Risk stratification for adjuvant radiotherapy in pathologic T3N0 rectal cancer using a DeepSurv-based survival model

The role of adjuvant radiotherapy in patients with pathologic T3N0 rectal cancer remains controversial. Although overall prognosis is favorable, marked heterogeneity exists, and reliable tools to identify patients who may derive meaningful benefit from postoperative radiotherapy are lacking. This study aimed to develop and validate a deep learning-based prognostic model to support individualized radiotherapy decision-making in this population.

We included 1,411 patients with pT3N0 rectal adenocarcinoma from the Surveillance, Epidemiology, and End Results (SEER) database for model development and internal validation. An independent real-world cohort of 118 patients was used for external validation. A DeepSurv-based survival model was constructed using clinicopathological variables to predict cancer-specific survival (CSS). Model performance was evaluated using time-dependent area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations (SHAP) were applied to improve model interpretability. Subgroup analyses were conducted to assess the survival impact of adjuvant radiotherapy across model-defined risk strata.

The DeepSurv model showed good discrimination, with 3-, 5-, and 10-year AUCs of 0.749, 0.739, and 0.769 in the training set and 0.819, 0.753, and 0.711 in the external validation cohort, respectively. Using the prespecified cutoff, risk stratification revealed differential survival patterns according to adjuvant radiotherapy status. Specifically, adjuvant radiotherapy was associated with improved survival in high-risk patients in both the SEER cohort (log-rank P < 0.001) and the external cohort (log-rank P = 0.032), but not in low-risk patients. Building on our prior Cox-based approach, DeepSurv captured nonlinear relationships and interactions to better individualize radiotherapy benefit. SHAP analysis identified carcinoembryonic antigen status as the most influential predictor.

This DeepSurv-based prognostic model effectively stratifies patients with pT3N0 rectal cancer and may help identify individuals who could potentially benefit from adjuvant radiotherapy. A risk-adapted postoperative treatment strategy may support individualized decision-making by prioritizing radiotherapy for high-risk patients while avoiding potential overtreatment in low-risk individuals.