Learning patient-specific spatial biomarker dynamics via operator learning for Alzheimer's disease progression

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Informatics

Background:

• Alzheimer's disease (AD) presents complex, heterogeneous progression, limiting current predictive models.
• Accurate forecasting of individual biomarker trajectories is crucial for effective AD management and therapeutic development.

Purpose of the Study:

• To develop a personalized machine learning framework for modeling Alzheimer's disease progression.
• To integrate multimodal data for accurate, patient-specific prediction of biomarker evolution.

Main Methods:

• Developed a machine learning-based operator learning framework integrating longitudinal multimodal imaging, biomarker, and clinical data.
• Employed geometry-aware neural operators using Laplacian eigenfunction bases to learn patient-specific disease dynamics.
• Integrated the framework within a digital twin paradigm for individualized predictions and simulations.

Main Results:

• Achieved prediction accuracy exceeding 90% across multiple Alzheimer's disease biomarkers.
• Demonstrated substantial performance improvement over existing predictive approaches.
• Enabled simulation of therapeutic interventions and in silico clinical trials.

Conclusions:

• The developed framework offers a scalable and interpretable platform for precision modeling in neurodegenerative diseases.
• This approach facilitates personalized therapeutic optimization for Alzheimer's disease.
• The digital twin paradigm enhances individualized prediction and intervention simulation capabilities.