Identification of Prognostic Biomarkers for Gastric Cancer Using a Machine Learning Method

Abstract

Background: Gastric cancer (GC) is one of the leading causes of cancer-related deaths worldwide. Therefore, identifying prognostic biomarkers for GC is important to improve the clinical outcomes of patients. Methods: Univariate Cox survival analysis and random survival forest (RSF) were performed on all genes in The Cancer Genome Atlas (TCGA) cohort I (n = 261) to screen for survival-related seed genes. A forward selection algorithm was used to further determine prognosis-related genes using ribonucleic acid sequencing (RNA-seq) or clinically integrated RNA-seq data, followed by the construction of prognostic models. The concordance index (C-index) and Akaike information criterion (AIC) were calculated to identify the optimal model, the performance of which was further validated in TCGA cohort II (n = 109) and the Gene Expression Omnibus series 84437 (GSE84437) cohort (n = 431), and compared with five previous prediction models. Results: Four prognostic models were constructed using the machine learning method. Model 3, based on the RSF model and RNA-seq data, was identified as the optimal model (AIC = 1050.76, C-index = 0.74, p = 2.39 × 10⁻¹³). Compared with models 1, 2, and 4, model 3 showed the highest predictive accuracy in both the internal (C-index = 0.73, p = 1.48 × 10⁻²) and external (C-index = 0.62, p = 0.020) validation cohorts. Receiver operating characteristic curves also confirmed the robust ability of the nine-gene signature in model 3 to assess GC prognosis in both TCGA and GSE84437 cohorts, with all areas under curves over 0.65. Furthermore, the prognostic performance of model 3 outperformed that of the other five existing prediction models (C-index = 0.74, p = 2.39 × 10⁻¹³). Conclusions: We propose a nine-gene marker with high sensitivity and specificity as a powerful tool for predicting the prognosis of GC.