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

Class imbalance poses a serious challenge in biomedical machine learning, particularly in genomics, where datasets are characterized by extremely high dimensionality and very limited sample sizes. In such settings, standard classifiers tend to favor the majority class, leading to biased predictions - an especially problematic issue in clinical diagnostics where rare conditions must not be overlooked. In this study, we introduce a Kernel Density Estimation (KDE)-based oversampling approach to rebalance imbalanced genomic datasets by generating synthetic minority class samples. Unlike conventional methods such as SMOTE, KDE estimates the global probability distribution of the minority class and resamples accordingly, avoiding local interpolation pitfalls. We evaluate our method on 15 real-world genomic datasets using three classifiers -Naïve Bayes, Decision Trees, and Random Forests- and compare it to SMOTE and baseline training. Experimental results demonstrate that KDE oversampling consistently improves classification performance, especially in metrics robust to imbalance, such as AUC of the IMCP curve. Notably, KDE achieves superior results in tree-based models while dramatically simplifying the sampling process. This approach offers a statistically grounded and effective solution for balancing genomic datasets, with strong potential for improving fairness and accuracy in high-stakes medical decision-making.