Synthetic Data Augmentation for Imbalanced miRNA Cancer Datasets Using WGAN-GP with a PUF-Based Security Layer for Privacy-Aware Deployment

Class imbalance remains a significant challenge in medical machine learning, particularly in genomic classification tasks involving rare cancer subtypes. In microRNA (miRNA) expression analysis, the underrepresentation of minority classes can bias model training and compromise the reliable classification of clinically relevant cancer subtypes. This study addresses the class imbalance problem in The Cancer Genome Atlas (TCGA) miRNA expression data for three kidney cancer subtypes: TCGA-KIRC, TCGA-KIRP, and TCGA-KICH. The original dataset comprises 1,642,113 miRNA–patient records distributed unevenly across the three subtypes, with 970,596, 547,371, and 124,146 records, respectively. To mitigate this imbalance, we propose three Wasserstein generative adversarial network with gradient penalty (WGAN-GP)-based dataset construction procedures: Progressive Balancing, Direct Balancing, and Pretrained Balancing. Progressive Balancing increases the representation of minority classes to match the majority class, producing a balanced dataset of 2,911,788 records, with 970,596 records per class. Ten machine learning classifiers were evaluated using the pre-balanced-pool protocol, under which synthetic records were generated before cross-validation. Random Forest achieved the highest weighted F1-score of 0.8788 under this protocol. To assess generalization more rigorously, we additionally conducted five-fold patient-level stratified grouped cross-validation, excluding patient identifiers and file-path information from all learning stages and restricting WGAN-GP augmentation to the training portion of each fold. Under this leakage-controlled protocol, Random Forest achieved a weighted F1-score of 49.84% ± 3.11% at the 50% balancing target. This lower score provides a more conservative estimate of performance on unseen patients than the pre-balanced-pool result. Furthermore, a fractional-balancing ablation using a standardized dataset of 15,000 samples identified the 50% balancing target as the best-performing evaluated augmentation level, suggesting a potential trade-off between augmentation volume and synthetic-data quality. For secure deployment, we incorporated a lightweight physical unclonable function (PUF)-based hardware security layer on the Nexys A7-100T field-programmable gate array (FPGA) to support device authentication and on-demand key generation. The hardware module exhibited an approximate latency of 350 clock cycles for PUF response generation, error correction, and hashing; dataset encryption/decryption and end-to-end inference latency were not benchmarked. Overall, this study presents a WGAN-GP-based framework for addressing class imbalance in kidney cancer miRNA classification and highlights the importance of patient-level leakage control when evaluating model generalization. However, the observed performance gap between the pooled and grouped protocols cannot be attributed to leakage alone, because the protocols also differ in training budget. Further validation against external balancing baselines and independent patient cohorts is required to establish the clinical generalizability of the proposed approach.