OPTIMIZATION OF DIABETES PREDICTION USING MACHINE LEARNING MODELS: THE ROLE OF PREPROCESSING AND TEST SIZE VARIATIONS

Abstract

Diabetes mellitus remains one of the fastest-growing non-communicable diseases in the world, with global prevalence pro-jected to rise from 463 million cases in 2019 to 578 million by 2030 and 700 million by 2045. Timely and accurate risk prediction is therefore essential for early clinical intervention. This study exam-ines how data preprocessing strategies and train–test split ratios jointly affect the predictive accuracy of three widely used machine learning classifiers Random Forest, Support Vector Machine (SVM), and Naïve Bayes applied to the Pima Indians Diabetes Database (PIDD). Physiologically implausible zero values in the Glucose, Blood Pressure, Skin Thickness, Insulin, and BMI attributes were treated as missing and imputed using the median owing to the pres-ence of substantial outliers. Three scaling techniques (Min-Max, Standard, and Robust Scaling) were then compared against the un-scaled dataset across three train–test split configurations (0.1, 0.2, and 0.3). The experimental results show that Random Forest con-sistently achieved the highest accuracy across every preprocessing scenario, peaking at 89.61% with Min-Max Scaling and a 0.1 test size. SVM benefited considerably from feature scaling, reaching 87.01% accuracy under the same configuration, while Naïve Bayes performed best on the untransformed data (81.82%) and slightly declined after scaling, confirming its comparative insensitivity to feature magnitude. Further analysis indicates that, because tree-splitting criteria are theoretically invariant to monotonic feature transformation, most of the accuracy gain attributed to “prepro-cessing” in Random Forest is more plausibly explained by the medi-an-imputation step than by scaling itself a distinction not previously articulated in this line of research. These findings demonstrate that the benefit of preprocessing is algorithm-dependent rather than uni-versal, and that smaller test sizes generally favor tree-based ensem-bles by providing more training instances. The study contributes empirical evidence and a practical decision framework for selecting preprocessing pipelines according to classifier characteristics, ra-ther than assuming a one-size-fits-all approach in diabetes predic-tion research.