A self-inspected adaptive SMOTE algorithm (SASMOTE) for highly imbalanced data classification in healthcare.

Area of Science:

• Machine Learning
• Bioinformatics
• Medical Informatics

Background:

• Healthcare datasets often exhibit class imbalance due to rare events like disease onset.
• The Synthetic Minority Over-sampling Technique (SMOTE) is used for imbalanced data but can generate low-quality, ambiguous samples.
• Existing SMOTE methods struggle with sample quality, impacting classification performance in critical healthcare applications.

Purpose of the Study:

• To introduce a novel Self-Inspected Adaptive SMOTE (SASMOTE) model for enhancing synthetic sample quality in imbalanced datasets.
• To improve the accuracy and reliability of machine learning models in healthcare by addressing data imbalance.
• To develop a method that generates higher-quality, separable synthetic minority class samples.

Main Methods:

• Proposed SASMOTE model utilizing an adaptive nearest neighborhood selection algorithm.
• Implemented an uncertainty elimination via self-inspection approach to filter ambiguous generated samples.
• Compared SASMOTE with existing SMOTE-based algorithms on real-world healthcare datasets.

Main Results:

• SASMOTE generates higher-quality synthetic samples compared to traditional SMOTE algorithms.
• Demonstrated improved prediction performance, particularly in F1 score, across two healthcare case studies.
• Effectively addressed challenges of ambiguous and non-separable samples in imbalanced data.

Conclusions:

• The proposed SASMOTE model significantly enhances the quality of synthetic samples for imbalanced healthcare data.
• SASMOTE offers a promising solution for improving machine learning model performance in critical applications like risk gene discovery and disease prediction.
• The self-inspection and adaptive neighborhood selection mechanisms contribute to better classification outcomes on highly imbalanced datasets.