Graph feature selection for enhancing radiomic stability and reproducibility across multiple institutions in head and neck cancer.

Radiomic biomarkers offer promise for precision oncology. However, their clinical utility is limited by variability from differing imaging protocols and the high dimensionality of radiomics data. Feature selection is key for better interpretability, accuracy, and efficiency, yet traditional methods lack stability and reproducibility. We investigate a Graph-Based Feature Selection (Graph-FS) approach that models feature interdependencies to identify stable radiomic signatures for head and neck squamous cell carcinoma (HNSCC) across institutions. We retrospectively analyzed 1,648 radiomic features extracted from the gross tumor volumes of 752 HNSCC patients from three institutions. After standard preprocessing and applying 36 radiomics parameter configurations to simulate variability, we compared Graph-FS with established methods: Boruta, Lasso, Recursive Feature Elimination (RFE), and Minimum Redundancy Maximum Relevance (mRMR). We evaluated feature selection stability and reproducibility using Pearson correlation, the Jaccard Index (JI), and the Dice-Sorensen Index (DSI) and assessed ranking consistency with Kendall's Coefficient of Concordance (W). Graph-FS achieved higher stability (JI = 0.46, DSI = 0.62, OP = 45.8%) versus baseline methods with JI of 0.005 (Boruta), 0.010 (Lasso), 0.006 (RFE) and 0.014 (mRMR). These results demonstrate that Graph-FS enhances feature stability, reproducibility, and predictive performance. This method could facilitate integration into AI-driven radiomics workflows for reliable, multi-center biomarker discovery.