Deep multimodal learning for domain-level cognitive decline prediction in Alzheimer's disease

Alzheimer's disease (AD) is characterized by significant variability in clinical progression; however, few studies have focused on developing models to predict cognitive decline. Anticipating these trajectories is essential for patient management, care planning, and developing new treatments. This study explores the potential of artificial intelligence (AI) techniques to model neurocognitive trajectories from multimodal neuroimaging data and further investigates different data representation frameworks.

Using information from 653 participants from the Alzheimer's Disease Neuroimaging Initiative (ADNI), we developed models to predict future clinical diagnoses and cognitive decline, both quantitatively (rate of decline) and qualitatively (presence or absence of decline). Input features included structural T1-weighted magnetic resonance imaging (MRI), [¹⁸F]-fluorodeoxyglucose positron emission tomography (FDG-PET), [¹⁸F]-florbetapir PET (AV45-PET), neuropsychological assessments, and demographic variables. Several information representation strategies were explored, including tabular data models, convolutional neural networks (CNNs), and graph neural networks (GNNs). Furthermore, to maximize the use of all available information, we proposed a modeling framework that performed modality-specific pre-training to learn feature embeddings, which were then integrated through a late-fusion layer to produce a unified representation for downstream prediction.

The modeling strategies demonstrated good predictive performance for future clinical diagnoses, consistent with previous studies (F1 = 0.779). Quantitative models explained approximately 29.4%-36.0% of the variance in cognitive decline. In the qualitative analysis, the models achieved AUC values above 0.83 when predicting cognitive deterioration in the memory, language, and executive function domains. Architecturally, CNN- and GNN-based models yielded the best performance, and the proposed pre-training strategy consistently improved predictive accuracy.

This study demonstrates that AI techniques can capture patterns of cognitive decline by exploiting multimodal neuroimaging data. These findings contribute to the development of more precise phenotyping approaches for neurodegenerative patterns in AD.