Objective: This study aimed to develop and validate a predictive model incorporating early VRR slope kinetics to predict long-term treatment outcomes.
Methods: This retrospective study included 78 patients with benign thyroid nodules treated with microwave ablation. VRR was measured at 3-, 6-, and 12-months post-ablation. Treatment success was defined as 12-month VRR ≥ 90%. Three slope parameters were calculated: K1 (0-3 months, %/month), K2 (3-6 months), and K3 (6-12 months). Machine learning algorithms (LASSO, Random Forest, Support Vector Machine) were employed for feature selection from baseline characteristics, contrast-enhanced ultrasound parameters, and ablation parameters. Multivariate logistic regression was used to develop the final predictive model.
Results: Overall, 38.5% of 78 patients achieved treatment success. VRR demonstrated progressive increase from 72.2 ± 5.5% (3 months) to 85.4 ± 8.5% (12 months), with decelerating slopes (K1: 0.241 ± 0.018, K2: 0.022 ± 0.014, K3: 0.011 ± 0.006). K1 showed a moderate positive correlation with 12-month VRR (R = 0.420, P < 0.001) and was a predictor of treatment success (AUC = 0.685). Machine learning identified multiple features, with 5 features (enhancement slope, FT4, K1, maximum diameter, TG) consistently selected by all three algorithms. The final logistic regression model incorporating 5 consensus features achieved an AUC of 0.908 (95% CI 0.845-0.971). Internal validation confirmed model robustness (tenfold cross-validation AUC = 0.873; bootstrap AUC = 0.881).
Conclusions: The early slope parameter K1 is a predictor of long-term ablation outcomes. Integration of early response kinetics with baseline characteristics substantially improves prediction accuracy, potentially guiding personalized treatment strategies and early intervention decisions.
Keywords: Machine learning; Microwave ablation; Predictive model; Thyroid nodule; Volume reduction rate.
© 2026. Società Italiana di Ultrasonologia in Medicina e Biologia (SIUMB).