Signals of Decline: Machine Learning driven Biomarkers for Alzheimer’s Disease

Marisa Ripoll, Moritz Nicolas Andres, Abbas Sohrabpour, Christoph Dinh, Klaus Diepold, Sheraz Khan

Keywords: Alzheimer’s Disease, MEG, Machine Learning, Shapley Additive Explanations (SHAP)

TL;DR: We present a MEG-based machine learning framework that achieves state-of-the-art MCI classification, leverages SHAP for interpretable biomarkers, and extends to modeling cognitive severity.

Abstract:

Alzheimer’s disease (AD) is a growing global health challenge, with pathological
changes beginning decades before clinical symptoms. Identifying non-invasive and
interpretable biomarkers is critical for early intervention. Magnetoencephalography
(MEG) provides access to brain oscillatory dynamics and connectivity patterns
that are disrupted in mild cognitive impairment (MCI), a prodromal stage of AD.
We evaluate five families of MEG-derived features and train an ensemble of 200
feature models, achieving MCI classification with F1 72.43%. We use Shapley
Additive Explanations (SHAP) to highlight discriminative regions and connections,
offering interpretable insights and pointing to potential new markers. Beyond
binary detection, model scores correlate with Mini-Mental State Examination
(MMSE) scores, suggesting potential for continuous disease staging. Together,
these results establish MEG-based machine learning as a promising avenue for
robust and clinically meaningful biomarkers.

Submission Number: 93