Machine learning based on event-related oscillations of working memory differentiates between preclinical Alzheimer's disease and normal aging.

Abstract

Objective: To apply machine learning approaches on EEG event-related oscillations (ERO) to discriminate preclinical Alzheimer's disease (AD) from age- and sex-matched controls.

Methods: Twenty-two cognitively normal preclinical AD participants with elevated amyloid and 21 cognitively normal controls without elevated amyloid completed n-back working memory tasks (n = 0, 1, 2). The absolute and relative power of ERO was extracted using the discrete wavelet transform in the delta, theta, alpha, and beta bands. Four machine learning methods were employed, and classification performance was assessed using three metrics.

Results: The low-frequency bands produced higher discriminative performances compared to high-frequency bands. The 2-back task yielded the best classification capability among the three tasks. The highest area under the curve value (0.86) was achieved in the 2-back delta band nontarget condition data. The highest accuracy (80.47%) was obtained in the 2-back delta and theta bands nontarget data. The highest F1 score (0.82) was in the 2-back theta band nontarget data. The support vector machine achieved the highest performance among tested classifiers.

Conclusion: This study demonstrates the promise of using machine learning on EEG ERO from working memory tasks to detect preclinical AD.

Significance: EEG ERO may reveal pathophysiological differences in the earliest stage of AD when no cognitive impairments are apparent.