Individualized estimation of benefit from adjuvant docetaxel plus S-1 in stage III gastric cancer using deep learning–based counterfactual survival analysis of the START-2 trial.

Abstract

4095 Background: The START-2 trial demonstrated that docetaxel plus S-1 (DS) was superior to S-1 alone in terms of relapse-free survival (RFS) and overall survival (OS) as adjuvant chemotherapy for patients (pts) with stage III gastric cancer (GC) (J Clin Oncol 2019; Gastric Cancer 2022). We conducted the START-2 AR study, a retrospective analysis, to develop a neural network–based model for predicting RFS and estimating individual-level treatment benefit. Methods: Among 912 pts enrolled in the START-2 trial, 599 provided written informed consent to participate in the START-2 AR study. Pts were randomly divided into training (70%) and test (30%) cohorts. Missing data were addressed using multiple imputation. A deep learning–based Cox proportional hazards model was trained in the training cohort and evaluated in the test cohort. Using this model, we estimated individual-level treatment benefit among pts in the DS group, quantified as the difference in restricted mean survival time (dRMST) between counterfactual survival curves comparing DS with S-1 alone. Subsequently, a LightGBM regressor was applied to identify clinicopathological factors associated with treatment benefit using SHapley Additive exPlanations (SHAP). In an exploratory analysis, the top-ranked variables were used to identify a subgroup with limited additional benefit from DS. Results: Among 599 pts (training cohort, n = 419; test cohort, n = 180), 249 RFS events were observed (174 and 75 events, respectively). The pooled Antolini C-index was 0.686 (95% confidence interval [CI], 0.643–0.726) in the training cohort and 0.628 (95% CI, 0.565–0.697) in the test cohort. The estimated dRMST in the DS group was 2.71 months (95% CI, 2.58–2.84). SHAP analysis of dRMST indicated that a lower number of metastatic lymph nodes, a higher platelet count, a higher number of dissected lymph nodes, and differentiated histology were the key factors associated with a smaller dRMST, indicating limited treatment benefit. When continuous variables were dichotomized at their median values and pts were stratified by the cumulative number of these four factors, the subgroup with < 3 factors (n = 454) showed significantly improved RFS with DS compared with S-1 alone (hazard ratio [HR], 0.67; 95% CI, 0.51–0.88; p = 0.004). In contrast, no significant benefit was observed in the subgroup with ≥3 factors (n = 145; HR, 1.32; 95% CI, 0.70–2.50; p = 0.396). Conclusions: Using a deep learning–based counterfactual dRMST framework, we identified a subgroup of pts with resected stage III GC who derived limited additional benefit from adjuvant DS therapy. Stratification based on four clinicopathological factors may facilitate personalized de-escalation of DS treatment. External validation in independent cohorts is warranted. Clinical trial information: UMIN000011438.