Predicting One-Year Survival and Feeding Tube Dependence in HNSCC via Machine Learning
Overview
This study developed machine learning models integrating clinical, blood, and MRI radiomic data to predict one-year overall survival and feeding tube dependence in patients with head and neck squamous cell carcinoma (HNSCC) treated surgically. The models demonstrated promising predictive performance, highlighting the potential of multimodal data integration for personalized prognosis and functional outcome prediction.
Background
Head and neck squamous cell carcinoma (HNSCC) is a prevalent malignancy with significant morbidity and mortality worldwide. Prognosis traditionally relies on TNM staging and HPV status, but these do not fully capture disease heterogeneity. Functional outcomes such as feeding tube dependence critically impact quality of life, necessitating accurate prediction tools. Advances in artificial intelligence enable integration of clinical, laboratory, and imaging data to improve outcome prediction and guide personalized management.
Data Highlights
Outcome
Data Modalities
Model Type
Performance Metrics
Statistical Significance
One-year overall survival
Clinical, blood, MRI radiomics
Random forest classifier
ROC-AUC, PR-AUC, F1-score (median across 10-fold CV)
p < 0.05 (Mann–Whitney U test with Fisher’s method)
Feeding tube dependence at one year
Clinical, blood, MRI radiomics
Random forest classifier
ROC-AUC, PR-AUC, F1-score (median across 10-fold CV)
p < 0.05 (Mann–Whitney U test with Fisher’s method)
Key Findings
Integration of clinical, laboratory, and MRI radiomic features enabled machine learning models to predict one-year survival and feeding tube dependence in HNSCC patients.
Random forest classifiers trained with multimodal data achieved statistically significant predictive performance, as measured by ROC-AUC, PR-AUC, and F1-score.
Handling of missing data via multivariate iterative imputation and class imbalance with SMOTE and Tomek links improved model robustness.
Inclusion of additional postoperative treatments as a binary feature allowed assessment of their influence on prediction accuracy.
Radiomic features were extracted from volumetric segmentations of primary tumors on contrast-enhanced T1-weighted MR images, providing quantitative imaging biomarkers.
SHAP values were used to interpret model decisions, enhancing transparency of feature contributions.
Clinical Implications
The study supports the feasibility of using routinely collected clinical, laboratory, and imaging data combined with machine learning to predict critical outcomes in HNSCC, such as survival and feeding tube dependence. This approach may facilitate personalized risk stratification, optimize rehabilitation planning, and improve resource allocation. Incorporation of such predictive models into clinical workflows could enhance decision-making and patient counseling.
Conclusion
Machine learning models integrating multimodal clinical data show promise in accurately predicting one-year survival and feeding tube dependence in surgically treated HNSCC patients. Further validation and refinement may enable their adoption for personalized prognostication and supportive care optimization.
References
Global Cancer Statistics 2020 -- Head and Neck Cancer Incidence and Mortality
TNM Classification and HPV Status in HNSCC Prognosis
Radiomic Feature Extraction and Machine Learning in Oncology
SMOTE and Tomek Links for Class Imbalance Handling
by Melda Yeghaian, Stefano Trebeschi, Marina Herrero-Huertas, Francisco Javier Mendoza Ferradás, Paula Bos, Maarten J. A. van Alphen, Marcel A. J. van Gerven, Regina G. H. Beets-Tan, Zuhir Bodalal, Lilly-Ann van der Velden
Most people know that tobacco use is a major risk factor for oral and head and neck cancers. Some even understand that excessive alcohol consumption also ups the odds.
A VHA study across 11 vendors finds AI-generated primary care notes score lower than clinician-written notes, with the largest deficits in thoroughness, organization, and usefulness
