Skip to main content
AAO-HNSF 2026 Annual Meeting & OTO EXPO Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Scientific Oral Presentation

Head and Neck Surgery

Development of a Machine Learning Model to Predict Treatment Failure in Head and Neck Cancer

Tuesday, October 20, 2026
8:45 AM - 8:51 AM PT
Location: West Exhibit Hall A, Learning Zone 3

    Presenting Author(s)

  • KC

    Kevin Z. Chen

    Research Fellow
    Washington University School of Medicine
    Palatine, IL, United States

    • Co-Author(s)

  • KG

    Kelly Gaudian, BS

    Medical Student
    Washington University School of Medicine, United States

    • Senior Author(s)

  • PZ

    Paul Zolkind, MD

    Assistant Professor of Otolaryngology—Head & Neck Surgery Chief of Otolaryngology, John Cochran VA M
    Washington University School of Medicine, United States

  • Matthew A. Shew, MD

    Assistant Professor
    Washington University School of Medicine
    St. Louis, MO, United States

Disclosure(s):
Kevin Z. Chen:

faculty for this accredited education activity has no relationship(s) with ineligible companies to disclose.

Introduction: Despite advances in multimodal therapy, treatment failure remains a major source of morbidity and mortality in head and neck squamous cell carcinoma (HNSCC). Current risk stratification relies on population-based linear models with predefined clinicopathologic features (e.g. extranodal extension (ENE), perineural invasion (PNI), advanced T stage). While these are pillars to current treatment paradigms, we still observe high rates of recurrence as they may not fully capture complex nonlinear relationships driving treatment response. Improved decision support tools are needed to more accurately identify patients at high risk for treatment failure and guide personalized treatment escalation.

Methods: We conducted a retrospective study of patients with HNSCC (n=993) treated with surgery followed by adjuvant therapy from January 2012 to July 2025. Nested cross-validation (5×3 folds) was used for model development and evaluation. Random Forest, XGBoost, and CatBoost models were trained using Boruta feature selection (90% threshold) with Optuna hyperparameter tuning. Logistic regression using clinical TNM stage and age served as a baseline model. Performance was assessed using mean AUC with 95% confidence intervals and SHAP analysis for interpretability.

Results: 5-year disease-free survival was 85.9% (95% CI 84.4-87.3%). Logistic regression achieved a mean AUC og 0.650. Machine learning models had better discrimination, with XGBoost performing best (AUC 0.744, 95% CI 0.700–0.784), surpassing Random Forest (AUC 0.720, 95% CI 0.678–0.763) and CatBoost (AUC 0.732, 95% CI 0.688–0.773). In the XGBoost model, SHAP bee swarm and importance plots identified ENE, p16 status, PNI, pathologic stage, and time between surgery and radiation as predictors with high importance and clear stratification of feature effects

Conclusions: An XGBoost-based machine learning model outperforms traditional clinicopathologic risk stratification in predicting HNSCC recurrence. These tools enable more precise selection of patients for treatment intensification and improved counseling following definitive surgery.