METHODS

published: 02 July 2018

doi: 10.3389/fnagi.2018.00184

Predicting Age From Brain EEG

Signals—A Machine Learning
Approach
Obada Al Zoubi 1,2 , Chung Ki Wong 1 , Rayus T. Kuplicki 1 , Hung-wen Yeh 1 , Ahmad Mayeli 1,2 ,
Hazem Refai 2 , Martin Paulus 1 and Jerzy Bodurka 1,3*
1
Laureate Institute for Brain Research, Tulsa, OK, United States, 2 Department of Electrical and Computer Engineering,
University of Oklahoma, Tulsa, OK, United States, 3 Stephenson School of Biomedical Engineering, University of Oklahoma,
Norman, OK, United States

Objective: The brain age gap estimate (BrainAGE) is the difference between the
estimated age and the individual chronological age. BrainAGE was studied primarily using
MRI techniques. EEG signals in combination with machine learning (ML) approaches
were not commonly used for the human age prediction, and BrainAGE. We investigated
whether age-related changes are affecting brain EEG signals, and whether we can predict
the chronological age and obtain BrainAGE estimates using a rigorous ML framework
with a novel and extensive EEG features extraction.
Methods: EEG data were obtained from 468 healthy, mood/anxiety, eating and
substance use disorder participants (297 females) from the Tulsa-1000, a naturalistic
longitudinal study based on Research Domain Criteria framework. Five sets of
preprocessed EEG features across channels and frequency bands were used with
different ML methods to predict age. Using a nested-cross-validation (NCV) approach
and stack-ensemble learning from EEG features, the predicted age was estimated. The
Edited by:
Christian Gaser,
important features and their spatial distributions were deduced.
Friedrich-Schiller-Universität-Jena, Results: The stack-ensemble age prediction model achieved R2 = 0.37 (0.06),
Germany
Mean Absolute Error (MAE) = 6.87(0.69) and RMSE = 8.46(0.59) in years. The age
Reviewed by:
Mihai Moldovan, and predicted age correlation was r = 0.6. The feature importance revealed that age
University of Copenhagen, Denmark predictors are spread out across different feature types. The NCV approach produced a
Safikur Rahman,
reliable age estimation, with features consistent behavior across different folds.
Yeungnam University, South Korea

*Correspondence: Conclusion: Our rigorous ML framework and extensive EEG signal features allow a
Jerzy Bodurka reliable estimation of chronological age, and BrainAGE. This general framework can be
[email protected]
extended to test EEG association with and to predict/study other physiological relevant
Received: 31 March 2018 responses.
Accepted: 01 June 2018
Keywords: aging, human brain, EEG, machine learning, feature extraction, BrainAGE
Published: 02 July 2018
Citation:
Al Zoubi O, Ki Wong C, Kuplicki RT,
Yeh H, Mayeli A, Refai H, Paulus M
INTRODUCTION
and Bodurka J (2018) Predicting Age
From Brain EEG Signals—A Machine
Brain changes due to age have been studied for decades (e.g., Lindsley, 1939; Harmony et al.,
Learning Approach. 1990; Lao et al., 2004) and more recently using genetics Lu et al. (2017). The term BrainAGE
Front. Aging Neurosci. 10:184. (the difference between predicted age—chronological age) was introduced to examine and capture
doi: 10.3389/fnagi.2018.00184 any disease-related deviations from natural aging, by comapring BrainAGE estimates in disease

Frontiers in Aging Neuroscience | www.frontiersin.org 1 July 2018 | Volume 10 | Article 184

