the narrow CIs indicate that the models’ discriminative ability is stable and the risk of overfitting to our specific cohort is low.

These results confirm that the reported performance metrics (discrimination, calibration) are reliable estimates for settings similar to our development context. However, as an internal validation, this primarily assesses model stability within the source data. External validation is needed to confirm generalizability to broader populations and different clinical environments.

Clinical implications

Based on easily accessible serological and US indicators, our model possesses potential for clinical translation pending external validation. In practical workflow, a ‘two-step’ assessment algorithm could be considered: (I) step 1: preliminary risk stratification. Use Model 1 (Tg + number of involved levels) to identify high-risk patients (Tg ≥2.57 ng/mL and involvement of ≥2 levels) for extended dissection. For example, a patient with Tg ≥2.57 ng/mL and 2 involved levels has a predicted probability of LNM ≥5 exceeding 80% based on the nomogram, potentially supporting consideration of an extended neck dissection. (II) Step 2: borderline case refinement. For intermediate-risk patients (e.g., Tg ≥2.57 ng/mL + 1 involved level, or Tg <2.57 ng/mL + 2 involved levels), further stratify by integrating the USscore: USscore ≥5 suggests high probability of high-volume metastasis, recommending extended dissection; USscore ≤4 allows a more limited dissection (e.g., restricted to levels containing ultrasonographically suspicious nodes) to balance efficacy and safety and potentially reduce complications.

If validated, this algorithm could help quantify the risk of LNM ≥5, potentially avoiding unnecessary surgical trauma for low-risk patients while improving therapeutic thoroughness for high-risk patients, advancing personalized care for recurrent DTC.

To contextualize our models’ performance, we compared them with published radiologist accuracy and our institutional ultrasound diagnostic data. A meta-analysis reported pooled sensitivity of 58% (95% CI: 46–69%) and specificity of 89% (95% CI: 79–95%) for ultrasound diagnosis of nodal metastasis in thyroid cancer (29). At our institution, review of 1,497 patients showed sensitivity of 49.8% (95% CI: 46.2–53.4%) and specificity of 88.4% (95% CI: 86.0–90.5%) for preoperative ultrasound detection of any nodal metastasis. Model 1 achieved sensitivity of 62% and specificity of 90%; Model 2 achieved sensitivity of 63% and specificity of 85%. Both models perform within or slightly above the range of expected radiologist accuracy, suggesting potential as decision-support tools for identifying high-volume disease. However, these indirect comparisons are limited by different clinical endpoints (any metastasis vs. high-volume disease) and lack of head-to-head evaluation.

Limitations

This study has several limitations: (I) its single-center, retrospective design and inclusion of only patients with recurrence after initial total thyroidectomy may introduce selection bias; the model’s generalizability requires validation in multicenter prospective studies with diverse patient populations and clinical settings; (II) the number of outcome events (n=38) relative to the number of predictors results in an events-per-variable (EPV) ratio of 7.6–9.5, which is below the ideal threshold of 10. While bootstrap validation suggested acceptable model stability (optimism-corrected AUC differences <0.011), this EPV ratio indicates that the models may be at risk for overfitting and that the regression coefficients and performance estimates should be interpreted with caution. The sample size is also limited (n=97), with a small subgroup of extremely high-volume metastasis (e.g., >10 nodes, n=6, 6.19%), potentially underestimating the model’s performance in this subgroup; larger samples are needed for optimized subgroup analysis and external validation; (III) it used only a single, static Tg value under TSH suppression, not incorporating dynamic Tg indicators (e.g., doubling time), which may have higher predictive sensitivity and could be integrated into future iterations of the model (30). The derived Tg threshold should therefore be considered exploratory and center-specific, requiring external validation; (IV) the US quantitative/qualitative indicators lacked multicenter, multi-observer agreement validation, as all images were reviewed by a single experienced sonographer. Although intra-observer agreement was high (Kappa >0.85), the reproducibility of the ultrasound scoring system across different operators remains. Standardization of measurements and multicenter studies with multiple observers are needed before clinical implementation; (V) the model primarily predicts the binary outcome of LNM ≥5, not providing continuous prediction of the exact number of metastatic nodes, nor differentiating between central versus lateral neck compartments, which have different surgical risks and strategies. Our use of a binary threshold (≥5 nodes) simplifies what is likely a continuous relationship between nodal count and prognosis. Additionally, qualitative prognostic markers such as ENE were not incorporated, as ENE is difficult to predict reliably on preoperative imaging and was not consistently documented in pathology reports. Future studies with larger cohorts should explore continuous prediction of metastatic node count and integrate both quantitative and qualitative (ENE) features for more comprehensive risk stratification; (VI) US assessment focused on the largest node per level, potentially omitting some morphological details; (VII) the construction of the USscore using standardized regression coefficients rounded to integers represents a heuristic, exploratory approach. LASSO regression excluded all five sonographic features, indicating weak individual associations, and the derived weights should be interpreted cautiously. The composite score’s observed association with the outcome requires validation in larger, independent cohorts using more robust methods such as penalized regression with cross-validation; (VIII) direct comparison with radiologist interpretation was not performed and is needed to evaluate clinical utility.

