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

Background/Objectives: Individualized treatment selection for head and neck cancer requires survival models that use routine clinical variables while accounting for censored time-to-event outcomes. This study developed an interpretable Cox proportional hazards baseline and a DeepSurv framework to estimate mortality risk and explore treatment-specific predictions among radiotherapy-based options. Methods: Clinical data were obtained from the RADCURE collection in The Cancer Imaging Archive. After preprocessing, stage harmonization, exclusion of sparse treatment categories, missing-data assessment, one-hot encoding, and standardization, 3266 patients were analyzed. Cox regression and DeepSurv were evaluated using a held-out 80%/20% split, paired bootstrap confidence intervals for C-index differences, and five-fold cross-validation. For treatment recommendation, treatment modality was hypothetically varied across radiotherapy alone, chemoradiotherapy, and radiotherapy plus epidermal growth factor receptor inhibitor while other covariates were held fixed. Results: On the held-out test set, Cox achieved a C-index of 0.684 and DeepSurv achieved a C-index of 0.695. The absolute difference was 0.011, with a paired bootstrap 95% CI of -0.010 to 0.030, indicating no statistically significant improvement. Five-fold cross-validation showed mean C-index values of 0.688 for Cox and 0.707 for DeepSurv. Cox regression identified older age and advanced tumor stage as higher-risk factors, whereas former and non-smoking status were associated with lower hazard than current smoking. Conclusions: DeepSurv provided only a modest numerical gain over the Cox baseline. The recommendation framework illustrates how survival models can generate treatment-specific risk estimates, but these outputs should be interpreted as decision-support signals rather than causal treatment effects. External validation and prospective evaluation are needed before clinical use.