Integrating MR dynamic radiomics and clinical parameters for machine learning-based prediction of short-term response to induction chemotherapy in nasopharyngeal carcinoma.

BACKGROUND AND PURPOSE: Accurate prediction of short-term response to induction chemotherapy (ICT) in nasopharyngeal carcinoma (NPC) remains a critical unmet clinical need. We aimed to construct a multimodal, interpretable machine learning (ML) framework integrating MR dynamic radiomics (delta radiomics) with clinical parameters to predict ICT response.

METHODS

A multicenter retrospective study enrolled 216 pathologically confirmed NPC patients from two institutions (Xiangyang No. 1 People's Hospital and Xiangyang Central Hospital, January 2015-September 2024; training cohort n=151, test cohort n=65) and an independent external validation cohort of 87 patients (Affiliated Hospital of Hubei University of Chinese Medicine, February 2017-June 2024). A Delta Radscore was computed from 2,637 candidate features extracted from paired pre- and arterial-phase post-contrast enhanced T1-weighted imaging (CE-T1WI) MRI, quantifying intratumoral microvascular dynamics. After two-stage feature selection (intraclass correlation coefficient [ICC] ≥0.80 stability screening followed by least absolute shrinkage and selection operator (LASSO)-random forest ensemble), eight ML algorithms-XGBoost, CatBoost, support vector machine (SVM), k-nearest neighbors (KNN), and Logistic Regression,etc-were systematically evaluated. SHAP (SHapley Additive exPlanations) analysis provided model interpretability. Independent clinical predictors were identified via uni- and multivariable logistic regression and integrated into a dynamic nomogram.

RESULTS

Multivariable logistic regression identified five independent predictors: T stage (OR = 4.29, 95% CI: 1.43-12.82, P = 0.009), lymphocyte count (OR = 2.08, 95% CI: 1.07-4.05, P = 0.031), lactate dehydrogenase (LDH; OR = 1.02, 95% CI: 1.01-1.03, P<0.001), Dynamic Rad_Score (OR = 0.65, 95% CI: 0.57-0.76, P<0.001), and Music Therapy and Psychological Counseling score (MTPC; OR = 0.33, 95% CI: 0.16-0.68, P = 0.026). XGBoost achieved the best performance across all cohorts: training AUC = 0.962 (95% CI: 0.937-0.998), test AUC = 0.861 (95% CI: 0.767-0.954), and external validation AUC = 0.926; significantly outperforming logistic regression (DeLong P<0.05). The nomogram C-index was 0.876, with a calibration Brier score of 0.082. SHAP analysis ranked Dynamic Rad_Score (mean |SHAP|=0.219), MTPC (0.199), LDH (0.177), and T stage (0.156) as the four most impactful predictors. Decision curve analysis confirmed net clinical benefit of XGBoost across threshold probabilities of 20%-85%.

CONCLUSION

An XGBoost-based multimodal interpretable model integrating MR dynamic radiomics with clinical parameters effectively predicts short-term ICT response in NPC, providing a quantifiable, explainable decision-support tool to guide individualized treatment strategies.