Machine learning models predict survival in lymph-node-positive medullary thyroid carcinoma

BackgroundMedullary thyroid carcinoma (MTC) carries a disproportionately high mortality among thyroid malignancies, and the risk is even greater once metastasis occurs; nevertheless, a dependable prognostic tool for lymph-node-positive MTC patients remains elusive. We aimed to derive and externally validate a machine learning model for predicting 3-year and 5-year overall survival (OS) and cancer-specific survival (CSS) in this high-risk population.MethodsRetrospective cohorts were assembled from the U.S. SEER database (n = 1,071) and Zibo Municipal Hospital (external validation, n = 198). After feature selection (Cox, Boruta, RFE), five algorithms (LightGBM, XGBoost, RF, MLP, KNN) were trained in 70% SEER data and tested in the remaining 30% and in the Chinese cohort. F1-score, MCC, sensitivity, specificity, AUC, calibration curve, and decision-curve analysis were evaluated; model explainability was assessed with SHAP.ResultsIn OS prediction, LightGBM achieved the highest AUC in both time horizons (SEER 3-year 0.833, 5-year 0.892; external 5-year 0.869), with superior accuracy. Calibration curves lay closest to the 45° diagonal, and decision-curve analysis demonstrated the greatest net benefit across clinically relevant risk thresholds. SHAP revealed the absence of surgery as the strongest adverse contributor for OS, followed by advanced age, larger tumour size, higher LNR, radiotherapy and chemotherapy demonstrated adverse effects. The same pattern emerges when predicting CSS. Based on these results, we developed an online calculator for predicting 3- and 5-year OS and CSS in patients with lymph-node-positive MTC.ConclusionLightGBM model provides an accurate, well-calibrated, and clinically useful tool for estimating survival in lymph-node-positive MTC. In addition, the decision to undergo surgery is considered the most important factor in the survival of MTC patients.