Genotypic Data Sparse Dimensionality Reduction and Variational Autoencoder-Based Feature Extraction

The analysis of epistasis in Alzheimer's disease (AD) is hindered by high-dimensional genotypic data and limited sample sizes, leading to computational inefficiency and overfitting risks. To address this, we propose a hybrid framework integrating sparse dimensionality reduction and Variational Autoencoder (VAE)-based feature extraction. First, a distance correlation-guided matrix groups correlated single nucleotide polymorphisms (SNPs), reducing redundancy. Second, the Intelligent Swarm-guided Sparse Coding Optimization (ISG-SCO) algorithm refines the feature space and significantly outperforms Orthogonal Matching Pursuit (OMP) in terms of convergence speed and optimization efficiency. Finally, a VAE compresses features into a low-dimensional latent space while preserving discriminative genetic information. Comparative experiments show the VAE model achieves superior performance (AUC = 0.67) over Principal Component Analysis (PCA, AUC = 0.43) and t-Distributed Stochastic Neighbor Embedding (t-SNE, AUC = 0.46). This approach effectively balances dimensionality reduction and feature retention, enabling reliable epistasis detection in AD genomic studies.