Improving classification on imbalanced genomic data via KDE-based synthetic sampling.

Area of Science:

• Biomedical Machine Learning
• Genomic Data Analysis
• Computational Biology

Background:

• Class imbalance is a significant challenge in high-dimensional genomic datasets.
• Standard machine learning models often exhibit bias towards the majority class.
• This bias is particularly problematic in clinical diagnostics for rare conditions.

Purpose of the Study:

• To introduce a novel Kernel Density Estimation (KDE)-based oversampling method.
• To rebalance imbalanced genomic datasets by generating synthetic minority class samples.
• To address limitations of conventional oversampling techniques like SMOTE.

Main Methods:

• Developed a KDE-based oversampling approach to estimate global minority class distributions.
• Generated synthetic minority class samples to rebalance imbalanced genomic datasets.
• Evaluated the method on 15 real-world genomic datasets using Naïve Bayes, Decision Trees, and Random Forests, comparing against SMOTE and baseline.

Main Results:

• KDE oversampling consistently improved classification performance across datasets and classifiers.
• Significant improvements were observed in imbalance-robust metrics like AUC of the IMCP curve.
• KDE demonstrated superior performance with tree-based models and simplified the sampling process.

Conclusions:

• KDE-based oversampling provides a statistically sound and effective solution for imbalanced genomic data.
• The method enhances fairness and accuracy in medical decision-making.
• Offers a promising alternative to existing oversampling techniques for complex biological data.