A Sequential Data Preprocessing Pipeline for Diabetes Prediction: A Data Leakage Prevention and Dual-Validation Approach

Abstract

Machine learning approaches for diabetes prediction face methodological challenges, including data leakage from preprocessing before data splitting, inconsistent handling of missing values, and class imbalance with varying validation methods. This study presents a systematic approach that prevents data leakage and establishes standardized benchmarks for diabetes prediction. Using the PIMA Indian Diabetes Dataset (768 patients), this study applied a preprocessing pipeline: MICE for missing values (652 missing, 9.43% of data), SMOTE for class balance (500 nondiabetic vs 268 diabetic cases), and z-score normalization for feature scaling. Two feature selection methods identified six important clinical variables: Glucose, Pregnancies, Glucose_BMI, Glucose_Age, BMI, and BloodPressure. Dual validation approaches were employed, single split (80:20) and 5-fold cross-validation, to compare five machine learning algorithms: Random Forest (RF), Multi-Layer Perceptron (MLP), XGBoost, Support Vector Machine (SVM), and Logistic Regression (LR). Experimental results demonstrated that RF achieved the highest accuracy (79.79%) in single split testing, whereas MLP performed best in cross-validation (77.81% accuracy, 84.43% ROC-AUC). All algorithms achieved ROC-AUC scores above 0.80. Cross-validation analysis revealed that RF showed consistent performance across data splits, whereas MLP demonstrated better adaptability to different data conditions.