Prognostic model for gastric cancer integrating immune biomarkers and machine learning approaches

Prognostication for advanced gastric cancer remains sub-optimal because conventional staging systems overlook systemic inflammation and tumour-immune contexture. We investigated whether incorporating routine immune biomarkers into machine-learning survival models improves risk stratification. A TRIPOD-compliant, single-centre cohort of 300 patients with histologically confirmed gastric adenocarcinoma (2020–2023) was analysed. Baseline variables comprised demographics, TNM stage, C-reactive protein (CRP), and four immune markers assayed on archival tissue: PD-L1 combined-positive score (CPS), microsatellite instability (MSI), tumour mutational burden (TMB), and quantitative CD8⁺-cell density. Missing values (< 15%) were multiply imputed. Feature selection combined univariable screening ( P < 0.10), clinical judgement, and LASSO. An eight-variable multivariable Cox model and a tuned Random Survival Forest (RSF) were developed; performance was assessed using optimism-corrected C-index, time-dependent AUC, calibration, decision-curve, and clinical-impact analyses. Age, T stage, N stage, and CRP were independent adverse predictors, whereas PD-L1 CPS ≥ 1, MSI-H, TMB-high (≥ 10 mut Mb⁻¹), and high CD8⁺ density were protective. The combined Cox model achieved an optimism-corrected C-index of 0.78, representing a 0.05 improvement over the clinical-only model ( P < 0.001). RSF further improved discrimination to 0.82, with AUCs of 0.85, 0.83, and 0.81 at 12, 24, and 36 months, respectively. Decision-curve analysis showed that the RSF conferred a net-benefit gain of 0.06 at a 30% treatment threshold, equating to six additional correctly managed patients per 100. Kaplan-Meier curves confirmed significant survival separation for RSF-defined risk strata and for each immune marker (all log-rank P ≤ 0.041). Integrating PD-L1, MSI, TMB, and CD8⁺ density with standard clinicopathological factors yields materially better survival prediction for gastric cancer, particularly when modelled with an explainable machine-learning approach. These readily assayable biomarkers could enable more precise prognostic stratification and treatment individualisation once validated externally.