Prediction of Diabetes Using Machine Learning on a Large Clinical Dataset with an Orange No-code Artificial Intelligence Tool

Background: Diabetes mellitus is increasing worldwide, and many individuals remain undiagnosed until complications arise, underscoring the need for scalable early risk prediction tools that go beyond conventional laboratory-based diagnosis.This study aimed to develop and evaluate supervised machine learning (ML) models for diabetes prediction using a publicly available diabetes prediction dataset of approximately 1,00,000 records comprising demographic and clinical variables such as age, sex, body mass index (BMI), hypertension, heart disease, smoking history, glycated hemoglobin (HbA1c), and blood glucose level, implemented through visual workflows in the Orange data mining platform.Materials and methods: Data preprocessing included cleaning missing or inconsistent values, encoding categorical variables, and normalizing continuous features, followed by internal 10-fold cross-validation to ensure robust performance estimation.Four classifiers were compared, namely Random Forest, Logistic Regression, support vector machine (SVM), and adaptive boosting, selected to balance interpretability, robustness, and ensemble learning strengths.Results: Among these, Random Forest and Logistic Regression demonstrated superior and clinically meaningful performance, achieving accuracy above 96%, high recall and F1-score, and strong discrimination, with Logistic Regression yielding the highest area under the receiver operating characteristic curve of 0.962, while SVM showed limited scalability and markedly lower discrimination.Visualization through confusion matrices and receiver operating characteristic curves confirmed the ability of Random Forest and Logistic Regression to identify diabetic cases while minimizing false negatives, which is critical for early detection. Conclusion:The findings highlight that accessible, no-code ML platforms such as Orange can enable health care professionals to construct reproducible predictive workflows without advanced programming expertise, and that classical, interpretable models can deliver state-of-the-art performance when trained on large, well-preprocessed datasets.The absence of external validation, clinical implementation, and prospective evaluation is an important limitation that may affect generalizability.However, the work establishes a practical baseline framework for diabetes risk prediction that can be extended with imbalance handling techniques, additional clinical features, and deployment into clinical decision support systems to support preventive care and reduce the global burden of diabetes.