Personalized Treatment Recommendation System in Head and Neck Cancer Using Survival Analysis and Deep Learning

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Individualized treatment selection for head and neck cancer necessitates accurate survival models using clinical data.
• Accounting for censored time-to-event outcomes is crucial for robust prognostic estimations.

Purpose of the Study:

• To develop and compare an interpretable Cox proportional hazards model and a DeepSurv framework for mortality risk estimation.
• To explore treatment-specific predictions for radiotherapy-based options in head and neck cancer.

Main Methods:

• Utilized the RADCURE dataset (3266 patients) with preprocessing, stage harmonization, and missing data handling.
• Evaluated Cox regression and DeepSurv using an 80%/20% held-out test set, paired bootstrap CIs, and five-fold cross-validation.
• Simulated treatment recommendations by varying radiotherapy modalities while keeping other covariates constant.

Main Results:

• DeepSurv achieved a C-index of 0.695 on the test set, compared to 0.684 for Cox, with no statistically significant difference.
• Cross-validation yielded mean C-indices of 0.707 for DeepSurv and 0.688 for Cox.
• Cox regression identified older age and advanced tumor stage as high-risk factors; current smoking status increased hazard.

Conclusions:

• DeepSurv demonstrated a marginal numerical advantage over the Cox model for head and neck cancer survival prediction.
• The developed framework can generate treatment-specific risk estimates, serving as decision-support tools.
• External validation and prospective studies are essential before clinical implementation of these survival models.