An interpretable machine learning model for predicting response to anti-GD2 combined with umbilical cord blood-derived natural killer cell immunotherapy in neuroblastoma

Background Neuroblastoma remains a major therapeutic challenge for children with high-risk or relapsed/refractory disease. Anti-GD2 therapy combined with umbilical cord blood-derived natural killer (UCB-NK) cell therapy represents an emerging strategy, but responses vary substantially. Imaging and bone marrow assessments require specialized resources, creating a need for complementary tools based on routinely available data. However, whether serial routine peripheral blood measurements can support response prediction in this treatment setting remains unclear. Methods Serial peripheral blood measurements from 42 patients were used to derive dynamic coupling features that reflected longitudinal trends and coordinated changes between biomarkers. These features were integrated with static variables to develop machine learning models for response prediction. Seven machine learning models were compared using the same repeated stratified cross-validation framework. Model selection considered discrimination, probability-estimation quality, and interpretability. SHapley Additive exPlanations (SHAP) was used to interpret the final model. Results Considering cross-validated performance, the quality of probability estimates, and clinical interpretability, L2-regularized logistic regression was selected as the final model for response prediction. The model leveraged dynamic peripheral blood features to predict response to anti-GD2 combined with UCB-NK cell immunotherapy, achieving a cross-validated AUC of 0.873 (95% bootstrap CI, 0.735–0.976) and a Brier score of 0.131 (95% bootstrap CI, 0.087–0.182). SHAP analysis characterized the direction and magnitude of feature contributions to model predictions. Conclusion This study developed an interpretable response-prediction model using serial routine peripheral blood tests obtained in clinical practice. The model may complement treatment-response assessment using routinely available data. However, validation in larger, independent cohorts is required before clinical use.