Al Zoubi et al. Predicting Age From Brain EEG

group to healthy control group. Magnetic Resonance Imaging group when performing mental tasks (Marciani et al., 1994). The
(MRI) has been widely used to build predictive models for age theta power was shown to increase from resting to arithmetic
by utilizing white matter (WM) and gray matter (GM) properties. task for younger group, while decreasing for the older group
Franke et al. (2010) employed T1-weighted (T1w) MRI structural (Widagdo et al., 1998). Moreover, the delta band beta-3 power
images to establish a framework (using a kernel method for showed an increase from resting to arithmetic tasks, while alpha
regression) for automatically and efficiently estimating the age was decreased (Widagdo et al., 1998). A more recent study used
of healthy individuals. This framework proved to be a reliable, four channels EEG recording to investigate age-related changes
scanner-independent, and efficient method for age estimation in EEG power from thousands of subjects throughout adulthood
in healthy subjects, yielding a correlation of r = 0.92 between (Hashemi et al., 2016). Their findings showed an overall age-
the estimated and the real age in the test samples and a mean related shift in band power from lower to a higher frequency and
absolute error of 5 years. Similarly, Cole et al. (2017) used Deep a gradual slowing of the peak α frequency with age. Furthermore,
Learning (DL) to study BrainAGE using both pre-processed studying the source of cortical rhythm suggested that occipital
and raw T1w MRI images. Their approach predicted age with delta and posterior cortical alpha rhythms decrease in magnitude
minimal efforts by achieving a correlation between age and during physiological aging with both linear and nonlinear trends
predicted-age of r = 0.96 and MAE = 4.16 years. Also, Valizadeh (Babiloni et al., 2006). Age prediction from EEG was studied
et al. (2017) obtained R2 = 0.77 from large healthy subjects in Dimitriadis and Salis (2017), where authors used functional
(n = 3,144) by training features from various anatomical brain connectivity features from EEG to predict age from 94 healthy
regions. Càmara et al. (2007) studied age-related changes in water subjects. Their results showed accuracy of R2 = 0.60 for eyes-
self-diffusion in cerebral white matter using Diffusion Tensor open and R2 = 0.48 for eyes-closed.
Imaging (DTI). Their results revealed white matter changes with The influence of diseases on EEG features were investigated
age in different brain regions, like the corpus callosum, prefrontal elsewhere. For instance, Saletu et al. (1995) used the mean EEG
regions, the internal capsule, the hippocampal complex, and power spectrum to study group differences between multi-infract
the putamen. Functional MRI (fMRI) imaging was also used to dementia (MID) and dementia of Alzheimer’s disease (AD) and
predict age alone, or combined with other imaging approach. compared it a healthy group. MID group showed a significant
For instance, Dosenbach et al. (2010) were able to explain increase of theta activity in occipital regions and decrease in
up to 55% of their sample variance from the functional MRI alpha activity. Abnormalities in cortical neural synchronization
connectivity (fcMRI) data. Likewise, Qin et al. (2015) related for subjects were observed in subjects with mild cognitive
the developmental changes in the amplitude of low-frequency impairment due to AD (ADMCI) and to Parkinson Disease
spontaneous fluctuations in resting-state fMRI to age. They (PDMCI) in delta and alpha (Babiloni et al., 2016). Differentiating
reported MAE of 4.6 years between chronological age and subjects with Alzheimer’s disease from healthy ones was studied
predicted-age. More recently, Liem et al. (2017) utilized cortical in Babiloni et al. (2016). Authors reported 70% accuracy using
anatomy and whole-brain functional connectivity for predicting the power and functional connectivity of cortical sources, which
brain-based age achieving MAE = 4.29 year. Several BrainAGE was later improved to 77% using Artificial Neural Network
studies revealed changes and differences among clinical groups. (Triggiani et al., 2017). Table 1 provides a summary of studies
For example, BrainAGE estimations in schizophrenia patients that specifically reported age prediction performance from brain
was attributed to accelerated aging when compared to healthy imaging data.
and bipolar subjects (Nenadić et al., 2017). In additions, In this study, we proposed a robust and rigorous framework
individuals diagnosed with medically refractory epilepsy had a to predict BrainAGE using different features of EEG signals
higher predicted age than health subjects (Pardoe et al., 2017). recorded during fMRI. First, we extended a recent open-source
Herein, we focus on studying BrainAGE using EEG signals. EEG feature extraction software in Matlab (Toole and Boylan,
Several studies have demonstrated that EEG features like EEG
rhythmic activity (e.g., delta, theta, alpha, beta, and gamma)
changes as a function of age (Matthis et al., 1980; Clarke et al., TABLE 1 | A summary of related work for predicting age from brain imaging data.
2001; Marshall et al., 2002; Ashburner, 2007; Cragg et al., 2011).
For instance, Benninger et al. (1984) found theta band showed Work Data No. of Performance
an increase in power spectra, while delta exhibited decrease for Samples
healthy children between 4 and 17 years. Gasser et al. (1988)
Franke et al., 2010 MRI 650 r = 0.92, MAE = 5 years
showed that: (i) the relative power increases with age in fast
Cole et al., 2017 MRI 2,001 r = 0.96, MAE = 4.16
bands, while decreases for the slow bands in healthy children years
and adolescent (6–17 years), (ii) all bands showed increase in the Dosenbach et al., fMRI 238 R2 = 0.55
absolute power except for alpha-2. Analyzing the coherence of 2010
EEG in resting state revealed that younger healthy subjects had Qin et al., 2015 fMRI 183 MAE = 4.6 years
a lower coherence than elderly ones for theta, alpha-3, beta-2, Valizadeh et al., MRI 3,144 R2 = 0.77
and beta-2 (Kikuchi et al., 2000). The beta relative power was 2017
positively correlated with age for older subjects for resting with Dimitriadis and EEG 94 R2 = 0.6 for eyes open
eye closed condition (Marciani et al., 1994). The alpha reactivity Salis, 2017 R2 = 0.48 for eyes closed
decreased and showed negative correlation with age in the older Liem et al., 2017 fMRI + MRI 2,354 MAE = 4.29 years

