A multiple combined method for rebalancing medical data with class imbalances.

Area of Science:

• Medical Informatics
• Machine Learning
• Data Science

Background:

• Medical datasets frequently exhibit class imbalance, negatively impacting classification model performance and minority class accuracy.
• Accurate identification is critical in medicine due to the high cost of misclassification and potential patient harm.

Purpose of the Study:

• To propose and validate a novel, multiple combined method for rebalancing imbalanced medical data.
• To enhance the accuracy and reliability of classification models in critical medical applications.

Main Methods:

• A hybrid approach combining resampling techniques (Synthetic Minority Oversampling Technique [SMOTE], Undersampling [US]), Particle Swarm Optimization (PSO), and MetaCost.
• Experimental validation using nine medical datasets and decision tree analysis for rule generation.
• Comparison against existing methods to evaluate performance improvements.

Main Results:

• The proposed ensemble learning method significantly improved Area Under the ROC Curve (AUC), recall, precision, and F1-score.
• MetaCost enhanced sensitivity, SMOTE boosted AUC, and US improved sensitivity, F1-score, and reduced misclassification costs in highly imbalanced data.
• PSO-based attribute selection increased sensitivity and reduced data dimensionality.

Conclusions:

• The combined method effectively addresses class imbalance in medical data, leading to more reliable diagnostic models.
• Specific strategies (US for imbalance ratio >9, combined SMOTE/US for <9) are recommended based on imbalance levels.
• The findings support the use of advanced rebalancing techniques for improving patient outcomes in data-driven healthcare.