MC-RVAE: Multi-channel recurrent variational autoencoder for multimodal Alzheimer's disease progression modelling

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Data Science

Background:

• Neurodegenerative diseases involve complex, multi-scale mechanisms.
• Predicting disease progression requires integrating diverse data over time, which is challenging.

Purpose of the Study:

• To develop a model capable of capturing cross-modal and temporal interactions for disease progression analysis.
• To improve the prediction and understanding of longitudinal disease trajectories.

Main Methods:

• A multi-channel recurrent variational autoencoder (MCRVAE) model was proposed.
• The model utilizes a shared latent variational space parameterized by a recurrent neural network.
• Evaluation was performed on synthetic and real longitudinal datasets (N=897) with imaging and non-imaging data.

Main Results:

• The MCRVAE outperformed baseline models (KNN, random forest, group factor analysis) in reconstructing missing modalities.
• Mean absolute error was reduced by 5% for subcortical volumes and cortical thickness.
• The model demonstrated robustness to missing features and generated realistic synthetic imaging biomarker trajectories.

Conclusions:

• The proposed MCRVAE effectively models complex interactions across modalities and time for neurodegenerative disease research.
• This approach enhances the prediction of disease progression and the generation of synthetic data.
• The model offers a robust tool for analyzing longitudinal biomedical data.