Predicting the prognosis of radical gastrectomy for patients with locally advanced gastric cancer after neoadjuvant chemotherapy using machine learning technology: a multicenter study in China.

Background: Neoadjuvant chemotherapy (NAC) can improve the prognosis of patients with locally advanced gastric cancer (LAGC). However, precise models for accurate prognostic predictions are lacking. We aimed to utilize Cox regression and integrate various machine learning (ML) algorithms to identify and prioritize key factors influencing LAGC overall survival to establish an efficient prognostic prediction model.

Methods: Data from 385 patients with LAGC who underwent NAC followed by radical gastrectomy at two centers between January 2016 and December 2020 were analyzed (internal training set, n = 167; internal validation set, n = 112; external validation set, n = 106). The internal cohort was randomly divided into training and validation sets in a 6:4 ratio.

Results: The support vector machine (SVM) model was identified as the best predictive model (AUC values: internal training set, 0.93; internal validation set, 0.74; external validation set, 0.74), outperforming the ypTNM staging system (AUC values: internal training set, 0.9330 vs. 0.7170; internal validation set, 0.7440 vs. 0.6700; external validation set, 0.7403 vs. 0.6960, respectively). In the internal cohort, patients in the HRG (High Risk Group) had significantly lower mean overall survival compared with patients in the LRG (Low Risk Group) (47.33 vs. 64.97 months, respectively; log-rank P = 0.006) and a higher recurrence rate (48.0% vs. 35.6%, respectively; P = 0.041).

Conclusions: The SVM model predicted postoperative survival and recurrence patterns in patients with LAGC post-NAC, and can address the limitations of the ypTNM staging system through providing more targeted decision-making for individualized treatment.