FIUS: Fixed partitioning undersampling method.

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Disease prediction modeling

Background:

• Predictive modeling for diseases like type 2 diabetes is crucial.
• Imbalanced datasets in electronic medical records (EMR) hinder accurate patient classification.
• Misclassification of high-risk patients is a significant challenge in diabetes prediction.

Purpose of the Study:

• To propose a novel data balancing method for imbalanced EMR datasets.
• To enhance the prediction performance of type 2 diabetes mellitus.
• To address the limitations of class imbalance in medical data analysis.

Main Methods:

• Developed a novel undersampling technique using a fixed partitioning distribution scheme.
• Applied the method to imbalanced electronic medical records (EMR) data.
• Retained valuable information during the data balancing process.

Main Results:

• Achieved an 80% Area Under the Curve (AUC) with logistic regression (LR) classifier.
• The proposed undersampling method outperformed existing techniques.
• Significantly improved LR classifier performance on imbalanced EMR data.

Conclusions:

• The novel balancing method demonstrates high effectiveness in predicting diabetes.
• The approach successfully addresses class imbalance issues in Canadian EMR datasets.
• This methodology is applicable to other domains facing imbalanced data challenges.