Improving Diabetes Prediction with a Structured Framework for Imbalanced Data

Handling imbalanced data in many real-world applications has always been a challenge, particularly in healthcare, where disease prediction is often inaccurate for such datasets. In this paper, we propose a novel approach for handling datasets with uneven class distribution, focusing on diabetes prediction as a case study. We worked on a real-life dataset for type-2 Diabetes prediction containing only 6% of diabetic instances. This underrepresentation of the positive class results in a greater challenge for traditional machine learning models to predict, which, in turn, leads to biased predictions and reduced model performance. To address this issue, we designed a custom architecture to optimize the training set where after splitting the dataset, the training set was partially balanced by dividing the majority class into subsets. Under sampling was applied for further refinement, followed by hyperparameter tuning, which improved the model performance. In this study, five classifiers, GradientBoost, CatBoost, AdaBoost, XGBoost, and Random Forest, were employed to evaluate the model’s performance using key evaluation metrics, including F1-score, precision, recall, and accuracy. Notably, a significant improvement was observed in recall and F1 score compared with the baseline model, highlighting the model’s enhanced ability to identify positive cases correctly. This study contributes to a broader understanding of how advanced techniques for handling imbalanced data can significantly enhance the performance of predictive models, thus offering valuable insights for healthcare practitioners in early disease diagnosis.