BLUE RIBBON POSTER IMAGING: Multimodal Prognostic Modeling In HNSCC: The Role of Radiomic Parameter Settings and Graph-Based Feature Selection
Abstract
Purpose
Radiomics models often exhibit unreliable performance and substantial variation across institutions. One aspect that has not been examined thoroughly is how parameter settings—such as normalization scale, outlier handling, and bin width—affect downstream results. In this study, we evaluated how adjustments to these radiomic settings, along with selected clinical variables, influence feature stability and predictive model performance across multiple HNSCC cohorts.
Methods
We conducted a retrospective analysis using 1,648 radiomic features extracted from the gross tumor volumes of 752 patients with HNSCC from three institutions. Pre-treatment CT features were extracted under 20 distinct parameter configurations. We applied five feature selection methods—Graph-FS, Boruta, Lasso, RFE-RF, and mRMR—and three models to predict 2-year survival using radiomics features alone or in combination with clinical variables (age, AJCC stage, T, N). We measured performance using ROC-AUC, with 95% confidence intervals derived from 1,000 bootstrap iterations. Additional metrics included F1 score, accuracy, and Brier score. Feature stability was quantified using Kuncheva’s index. We used our new metric, RobustScore, to summarize model consistency. Feature contributions were evaluated using Consensus SHAP.
Results
Among all pipelines, Graph-FS (Connected) with a Random Forest classifier achieved the strongest performance, particularly under the ns100_ro4_bw15 setting. Incorporating clinical predictors further increased performance, yielding an AUC of 0.82 (95% CI: 0.78–0.86), an F1 score of 0.79, an accuracy of 0.76, and a Brier score of 0.19. Relative to other feature selection methods, Graph-FS showed the most balanced behavior, with strong robustness (RobustScore ≈ 0.61) and moderate stability (Kuncheva ≈ 0.47).
Conclusion
Graph-based feature selection improved reproducibility and predictive performance across a range of radiomic parameter settings, supporting its value as a key component of multimodal prognostic modeling in HNSCC. Adding clinical variables further strengthened model performance, underscoring the benefit of a multimodal approach for HNSCC prognosis.