Predicting unfavorable response to initial radioactive iodine therapy in differentiated thyroid cancer: an explainable machine learning and survival analysis approach

Background Early post-treatment prognostic assessment may facilitate individualized follow-up in patients with differentiated thyroid cancer (DTC) after initial radioactive iodine (RAI) therapy (RAIT).This study aimed to develop and internally validate a dual-mode machine-learning framework integrating clinical information available from the pre-RAI baseline through the early post-RAI assessment period to predict subsequent unfavorable treatment response and persistence/recurrence-free survival (PRFS). Methods A retrospective cohort of 550 DTC patients who underwent total thyroidectomy followed by initial RAIT was included. Multidimensional clinical metrics, including demographic, pathological, imaging, and serum biochemical data, were collected. Post-therapy imaging and biochemical measurements obtained at the first follow-up approximately 1–3 months after RAIT were treated as early post-RAI predictors, whereas subsequent therapeutic response was evaluated dynamically during longitudinal follow-up according to the 2025 American Thyroid Association (ATA) dynamic risk stratification framework. Patients with excellent response (ER) or indeterminate response (IndR) were classified as having a favorable response, whereas those with biochemical incomplete response (BIR) or structural incomplete response (SIR) were classified as having an unfavorable response. PRFS was defined as the interval from initial RAIT to the first subsequent documentation of BIR or SIR; patients without such an event were censored at their last follow-up. Data were randomly split into training and testing cohorts at a 7:3 ratio. Boruta and Least Absolute Shrinkage and Selection Operator (LASSO) were utilized for feature selection in binary classification, while Random Survival Forest (RSF) feature importance ranking was applied for survival modeling. We developed Decision Tree (DT), Random Forest (RF), Light Gradient Boosting Machine (LightGBM), and Multi-Layer Perceptron (MLP)-based Artificial Neural Network (ANN) models for risk classification, alongside the Cox Proportional Hazards (Cox PH) model, Gradient Boosting Survival Analysis (GBSA), RSF, and Extra Survival Trees (EST) for survival prediction. Model performance was evaluated using the Area Under the ROC Curve (AUC), Concordance Index (C-index), calibration curves, and Decision Curve Analysis (DCA). The Shapley Additive exPlanations (SHAP) framework was employed to interpret the models’ decision-making mechanisms. Results Among 550 patients, 113 (20.5%) were classified into the unfavorable response group. The ANN classification model demonstrated superior performance on the testing cohort, with an AUC of 0.932 (95% CI: 0.888–0.967), accuracy of 0.879, specificity of 0.947, F1-score of 0.677, and the lowest Brier Score (0.091). Its clinical net benefit, as assessed by DCA, exceeded that of other models. Key predictors in the ANN risk classification model included Extrathyroidal Extension (ETE), T-stage, iodine-avid lymph node status (iodine_LN), RAI Dose, Metastatic Lymph Nodes (MLN), pre-radioiodine therapy stimulated thyroglobulin (pre_sTg), and pre-radioiodine therapy thyroglobulin antibody (pre_TgAb). Regarding PRFS prediction, the RSF model achieved the highest C-index (0.886) and a mean time-dependent AUC of 0.910 (95% CI: 0.867–0.940), outperforming other models. The core prognostic factors for the RSF model included pre_sTg, thyroglobulin at first follow-up (Tg_FU1), Pre-radioiodine Therapy Thyroglobulin/TSH Ratio (Pre_RAI_Tg_TSH_Ratio), MLN, RAI Dose, Thyroid Stimulating Hormone at first follow-up (TSH_FU1), and Thyroglobulin Antibody Reduction Rate (TgAbRR). Conclusion In this single-center cohort, the ANN and RSF models demonstrated favorable internal performance for unfavorable-response classification and PRFS prediction, respectively. These findings provide preliminary proof-of-concept evidence for an early post-RAI prognostic assessment framework; however, prospective multicenter external validation is required before its generalizability or clinical utility can be established.