A New Approach in Diabetes Diagnosis: Performance Comparison using Machine Learning Algorithms

According to the International Diabetes Federation (IDF), approximately 589 million adults are expected to be living with diabetes in 2025. This number continues to rise due to unhealthy eating habits and poor quality of life. Early and accurate diagnosis is crucial to prevent organ damage and improve patient outcomes. This study aims to develop a method for predicting diabetes without requiring individuals to visit healthcare facilities or undergo laboratory tests, using machine learning (ML) techniques to enable early intervention. A diabetes dataset containing 100,000 patient records was used. The dataset was reduced to 17,000 samples through undersampling. Ten different classification algorithms were applied to perform diabetes prediction. Initially, predictions were made using the full dataset, followed by a second phase where blood test data were excluded to assess the model's performance with fewer features. Among the ten algorithms tested, the BayesNet algorithm achieved the highest accuracy of 91% when using the full dataset. When blood-related features were removed, the BayesNet and Rotation Forest models still maintained a relatively high accuracy of 75.10%. The findings demonstrate that it is possible to predict diabetes with considerable accuracy even without blood test data. This approach can facilitate early diagnosis and preventive measures for diabetes, especially in settings where access to laboratory testing is limited.