Feed

Accurate diagnosis and classification of Central Nervous System (CNS) tumors, particularly meningiomas, pose significant challenges, especially when using DNA methylation profiling data. Limited sample availability and the high dimensionality of methylation data can limit robust analysis and the development of reliable algorithms. This study introduces a novel feature-based-driven augmentation strategy that effectively integrates the underlying probability density distribution, addressing class imbalance and distributional tail issues. We demonstrate the advanced utility of this method when applied to DNA methylation microarray datasets (raw IDAT files derived from archival meningioma tissue analyzed using the Illumina Infinium Methylation EPIC v2.0 BeadChip kit) for meningioma classification using a Wav2Vec2 transformer. Our data-aware approach outperforms a conventional copy-based benchmark augmentation technique, with significantly improved classification accuracies. Models trained on the red channel of methylation data, augmented by our method, achieved near-perfect accuracy of 98.75% (AUC=0.988), with the green channel achieving 75.3% accuracy, outperforming the benchmark augmentation method with 79.2% and 62.5% accuracies for the red and green signals, respectively. Notably, the enhanced performance in identifying intermediate meningiomas, an underrepresented class in our dataset, highlights the efficacy of our proposed augmentation technique.Clinical significance-Our proposed approach holds clinical significance by incorporating data-driven biologically informed augmentation to enhance CNS tumor classification, leveraging the full spectrum of DNA methylation diversity for more accurate tumor subtyping.