Gastric cancer is a highly prevalent malignant tumor with a poor prognosis, and its morbidity and mortality rates remain persistently high throughout the world, posing a serious threat to public health. To improve therapeutic outcomes and patient survival, accurately identifying pathological subtypes and implementing personalized treatment strategies has become a key focus of clinical research. However, traditional diagnostic approaches, such as endoscopy and biopsy, are often invasive and carry the risk of missing the diagnosis. To address these limitations, this study proposes a prediction model based on serum tumor markers and a stacking ensemble learning framework, with the aim of achieving early classification of gastric cancer pathological subtypes using machine learning techniques. A total of 7,382 gastric cancer patient records were included in the analysis. The stacking model was developed by optimizing the weights of the features through a combination of feature engineering and genetic algorithm-based selection. Experimental results demonstrate that the proposed model achieves superior performance in the test set in terms of accuracy, precision, and area under the ROC curve (AUC), significantly outperforming individual baseline classifiers. This study offers a novel non-invasive auxiliary diagnostic approach for gastric cancer and has the potential to reduce the need for repeated gastroscopic procedures.