Alzheimer's disease risk prediction via perceptual deformable attention generative adversarial network with large foundation models
Predicting the risk of Alzheimer's disease (AD) is fundamental for early-stage intervention. Nevertheless, most methods struggle to extract multi-omics associative patterns due to the limited feature perception and inflexible disease modeling. This paper proposes a novel evolutionary pattern mining framework for precise disease risk prediction. Firstly, large foundational models are employed to automatically construct high-quality features. Second, a perceptual deformable attention mathematical model is proposed, which combines multi-scale sparse attention and deformable attention mechanisms to capture evolutionary patterns of fused multi-omics features. Finally, a Perceptual Deformable Attention Generative Adversarial Network (PDAT-GAN) is developed. PDAT-GAN can precisely simulate the evolutionary procedure of AD using multi-omics data, thereby achieving robust risk prediction and pathogeny extraction for AD. We validate the advanced performance and interpretability of PDAT-GAN on public datasets, underscoring significance of PDAT-GAN in supporting clinical intervention and pathogenetic research. The code of PDAT-GAN can be accessed at: .