Investigation of Data Balancing Techniques for Diabetes Prediction

Diabetes prediction is critical for early intervention and effective disease management.However, the inherent class imbalance in medical datasets, such as the Pima Indians Diabetes dataset, often leads to biased predictions favoring the majority class.This study provides a systematic analysis of various data balancing techniques applied to diabetes prediction, examining their effects on different classifiers and validation techniques.This study uniquely evaluates the impact of prior balancing before data splitting, offering new insights into real-world deployment scenarios.The Pima dataset was used due to its clinical relevance in diabetes prediction and its widespread use as a benchmark, allowing for robust comparison and reproducibility of results.The developed framework consists of preprocessing, data balancing, classification, and evaluation.The performance of different balancing techniques across various classification algorithms and validation techniques was evaluated using accuracy, precision, recall, and Fmeasure.The results showed that applying data in cross-validation and balancing techniques fails to improve the prediction results, with the accuracy obtained with and without balancing around 90%.The accuracy improved slightly in the train-test percentage split, with the best accuracy of 91%.Finally, when balancing was applied prior to data splitting with cross-validation, the results were improved, as the combined sampling achieved an accuracy of 97.5% and undersampling achieved an accuracy of 94.1%.