Predicting Alzheimer's disease CSF core biomarkers: a multimodal Machine Learning approach

Area of Science:

• Neuroscience
• Biomarker Discovery
• Machine Learning

Background:

• Alzheimer's disease (AD) diagnosis relies on invasive cerebrospinal fluid (CSF) biomarkers, limiting their use in screening.
• Sleep disturbances are linked to AD pathology, suggesting potential for non-invasive biomarkers.
• Quantitative analysis of polysomnography (PSG) signals for AD detection is underexplored.

Purpose of the Study:

• To evaluate a multimodal machine learning (ML) approach for predicting core AD CSF biomarkers.
• To assess the effectiveness of non-invasive quantitative sleep electroencephalography (EEG) features as AD biomarkers.
• To determine if ML can identify AD neuropathology using accessible data.

Main Methods:

• Recruited mild-moderate AD patients for PSG, CSF, and blood biomarker analysis.
• Extracted quantitative features (non-linear, time, frequency domains) from PSG signals.
• Trained ML algorithms on clinical variables (CLINVAR), conventional PSG (SLEEPVAR), quantitative PSG (PSGVAR), and combined (ALL) features.

Main Results:

• Gradient Boosting Regressors showed best performance in estimating biomarker levels.
• The combined feature subset (ALL) generally yielded the lowest training errors.
• The SLEEPVAR subset excelled in predicting Aβ42, while ALL best predicted p-tau and t-tau.

Conclusions:

• Multimodal ML effectively predicts CSF biomarkers in early AD using non-invasive data.
• This approach provides a feasible and economical screening tool for AD risk.
• Integrating computational models aids clinical decisions and early AD patient identification.