Stomach cancer is a global health issue, requiring accurate pre­diction models for early diagnosis and effe­ctive treatment. This re­search article presents a new method to predict stomach cancer that tackles challenges such as outlie­rs and overfitting. The approach combines Gradient Boosting Machines (GBMs) with robust Z-Score transformation, model stacking, and Baye­sian optimization. The proposed model employs Z-Score transformation to address outliers ensuring that the data is standardized and robust to outliers. The next step is to select clinical, demographic, and molecular features using statistical analysis and domain expertise. These features are used to train a GBM model. Model stacking is used to improve the accuracy of the GBM model. Model stacking combines multiple GBM models trained on different subsets of the data. This leverages the strengths of individual models and reduces model variance. Lastly, the proposed model employs Bayesian optimization to tune the hyperparameters of the GBM models. Bayesian optimization uses previous evaluations to guide the search process, efficiently exploring the hyperparameter space and identifying configurations that maximize model performance. The proposed model achieved of 89% on the test set and the base model only achieve 85%. The proposed model achieved a high precision and recall indicating it was able to correctly identify patients with and without stomach cancer. The proposed model is effective in predicting stomach cancer and addresses outliers and overfitting. The proposed study can be used for early diagnosis of stomach cancer.