A hybrid sampling algorithm combining M-SMOTE and ENN based on Random forest for medical imbalanced data.

Area of Science:

• Machine Learning
• Medical Data Analysis
• Bioinformatics

Background:

• Medical diagnosis frequently encounters imbalanced data, where misclassifying patients carries a higher cost than healthy individuals.
• Traditional algorithms struggle with imbalanced datasets, often assuming equal misclassification costs and sample sizes.
• Accurate identification of patients in imbalanced medical datasets remains a critical challenge.

Purpose of the Study:

• To address the challenge of imbalanced data classification in medical diagnosis.
• To propose a novel hybrid sampling algorithm, RFMSE, for improved patient identification.
• To enhance classification performance without compromising the accuracy of healthy sample classification.

Main Methods:

• Developed RFMSE, a hybrid algorithm combining Misclassification-oriented Synthetic minority over-sampling technique (M-SMOTE), Edited Nearest Neighbor (ENN), and Random Forest (RF).
• M-SMOTE over-samples minority class (patients) based on RF misclassification rate.
• ENN removes noisy majority class (healthy) samples, followed by RF classification with MCC-based iteration stopping.

Main Results:

• RFMSE demonstrated effectiveness in solving imbalanced data classification problems across ten UCI datasets.
• The proposed method significantly improved F-value and Matthews Correlation Coefficient (MCC) compared to traditional algorithms.
• RFMSE enhances the identification of minority class samples while maintaining high classification performance.

Conclusions:

• RFMSE offers a robust solution for imbalanced data classification in medical diagnosis.
• The hybrid approach effectively balances the dataset and improves diagnostic prediction accuracy.
• This algorithm provides a valuable tool for improving patient identification in clinical settings.