Synthetic Expansion of Gene Expression Data: Enhancing Predictive Modeling through Augmentation Techniques

Abstract Background Gene expression microarrays offer valuable insights into disease mechanisms, yet their utility for predictive modeling is often constrained by small sample sizes. This study investigates the application of synthetic data augmentation techniques—particularly Gaussian noise injection—to expand two gene expression datasets (GDS3952 and GDS2771) and enhance model performance. Methods Feature selection methods, including ANOVA, Tukey tests, and Linear Discriminant Analysis, were used to reduce dimensionality and isolate discriminative gene subsets. Ensemble classifiers (Random Forest, Bagging, and Voting) were trained on both original and augmented datasets. Gaussian augmentation yielded statistically similar distributions as verified by Kolmogorov–Smirnov tests and significantly improved predictive performance across several tasks. Conclusions Synthetic data augmentation using Gaussian noise led to measurable improvements in model accuracy, with performance gains of up to 6% across tasks. In particular, binary classification of disease status ('healthy' vs. 'cancer') achieved accuracies as high as 98% under Leave-One-Out Cross-Validation (LOOCV). Permutation testing confirmed that these results were not artifacts of overfitting, but reflected the models' ability to learn generalizable patterns from the data. Importantly, the augmented datasets preserved statistical properties of the original data, as demonstrated by non-significant Kolmogorov–Smirnov tests. These findings suggest that even straightforward augmentation techniques, when applied with appropriate validation, can enhance both the accuracy and reliability of predictive models trained on limited gene expression datasets. This approach may be especially valuable in biomedical domains where data collection is costly or constrained.