Machine learning-based estimation of the mild cognitive impairment stage using multimodal physical and behavioral...

Area of Science:

• Neurology
• Biomarkers
• Artificial Intelligence

Background:

• Mild cognitive impairment (MCI) is a precursor to dementia, making early detection crucial for better patient outcomes.
• Current diagnostic methods like MRI and neuropsychological tests face challenges in accessibility and scalability.
• Developing accessible, scalable digital biomarkers is essential for early MCI detection and management.

Purpose of the Study:

• To evaluate the potential of multimodal physical and behavioral measures as digital biomarkers for estimating MCI severity.
• To assess the classification performance of machine-learning models using gait, body composition, and sleep data.
• To compare the efficacy of these digital biomarkers against traditional MRI-based assessments.

Main Methods:

• Recruited 80 MCI patients, categorized into early and late stages using Mini-Mental State Examination scores.
• Collected gait, body composition (DXA), sleep (polysomnography), and brain MRI data.
• Utilized machine learning models (SVM, Random Forest, MLP, CNN) to analyze unimodal and multimodal data for classification.

Main Results:

• Machine learning models using only physical and behavioral data achieved high accuracy (AUC up to 94%) in differentiating early and late MCI stages.
• Performance was comparable to MRI-based models, with marginal improvements when combining data types.
• Key predictors of cognitive function included gait velocity, lean body mass, and sleep efficiency.

Conclusions:

• Multimodal digital biomarkers derived from physical and behavioral signals can effectively estimate MCI severity.
• This approach presents a scalable and cost-effective strategy for the early detection and monitoring of cognitive decline.
• These findings support the use of digital biomarkers in real-world settings for managing cognitive health.