Screening for gastric cancer (GC) is critical for improving patient survival, yet current clinical methods suffer from invasiveness and limited sensitivity. The aim of this study is to develop a machine learning model for the non-invasive diagnosis of GC using the whole-genome methylation profile of circulating free DNA (cfDNA) in plasma. Plasma cfDNA from 61 patients with GC and 29 healthy controls. First, whole‑genome bisulfite sequencing (WGBS) was performed on 29 paired GC and adjacent normal tissues to identify differentially methylated regions (DMRs). These tissue DMRs were then mapped to plasma cfDNA methylation data, yielding 32 stable DMRs that changed consistently in both tissues and plasma. Using LASSO and Random Forest (RF) algorithms, eight core diagnostic markers were selected. Four machine learning algorithms (Support Vector Machine, RF, K-Nearest Neighbor, Naïve Bayes) were used to build diagnostic models. WGBS revealed global hypomethylation and CpG island hypermethylation in GC, consistent with plasma cfDNA methylation patterns. The RF model built on the eight cfDNA methylation markers (annotated to genes including ATP8A2) performed best, achieving AUC values of 0.978 and 0.938 in the training and independent test sets, respectively, with an accuracy of 88.46%, sensitivity of 94.44%, and specificity of 75.00%. After adjusting for age, the model achieved an AUC of 0.924 on the test set, still demonstrating high discriminatory performance. Functional analysis implicated these markers in pathways such as phagocytosis and axon guidance, and showed significant correlation with immune cell infiltration. Notably, these core DMRs are not covered by the probes on the methylation arrays (450K and EPIC v2), highlighting the advantage of WGBS for novel biomarker discovery. In conclusion, tissue-plasma concordance analysis identified a panel of plasma cfDNA methylation biomarkers with strong diagnostic potential. The resulting RF model provides a promising non-invasive strategy for GC screening.