Machine learning-based survival prediction and identification of adjuvant chemotherapy benefit in gastric cancer after first-line neoadjuvant chemotherapy: a multicenter, retrospective, cohort study with external validation

The benefit of adjuvant chemotherapy (ACT) for patients with gastric cancer (GC) after first-line neoadjuvant chemotherapy (NACT) is debated. This study developed and validated a machine learning (ML) model to predict disease-free survival (DFS) and identify patients who benefit from ACT. A total of 1150 patients treated with NACT and radical gastrectomy across four centers in China were retrospectively analyzed. Feature selection and model development employed multiple ML learners. The optimal model was determined using the concordance index (C-index), time-dependent receiver operating characteristic curves, time-dependent calibration curves, and decision curve analysis. ACT efficacy was evaluated in different risk groups using inverse probability of treatment weighting. Eleven ML learners identified eleven feature subsets. Subsequently, eleven feature subsets and eleven machine learning learners were combined, resulting in the development of 121 models. The GAMB-AORSF model, which combines a Generalized Additive Models via Gradient Boosting-selected feature subset with an Accelerated Oblique Random Survival Forest learner, exhibited the highest prediction performance. The model demonstrated robust discrimination with C-indices of 0.864 in the training cohort, and 0.813 and 0.789 in two validation cohorts. The GAMB-AORSF model successfully stratified patients to guide decision-making. High-risk patients derived significant survival benefits from ACT, with a 3-year restricted mean survival time extension of 5–7 months and an absolute recurrence risk reduction of 14–20% across cohorts. Low-risk patients showed no significant survival improvement from ACT. The GAMB-AORSF model has the potential to predict DFS and guide ACT in GC patients who underwent NACT and radical surgery. This tool facilitates postoperative decision-making.