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

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Alzheimer's disease (AD) presents significant variability in clinical progression, necessitating better methods for predicting cognitive decline.
• Accurate prediction of AD trajectories is crucial for patient management, care planning, and therapeutic development.

Purpose of the Study:

• To develop and evaluate artificial intelligence (AI) models for predicting cognitive decline in Alzheimer's disease.
• To explore various data representation frameworks, including deep learning approaches like CNNs and GNNs, for modeling neurocognitive trajectories.

Main Methods:

• Utilized multimodal neuroimaging data (MRI, FDG-PET, AV45-PET) and clinical data from 653 participants in the Alzheimer's Disease Neuroimaging Initiative (ADNI).
• Developed predictive models using tabular data, CNNs, and GNNs, incorporating a novel framework with modality-specific pre-training and late-fusion for integrated feature representation.
• Predicted both quantitative (rate of decline) and qualitative (presence/absence of decline) cognitive changes.

Main Results:

• AI models demonstrated strong predictive performance for future clinical diagnoses (F1 = 0.779).
• Models explained 29.4%-36.0% of the variance in cognitive decline rates and achieved AUC > 0.83 for predicting cognitive deterioration across domains (memory, language, executive function).
• CNN- and GNN-based models, coupled with the proposed pre-training strategy, consistently yielded superior predictive accuracy.

Conclusions:

• AI techniques effectively capture patterns of cognitive decline in Alzheimer's disease by integrating multimodal neuroimaging data.
• These findings support the development of AI-driven precision phenotyping for neurodegenerative patterns in AD, aiding clinical management and research.