An interpretable multimodal framework using compact biomarkers and Kolmogorov-Arnold networks improves the early...

Early diagnosis of Alzheimer's disease (AD), especially accurate identification at the mild cognitive impairment (MCI) stage, is crucial for slowing disease progression. Although deep learning has achieved promising performance in AD diagnosis, existing multimodal models often operate as "black boxes," lacking the transparency required for clinical practice and failing to explicitly model deep interactions between imaging and clinical features. To address these limitations, this study proposes an interpretable multimodal framework, namely the Compact Biomarker Kolmogorov-Arnold Network (CBKAN). Specifically, we introduce EHCTNet with disease-specific attention to extract features from 3D MRI data, and innovatively constrain the encoder to output a set of compact biomarkers instead of traditional high-dimensional abstract vectors, mimicking the diagnostic logic of clinicians (e.g., judging brain atrophy). In addition, a hybrid feature Transformer is used to fuse these imaging biomarkers with clinical and genetic data, explicitly capturing complementary relationships across modalities. Finally, the Kolmogorov-Arnold Network (KAN) is adopted as the classifier to effectively model the highly nonlinear characteristics of AD progression. Experiments on 800 subjects from the ADNI dataset show that CBKAN achieves 91.1% accuracy and an F1-score of 0.910 in the AD/MCI/CN classification task, significantly outperforming existing mainstream methods. Statistical analyses validate the effectiveness of each component of the model. The proposed model provides a potentially interpretable and high-performing decision-support framework for early Alzheimer's disease diagnosis, although its cross-cohort generalizability and real-world clinical utility require further validation in independent external datasets.