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

Area of Science:

• Endocrinology
• Medical Imaging
• Oncology

Background:

• Thyroid nodules are common, and accurate malignancy risk assessment is crucial for patient management.
• Current assessment often relies on ultrasonographic features and clinical factors, but improved prediction models are needed.

Purpose of the Study:

• To develop and validate prediction models integrating ultrasonographic features and laboratory indicators for thyroid nodule malignancy risk.
• To evaluate the clinical utility of these models in supporting decision-making for C-TIRADS 3-4 nodules.

Main Methods:

• A multicenter retrospective study involving 631 nodules (modeling cohort) and 193 nodules (external validation cohort).
• Logistic regression, random forest, SVM, XGBoost, and LightGBM models were built using ultrasound and laboratory data.
• Model performance was evaluated using AUC in internal and external validation sets.

Main Results:

• Logistic regression achieved an AUC of 0.924 internally and 0.929 externally, demonstrating stable performance.
• In external validation, logistic regression showed significantly higher AUCs than other machine learning models after correction.
• The logistic regression model exhibited good calibration and clinical utility, leading to a nomogram for risk assessment.

Conclusions:

• A logistic regression model using routine clinical and ultrasound data provides a stable, interpretable tool for assessing malignancy risk in C-TIRADS 3-4 thyroid nodules.
• The developed nomogram aids in individualized risk assessment and supports clinical decision-making.