Preoperative neck CT scans were used to identify the thyroid isthmus, the medial and lateral tumor margins, and the tumor centroid. The following definitions were applied: the tumor centroid was defined as the intersection of the tumor’s longest and shortest diameters; the thyroid isthmus was defined as the thyroid tissue anterior to the trachea connecting both thyroid lobes. Tumor location was classified according to the method described by Lim et al. (9): on axial CT images, vertical projection lines were drawn from the most lateral points of the trachea to the anterior skin surface. Tumors were categorized as: isthmus thyroid cancer, if the centroid fell between the two projection lines; isthmus-proximity PTC, if the centroid lay outside the projection lines but the medial tumor margin remained between them; or normally located thyroid cancer, if both the centroid and medial margin lay outside the projection lines (Figure 1). The central lymph node compartment was defined in accordance with the 2015 American Thyroid Association guidelines (10), which include the pretracheal, paratracheal, and prelaryngeal nodal groups. For statistical analysis, pretracheal and paratracheal lymph nodes were grouped and analyzed as part of the Ipsi-CLNs region.

This retrospective study enrolled patients with solitary PTC who underwent surgery at the Affiliated Hospital of Jiangsu University between January 2012 and June 2025. Inclusion criteria were: (1) postoperative pathological confirmation of PTC; (2) presence of a single tumor focus; (3) first-time surgery consisting of total or near-total thyroidectomy with bilateral central lymph node dissection; and (4) availability of complete preoperative imaging including thyroid ultrasound and neck CT. Exclusion criteria comprised: (1) tumors located in both lobes or the isthmus; (2) multifocal lesions within the same lobe; (3) previous history of other malignant neck tumors; (4) locally advanced PTC (stage T4a or higher) with invasion of critical structures such as the trachea, esophagus, prevertebral fascia, carotid artery, or mediastinal vessels; and (5) incomplete clinical or follow-up records that would preclude accurate analysis.

The following clinical and pathological variables were recorded and analyzed: age, gender, laterality of the tumor, presence of Hashimoto’s thyroiditis, BRAF V600E mutation status, findings from preoperative ultrasound and CT assessments, status of Ipsi-CLNs, aggressive pathological features (if present), tumor size, and extrathyroidal extension (ETE) (11).For preoperative imaging evaluation, lymph nodes were considered suspicious for metastasis on ultrasound based on established criteria: loss of the fatty hilum, cortical hyperechogenicity, calcifications, cystic changes, or abnormal peripheral vascularity (12). On CT, suspicion was raised by features including heterogeneous enhancement, necrosis or cystic degeneration, short-axis diameter ≥1.3 cm, clustering (≥3 nodes), irregular contours, indistinct margins, or fine granular calcifications (13).All ultrasound and CT images were reviewed independently by two radiologists, each with more than five years of subspecialty experience in thyroid imaging. Interobserver agreement was assessed. Any discrepancies in interpretation were resolved through a consensus review involving a third senior radiologist, whose judgment served as the final reference standard.

Patients were first categorized as having isthmus-proximity or non−isthmic PTC based on tumor location, and the rates of contralateral central lymph node metastasis were compared between the two groups. From the isthmus−proximity cohort, patients were randomly divided into a training set (70%) and a test set (30%). Feature selection was performed within the training set using both LASSO regression and the Boruta algorithm. Variables identified by both methods were included in the final predictive model. Seven machine learning algorithms were implemented: logistic regression, decision tree, random forest, XGBoost, LightGBM, support vector machine (SVM), and artificial neural network (ANN). During model development, hyperparameters were optimized via five−fold cross−validation coupled with grid search, aiming to maximize the AUC. To address class imbalance, the synthetic minority oversampling technique (SMOTE) was applied separately within each cross−validation training fold; the corresponding validation fold remained unaltered to avoid data leakage. After hyperparameter tuning, each model was refitted on the entire training set and evaluated on the hold−out test set. Performance was assessed using accuracy, sensitivity, precision, F1−score, and AUC. For the best−performing model, SHAP (Shapley Additive Explanations) analysis was used to interpret feature importance. SHAP summary plots and dependence plots were generated to illustrate the direction and magnitude of each variable’s contribution to model predictions.

