MoX: an explainable hybrid deep-learning model for integrating multi-omics data to predict event-free survival in neuroblastoma prognosis

Neuroblastoma, a challenging pediatric cancer, requires effective prediction models to improve patient outcomes. This paper introduces MoX, a hybrid deep-learning model designed to predict event-free survival (EFS) in neuroblastoma patients by integrating both clinical and genetic data. The MoX model combines feature selection via Random Forest with a dual-branch deep learning architecture that processes gene expression and clinical features separately before integrating them for final predictions. It was evaluated using the GSE85047 dataset, encompassing gene expression profiles and clinical attributes. The model’s performance is evaluated through K-Fold Cross-Validation, and SHAP (SHapley Additive exPlanations) is utilized to ensure interpretability by elucidating the contribution of individual features. MoX demonstrated exceptional performance with a training accuracy of 96.79%, validation accuracy of 98.15%, and test accuracy of 98.89%. The model achieved a high Area Under the Curve (AUC) of 0.992 for the test data ROC curve and a Precision-Recall (PR) curve with an AUC of 0.989, highlighting its ability to distinguish between event-free and non-event-free survival effectively. The SHAP analysis revealed that genes such as CD9, GABRB3, FRAS1, and IGF2BP1 were the most influential in predicting outcomes, offering valuable insights into the model's interpretability. MoX's low training and validation loss values (0.082 and 0.099, respectively) and high F1 score of 0.989 underscore its robust performance and reliability. MoX provides a powerful tool for enhancing personalized treatment strategies and guiding future research in multi-omic data integration. This paper marks a significant advancement in neuroblastoma prognosis by presenting a model that not only enhances predictive performance but also provides a high level of interpretability, supporting precision oncology and guiding future research in multi-omic data integration.