ESTRO 2025 - Abstract Book

S1114

Clinical – Head & neck

ESTRO 2025

In the independent testing cohort for OS, the ANN identified 46 high-risk and 38 low-risk patients, showing a significant survival difference (p = 0.006) with ROC-AUC of 0.67 and PR-AUC of 0.62. SHAP analysis identified ECOG status, kidney function (eGFR), and HPV-status, nodal and tumor staging as key predictive factors for OS with positive SHAP values of 0.058, 0.045, 0.038, 0.035 and 0.026 respectively. For PFS, 51 patients were classified as high-risk and 38 as low-risk; no significant survival difference was observed (p = 0.166), with a ROC-AUC of 0.62 and PR-AUC of 0.70. SHAP analysis highlighted HPV-status, eGFR, ECOG and nodal status as important factors. Conclusion: Our machine learning stratification model may aid clinical decisions for elderly HNSCC patients by effectively predicting OS, though it did not achieve significant risk separation for PFS. The OS ANN successfully identified high- and low-risk groups with considerable survival differences, and SHAP analysis highlighted key prognostic factors for stratification. These findings offer valuable insights into survival risks, aiding in the development of more tailored treatment strategies for elderly HNSCC patients.

Keywords: Artificial Neural Network, HNSCC, elderly