A multicenter, clinically interpretable prediction model for malignancy risk in C-TIRADS 3-4 thyroid nodules.

Objectives: To develop and validate clinically interpretable prediction models that integrate ultrasonographic features and laboratory indicators for malignancy risk assessment in thyroid nodules, and to evaluate their potential value in supporting clinical decision-making. Methods: This was a multicenter retrospective study. 631 thyroid nodules from 478 patients at Yueyang People's Hospital and 193 thyroid nodules from 144 patients at Hunan Cancer Hospital were included. Data from Yueyang People's Hospital was used as the modeling cohort, randomly divided into a training set and an internal validation set (approximately 7:3) based on patient data. Data from Hunan Cancer Hospital served as an independent external validation set. Logistic regression, random forest, support vector machine(SVM), XGBoost, and LightGBM models were constructed based on ultrasound imaging features and clinical laboratory indicators, and their predictive performance was evaluated in the internal and external validation sets. Results: In the internal validation cohort, the logistic regression model achieved an area under the receiver operating characteristic curve (AUC) of 0.924 (95% CI: 0.881-0.959), comparable to the support vector machine (AUC = 0.922) and other machine learning models. In the external validation cohort, the AUCs ranged from 0.871 to 0.929, with logistic regression demonstrating strong and stable discriminative performance (AUC = 0.929, 95% CI: 0.887-0.965). Pairwise comparisons using the DeLong test showed that no statistically significant differences were observed between models in the internal validation cohort after Bonferroni correction. In contrast, in the external validation cohort, logistic regression demonstrated significantly higher AUCs than support vector machine, random forest, XGBoost, and LightGBM after correction. The logistic regression model also demonstrated balanced classification performance, good calibration, and favorable clinical utility in decision curve analysis. A nomogram was constructed based on this model for individualized malignancy risk assessment. Conclusions: This study demonstrates that, under strict variable selection, a logistic regression model based on routine clinical and ultrasound information can provide a stable and clinically interpretable risk assessment tool for C-TIRADS 3-4 thyroid nodules.