An interpretable generative multimodal neuroimaging-genomics framework for decoding Alzheimer’s disease

Background: Alzheimer’s disease (AD) is the most prevalent form of dementia, with a prodromal stage known as mild cognitive impairment (MCI), where patients may either progress to AD or remain stable. Understanding the structural and functional modulations of the brain during this progression is critical. Multimodal MRI data combined with single nucleotide polymorphisms (SNPs) can provide valuable insights; however, missing data and interpretability remain key challenges. Methods: We developed a multimodal deep learning (DL)–based classification framework integrating MRI and SNP data to classify AD patients versus healthy controls and to predict MCI conversion. A generative adversarial network (GAN) module using cycle consistency was introduced in the latent space to impute missing data, addressing a common issue in multimodal analysis. An explainable AI method was employed to extract feature relevance, enabling post-hoc validation and improving interpretability of the learned representations. Results: Experimental evaluation on two tasks, AD detection and MCI conversion, demonstrated competitive state-of-the-art performance, achieving accuracies of 0.926 ± 0.02 (CI [0.90, 0.95]) and 0.711 ± 0.01 (CI [0.70, 0.72]), respectively. Interpretability analysis revealed gray matter modulations in cortical and subcortical brain regions commonly associated with AD. Additionally, impairments were observed in sensory-motor and visual resting-state networks, and genetic mutations linked to endocytosis, amyloid-beta, and cholesterol pathways were identified. Conclusion: The proposed integrative and interpretable DL framework effectively handles missing multimodal data and enhances biological interpretability. It demonstrates strong performance in AD detection and MCI prediction, offering valuable insights into the structural, functional, and genetic underpinnings of Alzheimer’s disease.