Diabetes can be managed and avoided through early detection, which is essential to the prevention of serious health problems. This paper is aimed at creating machine learning predictive models to detect people who are at risk of diabetes based on a publicly available dataset. The data includes demographical, clinical, and lifestyle related variables such as age, body mass index (BMI), glucose level, blood pressure and insulin level. An extensive data preprocessing pipeline was used, which included missing values treatment, features scaling, outlier's identification, and class balancing. Several machine learning methods were trained and tested such as Decision Tree and Random Forest using general performance metrics of accuracy, precision, recall, and F1-score. The results of the experiments show that the model of the Random Forest outperformed other methods achieving the high level of prediction of diabetes risks 96.83, which proves the high level of its strength and reliability in determining the risk of diabetes. The results underline the relevance of the methods of ensemblebased machine learning in the medical domain and reveal the prospect of the suggested model as the basis of a creation of the decision-support system to help in the early diagnosis and treatment of diabetes.