Objective: This retrospective and exploratory study tested the accuracy of artificial neural networks (ANNs) at detecting Alzheimer's disease patients with dementia (ADD) based on input variables extracted from resting-state electroencephalogram (rsEEG), structural magnetic resonance imaging (sMRI) or both.

Methods: For the classification exercise, the ANNs had two architectures that included stacked (autoencoding) hidden layers recreating input data in the output. The classification was based on LORETA source estimates from rsEEG activity recorded with 10-20 montage system (19 electrodes) and standard sMRI variables in 89 ADD and 45 healthy control participants taken from a national database.

Results: The ANN with stacked autoencoders and a deep leaning model representing both ADD and control participants showed classification accuracies in discriminating them of 80%, 85%, and 89% using rsEEG, sMRI, and rsEEG + sMRI features, respectively. The two ANNs with stacked autoencoders and a deep leaning model specialized for either ADD or control participants showed classification accuracies of 77%, 83%, and 86% using the same input features.

Conclusions: The two architectures of ANNs using stacked (autoencoding) hidden layers consistently reached moderate to high accuracy in the discrimination between ADD and healthy control participants as a function of the rsEEG and sMRI features employed.

Significance: The present results encourage future multi-centric, prospective and longitudinal cross-validation studies using high resolution EEG techniques and harmonized clinical procedures towards clinical applications of the present ANNs.