Future research could focus on the following: conducting external validation to test model generalizability; integrating Tg kinetics, molecular pathological features (e.g., BRAF, TERT), and advanced imaging features (e.g., US elastography); developing more refined predictive tools (e.g., for metastatic count, specific compartment); and creating online calculators to enhance accessibility.

Conclusions

This study developed and validated a non-invasive multimodal model integrating serum Tg and sonographic features to predict high-volume cervical LNM in recurrent DTC post-total thyroidectomy. The model shows moderate to good discrimination (AUC =0.865) and calibration. The proposed two-step algorithm, pending external validation, may inform individualized secondary LND planning. Future multicenter validation and integration of dynamic/molecular biomarkers may further improve precision and translation, potentially supporting more precise management of recurrent DTC.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0082/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0082/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0082/prf

Funding: This study was supported by the China-Japan Friendship Hospital Talent Introduction Project (No. 2019-RC-2).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0082/coif). All authors report that this study was supported by the China-Japan Friendship Hospital Talent Introduction Project (No. 2019-RC-2). The authors have no other 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.The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Ethics Review Board of China-Japan Friendship Hospital (No. 2019-103-K71). Informed consent was waived due to the retrospective nature of the study.

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. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Baloch ZW, Asa SL, Barletta JA, et al. Overview of the 2022 WHO Classification of Thyroid Neoplasms. Endocr Pathol 2022;33:27-63. [Crossref] [PubMed]
3. Ringel MD, Sosa JA, Baloch Z, et al. 2025 American Thyroid Association Management Guidelines for Adult Patients with Differentiated Thyroid Cancer. Thyroid 2025;35:841-985. [Crossref] [PubMed]
4. Liu T, Meng K, Zhu Y, et al. Clinical Treatment Status of Papillary Thyroid Microcarcinoma. Advanced Ultrasound in Diagnosis and Therapy 2025;9:41-6.
5. Guidelines Working Committee of Chinese Society of Clinical Oncology. Guidelines of Chinese Society of Clinical Oncology (CSCO) Differentiated Thyroid Cancer. Journal of Cancer Control and Treatment 2021;1164-201.
6. Cavalheiro BG, Shah JP, Randolph GW, et al. Management of Recurrent Well-Differentiated Thyroid Carcinoma in the Neck: A Comprehensive Review. Cancers (Basel) 2023;15:923. [Crossref] [PubMed]
7. Bosset M, Bonjour M, Castellnou S, et al. Long-Term Outcome of Lobectomy for Thyroid Cancer. Eur Thyroid J 2021;10:486-94. [Crossref] [PubMed]
8. Kuba S, Yamanouchi K, Hayashida N, et al. Total thyroidectomy versus thyroid lobectomy for papillary thyroid cancer: Comparative analysis after propensity score matching: A multicenter study. Int J Surg 2017;38:143-8. [Crossref] [PubMed]
9. Chinese Society of Clinical Oncology (CSCO) diagnosis and treatment guidelines for persistent/ recurrent and metastatic differentiated thyroid cancer working group. Chinese Society of Clinical Oncology (CSCO) diagnosis and treatment guidelines for persistent/recurrent and metastatic differentiated thyroid cancer 2018 (English version). Chinese Journal of Cancer Research 2019;31:99-116. [Crossref] [PubMed]
10. Samaan NA, Schultz PN, Hickey RC, et al. The results of various modalities of treatment of well differentiated thyroid carcinomas: a retrospective review of 1599 patients. J Clin Endocrinol Metab 1992;75:714-20. [Crossref] [PubMed]
11. Scharpf J, Tuttle M, Wong R, et al. Comprehensive management of recurrent thyroid cancer: An American Head and Neck Society consensus statement: AHNS consensus statement. Head Neck 2016;38:1862-9. [Crossref] [PubMed]
12. Hughes DT, Rosen JE, Evans DB, et al. Prophylactic Central Compartment Neck Dissection in Papillary Thyroid Cancer and Effect on Locoregional Recurrence. Ann Surg Oncol 2018;25:2526-34. [Crossref] [PubMed]
13. Levine RA. History of Thyroid Ultrasound. Thyroid 2023;33:894-902. [Crossref] [PubMed]