Data were processed and analyzed using R (version 4.4.2). Variables with missing values exceeding 30% were excluded from the analysis. For the remaining variables with missing data, multiple imputation was performed using the mice package in R. Continuous variables with normal distribution are presented as mean ± standard deviation and were compared between groups using the independent samples t−test. Non−normally distributed continuous variables are reported as median (interquartile range) and were compared using the Mann−Whitney U test. Categorical variables are expressed as frequency (percentage) and were compared using the chi−square test or Fisher’s exact test, as appropriate. All statistical tests were two−sided, with a significance level set at α = 0.05.

Between January 2012 and June 2025, a total of 2,711 patients with papillary thyroid carcinoma (PTC) were initially assessed. After applying the exclusion criteria, 1,672 patients were included in the final analysis (Figure 2). Reasons for exclusion were as follows: bilateral or isthmic PTC (n=358), unilateral multifocal carcinoma (n=129), previous history of neck malignancy (n=20), locally advanced disease (n=4), and incomplete clinical data (n=528). Of the 1,672 included patients, 397 were diagnosed with unilateral PTC with isthmus proximity, and the remaining had PTC in a typical (non-isthmic) location. The rate of Cont-CLNs metastasis was significantly higher in the isthmus-proximity group than in the non-isthmic group (P < 0.05; Table 1). These 397 patients with unilateral isthmus-proximity PTC were then randomly divided into a training set (n=278, 70%) and a test set (n=119, 30%) in a 7:3 ratio (Table 2). No statistically significant differences were observed in baseline characteristics between the training and test sets, confirming the comparability of the split.

The feature selection results from both LASSO regression and the Boruta algorithm consistently identified five key predictors: preoperative CT assessment, extrathyroidal extension, preoperative ultrasound assessment, Ipsi-CLNs metastasis, and tumor size (Figure 3).

Based on the five selected predictors, seven machine learning algorithms were applied to build prediction models: logistic regression, decision tree, random forest, XGBoost, LightGBM, support vector machine, and artificial neural network. Model performance was systematically evaluated on the test set using the AUC, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), precision, recall, and F1-score (Table 3).

The random forest model showed the best overall performance, attaining the highest F1-score and AUC values of 0.942 in the training set and 0.861 in the validation set. Furthermore, decision curve analysis confirmed its favorable clinical net benefit across a range of threshold probabilities, supporting its practical utility in clinical decision-making (Figure 4).

To interpret how the random forest model predicts Cont-CLNs metastasis in patients with isthmus-proximity PTC, SHAP analysis was performed. The five most influential predictors, ranked in descending order of SHAP importance, were: preoperative CT assessment, extrathyroidal extension, Ipsi-CLNs metastasis, preoperative ultrasound assessment, and tumor size (Figure 5A). This order suggests that imaging and pathological markers of local invasion and nodal involvement carry greater predictive weight than tumor size alone.

SHAP summary and dependence plots (Figures 5B, C) illustrated the direction and magnitude of each feature’s effect. A clear “dose–response” pattern was observed for preoperative CT assessment: negative findings corresponded to negative SHAP values, lowering the predicted probability of metastasis, whereas positive findings produced strongly positive SHAP values, making this the strongest driver of the model’s risk prediction. Extrathyroidal extension and Ipsi-CLNs metastasis showed a similar directional influence, with presence of either feature substantially increasing predicted risk, though the effect of Ipsi-CLNs was slightly weaker. Preoperative ultrasound assessment followed the same trend, with positive findings raising the estimated probability of metastasis. By contrast, tumor size exhibited a more complex, nonlinear relationship with predicted risk. Although larger tumors (especially >10 mm) were weakly associated with increased risk, size alone was not a stable or strong independent predictor in this model. These interpretability findings align closely with established clinical understanding, supporting the biologic plausibility and practical relevance of the model.