Modeling autosomal dominant Alzheimer's disease with machine learning

Area of Science:

• Neuroscience
• Biomedical Engineering
• Genetics

Background:

• Autosomal dominant Alzheimer's disease (ADAD) represents a unique genetic model for studying Alzheimer's disease (AD) pathogenesis.
• Understanding the precise trajectories of molecular and structural changes in ADAD is crucial for developing targeted interventions.

Purpose of the Study:

• To apply machine learning models for discovering novel disease trajectories in ADAD.
• To identify key predictors of disease progression across multiple imaging modalities.

Main Methods:

• Longitudinal structural MRI, amyloid PET, and fluorodeoxyglucose PET data were collected from 131 ADAD mutation carriers and 74 non-carriers.
• A deep neural network was trained to predict disease progression, and Relief algorithms identified significant predictors.
• The study utilized data from the Dominantly Inherited Alzheimer Network (DIAN).

Main Results:

• The Relief algorithm identified the caudate, cingulate, and precuneus as the strongest predictors of mutation status.
• The predictive model achieved high accuracy (R² = 0.95 for Pittsburgh compound B and atrophy, R² = 0.93 for fluorodeoxyglucose) in forecasting future changes.
• Disease progression was characterized by sigmoidal amyloid accumulation, biphasic metabolic changes, and gradual volume decrease.

Conclusions:

• Machine learning effectively models ADAD progression, revealing distinct sigmoidal and biphasic trajectories for key biomarkers.
• The findings highlight specific brain regions (subcortical, middle frontal, posterior parietal) most affected during disease progression.
• This approach offers a powerful tool for understanding ADAD and potentially other neurodegenerative diseases.