14. Sipos JA. The History of Thyroid Ultrasound: Past, Present, and Future Directions. Endocr Pract 2024;30:1220-6. [Crossref] [PubMed]
15. Bachelot A, Cailleux AF, Klain M, et al. Relationship between tumor burden and serum thyroglobulin level in patients with papillary and follicular thyroid carcinoma. Thyroid 2002;12:707-11. [Crossref] [PubMed]
16. Mahajan A, Kurki V, Rai P, et al. Post-thyroidectomy ultrasonography versus thyroglobulin as a surveillance tool for locoregional recurrence in patients with differentiated thyroid carcinoma: A single centre 10-year study. Front Endocrinol (Lausanne) 2025;16:1594721. [Crossref] [PubMed]
17. Haugen BR, Alexander EK, Bible KC, et al. 2015 American Thyroid Association Management Guidelines for Adult Patients with Thyroid Nodules and Differentiated Thyroid Cancer: The American Thyroid Association Guidelines Task Force on Thyroid Nodules and Differentiated Thyroid Cancer. Thyroid 2016;26:1-133. [Crossref] [PubMed]
18. Li S, Tang W, Feng M, et al. Ultrasound features and thyroid function of patients with papillary thyroid carcinoma for predicting metastatic cervical lymph nodes with atypical ultrasound features. Front Oncol 2025;15:1628205. [Crossref] [PubMed]
19. Wu Y, Su Y, Zhao Y, et al. Precision of predictive nomograms for lymph node metastasis of thyroid cancer from Chinese real-world study: a systematic review and meta-analysis. Front Endocrinol 2025;16:1617563. [Crossref] [PubMed]
20. Yu Y, Ouyang W, Huang Y, et al. Artificial intelligence-based multi-modal multi-tasks analysis reveals tumor molecular heterogeneity, predicts preoperative lymph node metastasis and prognosis in papillary thyroid carcinoma: a retrospective study. Int J Surg 2025;111:839-56. [Crossref] [PubMed]
21. Chinese Society of Endocrinology, Thyroid and Metabolism Surgery Group of the Chinese Society of Surgery, China Anti-Cancer Association Committee of Head and Neck Oncology, et al. Guidelines for the diagnosis and management of thyroid nodules and differentiated thyroid cancer (Second edition). Chin J Endocrinol Metab 2023;39:181-226.
22. Giovanella L, Milan L, Roll W, et al. Thyroglobulin measurement is the most powerful outcome predictor in differentiated thyroid cancer: a decision tree analysis in a European multicenter series. Clin Chem Lab Med 2024;62:2307-15. [Crossref] [PubMed]
23. Petrovic I, LoPresti J, Fatemi S, et al. Influence of Thyroglobulin Autoantibodies on Thyroglobulin Levels Measured by Different Methodologies: IMA, LC-MS/MS, and RIA. J Clin Endocrinol Metab 2024;109:3254-63. [Crossref] [PubMed]
24. Fatemeh A, Mohammad Ali Y, Hassan MM, et al. The Relationship Between Serum Levels of Thyroglobulin Antibody and the Risk of Recurrence in Patients With Differentiated Thyroid Cancer. Cancer Rep (Hoboken) 2025;8:e70191. [Crossref] [PubMed]
25. Hafez LG, Elkomos BE, El-Shafaei MAM, et al. The risk of central nodal metastasis based on prognostic factors of the differentiated thyroid carcinoma: a systematic review and meta-analysis study. Eur Arch Otorhinolaryngol 2023;280:2675-86. [Crossref] [PubMed]
26. Zhu J, Tian M, Zhang T, et al. Diagnostic value of CT enhancement degree in lymph node metastasis of papillary thyroid cancer: A comparison of enhancement, ratio, and difference. Front Endocrinol (Lausanne) 2023;14:1103434. [Crossref] [PubMed]
27. Abbasian Ardakani A, Reiazi R, Mohammadi A. A Clinical Decision Support System Using Ultrasound Textures and Radiologic Features to Distinguish Metastasis From Tumor-Free Cervical Lymph Nodes in Patients With Papillary Thyroid Carcinoma. J Ultrasound Med 2018;37:2527-35. [Crossref] [PubMed]
28. Abbasian Ardakani A, Mohammadi A, Mirza-Aghazadeh-Attari M, et al. Diagnosis of Metastatic Lymph Nodes in Patients With Papillary Thyroid Cancer: A Comparative Multi-Center Study of Semantic Features and Deep Learning-Based Models. J Ultrasound Med 2023;42:1211-21. [Crossref] [PubMed]
29. Chen YH, Zhang YQ. Diagnostic potential of ultrasonography and computed tomography in differentiating cervical lymph node metastasis of thyroid cancer: a systematic review and meta-analysis. Arch Med Sci 2020;19:965-75. [Crossref] [PubMed]
30. Giovanella L, Garo ML, Albano D, et al. The role of thyroglobulin doubling time in differentiated thyroid cancer: a meta-analysis. Endocr Connect 2022;11:e210648. [Crossref] [PubMed]

(English Language Editor: J. Jones)