Development of a machine learning algorithm to predict the residual cognitive reserve index

Area of Science:

• Neuroscience
• Gerontology
• Machine Learning

Background:

• Cognitive decline in late life is linked to neurodegeneration.
• Cognitive reserve explains individual differences in resilience to brain changes.
• Current methods for measuring cognitive reserve lack accessibility, validity, and mechanistic insight.

Purpose of the Study:

• To develop and validate machine learning models for estimating cognitive reserve using accessible clinical data.
• To assess if these models can predict a criterion standard of cognitive reserve.
• To determine if the models can prospectively moderate the association between brain changes and cognitive decline.

Main Methods:

• Utilized a training sample (N=1665) from UC Davis and ADNI-2, operationalizing cognitive reserve via an MRI-based residual approach.
• Trained eXtreme Gradient Boosting models (Minimal, Extended, Full) to predict the residual reserve index (RRI) using varying sets of clinical variables.
• Externally validated models in an independent ADNI 1/3/GO sample (N=1640) to test moderation of brain-cognition associations.

Main Results:

• The Minimal model (basic clinical data) showed poor accuracy (r=0.23) and failed to moderate brain-cognition effects.
• Extended and Full models (including cognitive performance and informant data) demonstrated modest accuracy (r=0.49, 0.54) and successfully moderated longitudinal brain-cognition associations.
• These machine learning models outperformed traditional proxies like education and word reading.

Conclusions:

• Accessible machine learning models incorporating cognitive performance and informant data can effectively estimate cognitive reserve without neuroimaging.
• These models offer a valid and dynamic proxy for cognitive reserve, providing insights into resilience mechanisms.
• The findings highlight the importance of cognitive and functional data, beyond basic demographics, for accurate cognitive reserve assessment.