Interpretable four-class machine learning prediction of initial I-131 therapy responses in differentiated thyroid cancer

Background Differentiated thyroid carcinoma (DTC) responses to initial post-operative I-131 therapy are heterogeneous. Standard binary models combine incomplete responses into a ‘non-excellent’ category, obscuring transitional states (indeterminate response [IDR] and biochemical incomplete response [BIR]) and outcome-specific drivers. We developed and validated an interpretable four-class machine learning framework to separate these dynamic trajectories and support pre-therapeutic decisions.Materials and methods Data from 948 DTC patients were partitioned into training (n = 663) and test (n = 285) cohorts, with an independent temporal validation cohort (n=150). Four algorithms (Random Forest [RF], eXtreme Gradient Boosting [XGBoost], Light Gradient Boosting Machine [LightGBM] and Multi-Layer Perceptron [MLP]) were evaluated to predict excellent response (ER), IDR, BIR and structural incomplete response (SIR). Performance was assessed via area under the receiver operating characteristic curve (AUC), calibration curves and decision curve analysis (DCA). Shapley Additive exPlanations (SHAP) and multivariable logistic regression quantified category-specific drivers.Results Random forest achieved a micro-average AUC of 0.894 (95% CI: 0.842–0.901) in the test cohort and 0.885 (95% CI: 0.832–0.912) in validation. SHAP analysis revealed that pre-treatment stimulated thyroglobulin (pre-sTg) and stimulated thyroglobulin to thyroid-stimulating hormone ratio (LOG(sTg/TSH)) dominated ER, IDR and BIR predictions (pre-sTg peak weight: 31.9% in BIR), whereas diagnostic whole-body scintigraphy (Dx-WBS) was definitive for SIR (45.1% weight). Distant metastasis on Dx-WBS dramatically elevated SIR risk (OR: 171.89); pre-sTg ≥ 10 ng/mL significantly increased both SIR (OR: 7.18) and BIR (OR: 5.48) risks.Conclusions This four-class framework separates biochemical and structural drivers of post-I-131 responses, providing a practical risk-stratification tool for personalized management before therapy.