Serum Thyroid Biomarkers for Diagnosing Malignant Thyroid Nodules: A Machine Learning Approach with External and Causal Validation

Background: Thyroid nodules (TNs) are common, and accurate malignancy risk stratification is critical for clinical decision-making.Although serum thyroid biomarkers are routinely measured, their contribution to malignancy discrimination remains insufficiently characterized.Thus, we aimed to develop a practical and biologically grounded model for identifying malignant TNs using routinely clinical data.Methods: We retrospectively analyzed 5537 patients initially diagnosed with TNs from two hospitals, including 66.23% with malignant TNs.Diagnostic models were developed using logistic regression and machine learning approaches with internal and external validation.Causal associations were assessed using Mendelian randomization (MR), and biological relevance was further examined based on TCGA data.Results: Younger age (OR = 0.96 [0.95-0.97],P < 0.001) and lower thyroglobulin (Tg) levels (OR = 0.64 [0.61-0.67],P < 0.001) were identified as key risk factors for malignant TNs.The two-factor model achieved AUCs of 0.755 (training), 0.735 (internal validation), 0.76 (temporal validation), and 0.784 (external validation) among various cohorts, and demonstrated incremental value to TI-RADS.An interactive web-based nomogram was developed as a proof-of-concept tool to support clinical utility.MR analysis supported a positive causal relationship between Tg and benign TNs ( = 0.62, P = 0.006), while TCGA data confirmed the inverse association between Tg and malignant TNs (OR = 0.14, P < 0.001). Conclusion:A simple model integrating age and routine serum Tg demonstrates robust and generalizable performance for malignancy risk stratification of TNs.By integrating clinical prediction with causal inference and biological validation, this study provides a complementary tool to existing ultrasound-based risk stratification systems.