Development and validation of machine learning-based survival analysis to predict outcome in gastric cancer with...

The previously integrated tumor-inflammation-nutrition (HI-GC) score has demonstrated dynamic monitoring value for recurrence and clinical decision-making in patients with postsurgical gastric cancer (GC). However, its failure to incorporate clinical-pathological factors limits its capacity for baseline risk assessment. This study aimed to develop a model that accurately identifies patients for adjuvant chemotherapy and dynamically evaluates recurrence risk.

This retrospective, multicenter, longitudinal cohort study, spanning nine hospitals, included 7,085 patients with GC post-radical gastrectomy. A baseline prognostic model was constructed using 117 machine-learning algorithms. The dynamic survival decision tree model (dySDT) was employed to combine the baseline model with the HI-GC score.

A Cox regression model incorporating six factors was used to create a nomogram [Harrell's C-index: training cohort: 0.765; 95% confidence interval (95% CI): 0.747, 0.783; validation set: 0.810; 95% CI: 0.747, 0.783], including pT stage, positive lymph node ratio, pN stage, tumor size, age, and adjuvant chemotherapy. The best-performing machine learning model exhibited similar predictive accuracy to the nomogram (C-index: 0.770). For the short-term dySDT at 1 month, the mortality hazard ratios (HRs) for groups IIa, IIb, and III were 2.61 (95% CI: 2.24, 3.04), 5.02 (95% CI: 4.15, 6.06), and 8.88 (95% CI: 7.57, 10.42), respectively, compared to group I. Stratified analysis revealed a significant interaction between adjuvant chemotherapy and overall survival in each subgroup (P<0.001). The long-term dySDT at 1 year showed HRs of 3.25 (95% CI: 2.12, 4.97) for group II, 6.73 (95% CI: 4.29, 10.56) for group IIIa, and 17.88 (95% CI: 10.71, 29.84) for group IIIb.

The dySDT effectively stratifies mortality risk and provides valuable assistance in clinical decision-making after gastrectomy.