Development and internal validation of preoperative clinical and ultrasonographic models for cervical lymph node metastasis in papillary thyroid carcinoma
Year: 2026
Type: article
Study design: Observational Study
Abstract: Background: Cervical lymph node metastasis (CLNM) is common in papillary thyroid carcinoma (PTC), yet preoperative ultrasonography alone has limited sensitivity, particularly for central compartment disease. We developed and internally validated preoperative prediction models for CLNM and examined whether a constrained nonlinear machine-learning approach added predictive value beyond regularized logistic regression. Methods: This single-center retrospective prediction-model study included 184 patients with surgically and pathologically confirmed PTC who had preoperative thyroid and cervical lymph-node ultrasonography and postoperative pathological assessment of cervical lymph nodes. The primary outcome was histopathologically confirmed central or lateral CLNM. A fixed set of nine preoperative predictors-age, maximum tumor diameter, number of lesions, ultrasonographic extrathyroidal extension, margin abnormality, taller-than-wide shape, microcalcification, intranodular vascularity, and suspicious/enlarged cervical lymph nodes-was used for both models; univariable P values were not used as predictor-selection thresholds. Ridge logistic regression and a complexity-constrained extreme gradient boosting (XGBoost) model were evaluated using repeated nested stratified cross-validation (outer five-fold cross-validation repeated five times; inner three-fold tuning). Performance was based on patient-level averaged out-of-fold (OOF) predicted probabilities. Results: CLNM was present in 63 of 184 patients (34.2%); 121 patients had no CLNM. Ridge logistic regression achieved an area under the receiver operating characteristic curve (AUC) of 0.766 [95% confidence interval (CI): 0.694-0.836], compared with 0.740 (95% CI: 0.661-0.818) for XGBoost. Brier scores were 0.179 and 0.190, respectively. Calibration intercept/slope were 0.091/1.152 for ridge logistic regression and 0.409/1.570 for XGBoost. At a fixed probability threshold of 0.50, sensitivity/specificity were 0.413/0.901 for ridge logistic regression and 0.317/0.917 for XGBoost. Paired performance differences were uncertain, with CIs including no difference. Conclusions: In this limited single-center cohort, XGBoost did not demonstrate clear added predictive value over ridge logistic regression. Both models had insufficient sensitivity at a fixed 0.50 threshold to support stand-alone surgical decision-making. The findings support exploratory preoperative risk stratification, but larger external cohorts are required for validation and recalibration before clinical implementation.
Source: Gland Surgery
Institution: Xiangnan University
Language: English
Subfield: Endocrinology, Diabetes and Metabolism
Field: Medicine
Domain: Health Sciences
Open Access status: diamond
OpenAlex ID: W7223979588