Balanced Multiclass Diabetes Dataset

This refined dataset is originally based on the "Diabetes Dataset" uploaded by Ahlam Rashid in Mendeley Data. The link to the original dataset is:https://data.mendeley.com/datasets/wj9rwkp9c2/1. The original dataset contains a total of 1000 subjects divided into three classes: diabetic, non-diabetic, and predict-diabetic.

Among the 1000 subjects, 844 are diabetic, 103 are non-diabetic, and 53 are predict-diabetic, resulting in an extreme class imbalance. We found a total of 170 duplicate and 5 mislabeled subjects in the original dataset, which we subsequently removed or correctd. After initial processing, there were 694 diabetic, 96 non-diabetic, and 40 predict-diabetic subjects remaining.

Following initial preprocessing, SMOTE oversampling was applied to generate 2,079 relatively balanced samples. Subsequently, Tomek Links undersampling was employed to eliminate 2 boundary noise instances, yielding a balanced dataset comprising 692 diabetic, 694 non-diabetic, and 691 predict-diabetic samples.