Real-World Multimodal Machine Learning for Risk Enrichment Across the Alzheimer's Disease Spectrum.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Mild cognitive impairment (MCI) is a heterogeneous condition within the Alzheimer's disease (AD) spectrum.
• Current diagnostic labels may not fully capture biological variability in MCI and AD.
• There is a need for methods to biologically enrich patient populations in memory clinics.

Purpose of the Study:

• To evaluate multimodal machine learning for biologically informed enrichment of MCI and AD patients.
• To assess the utility of routine clinical data and neuroimaging in a real-world memory clinic cohort.
• To determine if machine learning can support a more nuanced understanding of the Alzheimer's disease continuum.

Main Methods:

• Analysis of 474 patients (1547 visits) with comprehensive clinical, cognitive, and laboratory data.
• Integration of magnetic resonance imaging (MRI) regional volumes and fluorodeoxyglucose-positron emission tomography (FDG-PET) uptake.
• Training of Elastic Net and gradient boosting models using nested cross-validation with patient-level separation.

Main Results:

• Model discrimination improved with the addition of data modalities, with FDG-PET showing the largest impact.
• Posterior default mode network hypometabolism was identified as a key predictor.
• The MCI subgroup exhibited a continuous distribution of AD-like scores, indicating biological enrichment.

Conclusions:

• Multimodal machine learning models offer an interpretable framework for enriching heterogeneous memory clinic populations.
• These models can potentially improve the biological stratification of patients within the Alzheimer's disease continuum.
• Integrating diverse data types enhances the understanding of neurodegenerative disease heterogeneity.