A lightweight triple-modal fusion network for progressive mild cognitive impairment prediction in Alzheimer's disease.

Introduction: As a progressive neurodegeneration, Alzheimer's disease (AD) represents the primary etiology of dementia among the elderly. Early identification of individuals with mild cognitive impairment (MCI) who are likely to convert to AD is essential for timely diagnosis and therapeutic intervention. Although multimodal neuroimaging and clinical data provide complementary information, existing fusion models often face challenges such as high computational complexity and limited interpretability. Methods: To address these limitations, we introduce TriLightNet, an innovative lightweight triple-modal fusion network designed to integrate structural MRI, functional PET, and clinical tabular data for predicting MCI-to-AD conversion. TriLightNet incorporates a hybrid backbone that combines Kolmogorov-Arnold Networks with PoolFormer for efficient feature extraction. Additionally, it introduces a Hybrid Block Attention Module to capture subtle interactions between image and clinical features and employs a MultiModal Cascaded Attention mechanism to enable progressive and efficient fusion across the modalities. These components work together to streamline multimodal data integration while preserving meaningful insights. Results: Extensive experiments conducted on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset demonstrate the effectiveness of TriLightNet, showcasing superior performance compared to state-of-the-art methods. Specifically, the model achieves an accuracy of 81.25%, an AUROC of 0.8146, and an F1-score of 69.39%, all while maintaining reduced computational costs. Discussion: Furthermore, its interpretability was validated using the Integrated Gradients method, which revealed clinically relevant brain regions contributing to the predictions, enhancing its potential for meaningful clinical application. Our code is available at https://github.com/sunyzhi55/TriLightNet.