Developed a sparse 10-feature signature (4 clinical + 6 radiomic) for locoregional recurrence prediction in head and neck cancer.

Achieved AUC 0.81 [0.62-0.95] on test set with minimal overfitting (train AUC 0.79).

Clinical features + radiomics significantly outperformed radiomics alone (0.81 vs 0.73).

This interpretable model enables risk-stratified treatment planning and has potential for clinical deployment following multi-center validation and federated learning-based collaborative refinement.