Abstract Alzheimer's disease (AD) classification using machine learning has increasingly relied on multimodal inputs such as Magnetic Resonance Imaging (MRI), cognitive assessments, and biological markers. This study evaluates whether integrating these sources enhances predictive performance compared to using them independently. Neural networks were trained on data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) to classify subjects into Cognitively Normal (CN), Mild Cognitive Impairment (MCI), and AD categories using unimodal, bimodal, and trimodal input configurations. Contrary to expectations, multimodal models did not consistently outperform unimodal ones. The highest test accuracy (81%) was achieved by both the cognitive-only and trimodal models, with the former also demonstrating superior class-wise performance. These findings suggest that neuropsychological features may carry greater diagnostic value than imaging or fluid biomarkers, underscoring the importance of more targeted data fusion strategies. Furthermore, the inclusion of biological markers did not significantly improve early MCI detection, likely due to their limited dimensionality and the model's constrained ability to extract meaningful patterns from such inputs.
Khan et al. (Mon,) studied this question.