DeepSurv-based deep learning model for survival prediction and personalized treatment recommendation in tongue squamous cell carcinoma.

We developed a DeepSurv-based deep neural network for survival prediction and treatment recommendation in tongue squamous cell carcinoma (TSCC). The model was trained on 2,015 patients from the Surveillance, Epidemiology, and End Results (SEER) database (2000-2021) and validated on internal (n = 504) and external (n = 249) cohorts. Using dual residual blocks with batch normalization and a composite Cox-ranking loss function, the model achieved concordance indices of 0.781 and 0.744 in internal and external validation, respectively. Time-dependent AUCs ranged from 0.85 to 0.72 (internal) and 0.76 to 0.71 (external) for 1- to 5-year predictions. Risk stratification demonstrated significant survival differences between predicted risk groups (log-rank P < 0.001). SHAP analysis identified age, tumor size, and histological grade as key prognostic factors. The model effectively identified patients benefiting from adjuvant therapies, though prospective validation is warranted for clinical implementation.