An attention-augmented multimodal classification of alzheimer's disease and parkinson's disease vs healthy controls using MRI, EEG, and SNP data.

Due to the late manifestation of structural symptoms and symptomatic overlap, neurodegenerative diseases such as Parkinson's Disease (PD) and Alzheimer's Disease (AD) remain difficult to diagnose accurately. In order to categorize AD and PD in comparison to Healthy Controls (HC), this study suggests a multimodal classification framework that combines genetic Single Nucleotide Polymorphism (SNP) data, structural Magnetic Resonance Imaging (MRI), and functional Electroencephalography (EEG). To improve the model's accuracy and interpretability, the method makes use of an uncertainty estimate module and a novel cross-modality attention mechanism. The framework strives for diagnosis, concentrating on detecting Parkinson's disease (PD) and Alzheimer's disease (AD) in individuals exhibiting modest motor symptoms or early cognitive impairments, which are indicative of the prodromal stage of both conditions. A dataset of 2,500 MRI images, 1,500 EEG recordings, and SNP data for 1,000 subjects drawn from OpenNeuro, PPMI, and the UK Biobank was utilized in extensive analyses. The developed model was contrasted with recent unimodal and multimodal techniques. Our findings exhibit statistically significant increases of 6-12% compared to similar methods, with 95.6% average classification accuracy on AD and 94.8% on PD. The importance of the attention mechanism and both modalities to overall performance is quantified using ablation studies. Quantification of uncertainty also improves interpretability for possible clinical use. These results demonstrate the proper neurodegenerative disease diagnosis when explainable AI elements are paired with stable multimodal fusion.