Page 102 - MEGIN Book Of Abstracts - 2023

BACKGROUND Electrophysiological studies show that Detection of early stages of Alzheimer's disease
reductions in power within the alpha band are associat- based on MEG activity with a randomized
ed with the Alzheimer's disease (AD) continuum. Physi- convolutional neural network (2020)
cal activity (PA) is a protective factor that has proved to
reduce AD risk and pathological brain burden. Previous Lopez-Martin, Manuel; Nevado, Angel; Carro, Belen
research has confirmed that exercise increases power
in the alpha range. However, little is known regarding Dpto., TSyCeIT, ETSIT, Universidad de Valladolid, Paseo de
whether other non-modifiable risk factors for AD, such Belén 15, Valladolid, 47011, Spain. Electronic address: mlo-
as increased age or APOE ε4 carriage, alter the associa- [email protected]; Laboratory for Cognitive and Computation-
tion between PA and power in the alpha band. al Neuroscience, Center for Biomedical Technology, Campus
Montegancedo, Pozuelo de Alarcón, Madrid, 28223, Spain;
METHODS The relationship between PA and alpha Experimental Psychology, Department of the Complutense,
band power was examined in a sample of 113 healthy University of Madrid, Spain
adults using magnetoencephalography. Additionally,
we explored whether ε4 carriage and age modulate ABSTRACT The early detection of Alzheimer's disease
this association. The correlations between alpha power can potentially make eventual treatments more ef-
and gray matter volumes and cognition were also fective. This work presents a deep learning model to
investigated. detect early symptoms of Alzheimer's disease using
synchronization measures obtained with magneto-
RESULTS We detected a parieto-occipital cluster in encephalography. The proposed model is a novel
which PA positively correlated with alpha power. The deep learning architecture based on an ensemble of
association between PA and alpha power remained randomized blocks formed by a sequence of 2D-con-
following stratification of the cohort by genotype. volutional, batch-normalization and pooling layers. An
Younger and older adults were investigated separately, important challenge is to avoid overfitting, as the num-
and only younger adults exhibited a positive relation- ber of features is very high (25755) compared to the
ship between PA and alpha power. Interestingly, when number of samples (132 patients). To address this issue
four groups were created based on age (younger-older the model uses an ensemble of identical sub-models
adult) and APOE (E3/E3-E3/E4), only younger E3/E3 all sharing weights, with a final stage that performs an
(least predicted risk) and older E3/E4 (greatest predict- average across sub-models. To facilitate the explora-
ed risk) had associations between greater alpha power tion of the feature space, each sub-model receives a
and higher PA. Among older E3/E4, greater alpha power random permutation of features. The features corre-
in these regions was associated with improved memory spond to magnetic signals reflecting neural activity and
and preserved brain structure. are arranged in a matrix structure interpreted as a 2D
image that is processed by 2D convolutional networks.
CONCLUSION PA could protect against the slowing The proposed detection model is a binary classifier
of brain activity that characterizes the AD continuum, (disease/non-disease), which compared to other deep
where it is of benefit for all individuals, especially E3/E4 learning architectures and classic machine learning
older adults. classifiers, such as random forest and support vector
machine, obtains the best classification performance
Keywords: APOE ε4, Alpha power, Alzheimer’s disease, results with an average F1-score of 0.92. To perform the
Magnetoencephalography, Physical activity comparison a strict validation procedure is proposed,
and a thorough study of results is provided.
Alzheimer's research & therapy (2020), Vol. 12, No. 1
(32962736) (6 citations)

 ontents  Index 81
C