An interpretable machine learning-based approach: development, validation, and clinical utility for distant metastasis prediction in PTC

Background​ Papillary thyroid carcinoma (PTC) constitutes 80–90% of all thyroid malignancies. Despite its generally favorable prognosis, 20–30% of PTC patients present intermediate/high-risk features, increasing distant metastasis risk. Traditional clinicopathological predictors (e.g., TNM stage, tumor size) have limited accuracy in forecasting metastasis, creating a need for more precise prediction tools.​. Methods​ A total of 2,452 PTC patients (diagnosed 2015–2023, follow-up ≥6 months) were enrolled, with data including clinical, pathological, laboratory, and ultrasound indices. Feature selection integrated LASSO, RFE, and ReliefF, identifying 7 core features. Nine machine learning (ML) algorithms were compared; SHAP analysis was used for interpretability. External validation included 432 patients, and a Django-based prediction website was developed.​. Results​ The LightGBM model exhibited optimal performance: test-set AUC = 0.886, accuracy = 0.887, and external validation AUC = 0.758. SHAP analysis identified extrathyroidal invasion (Mean |SHAP| = 0.1329) and thyroglobulin antibody (TgAb, Mean |SHAP| = 0.0981) as top predictors. Tumor size showed a nonlinear association with metastasis, and the model had a 93.6% negative predictive value (NPV) for excluding low-risk patients.​. Conclusions​ This interpretable ML model outperforms traditional predictors, effectively supporting clinical risk stratification and personalized treatment decision-making for PTC patients, with potential for broad clinical application.​.