Frontiers in Aging Neuroscience | www.frontiersin.org 2 July 2018 | Volume 10 | Article 184

Al Zoubi et al. Predicting Age From Brain EEG

2017) to provide a feature representation of individual subjects. We included EEG data collected from 468 subjects (mean
Then, we applied a set of machine learning (ML) methods age: 34.8 years, 297 females). One resting EEG-fMRI run was
to predict age from features. Our proposed framework and conducted for each subject; lasting 8 min. The participants
a proof-of-concept analysis revealed that robust BrainAGE were instructed to relax and keep their eyes open and fixate
predictors span multiple EEG signal features, including separate on a cross.
channels, and frequencies. The overall accuracy elaborated Magnetic resonance (MR) images were acquired
that EEG BrainAGE is a promising approach to study brain simultaneously via a General Electric Discovery MR750
maturity and has capacity to reveal different factors that affect whole-body 3 T MRI scanner with a standard 8-channel,
natural aging. receive-only head coil array. A single-shot gradient-recalled EPI
sequence with Sensitivity Encoding (SENSE) was employed for
the fMRI acquisition. The fMRI data has not been used in this
METHODS paper.
Participants EEG Data Preprocessing
Participants were selected from the first 500 subjects of the
For each scan the EEG data was preprocessed with an in-house
Tulsa 1000 (T-1000), a naturalistic study that is assessing
script developed in MATLAB. The script was designed to remove
and longitudinally following 1,000 individuals, including
the MR gradient artifact and cardioballistic artifact from the
healthy comparisons and treatment-seeking individuals with
EEG data. The details about the preprocessing script are given
mood disorders and/or anxiety, substance use, and eating
as follow. The MR gradient artifact was first removed from the
disorders. The T-1000 aims to determine how disorders of
EEG data using optimal basis sets (Allen et al., 2000; Delorme
affect, substance use, and eating behavior organize across
and Makeig, 2004; Niazy et al., 2005). Then the EEG data was
different levels of analysis with a focus on predictors of long-
band-pass filtered between 1 and 70 Hz, down-sampled to 4 ms
term prognosis, symptom severity, and treatment outcome
temporal resolution, and band-stop filtered (1 Hz bandwidth) at
(Victor et al., 2018). The T-1000 study is conducted at the
the harmonics of the fMRI slice selection frequency (19.5 Hz),
Laureate Institute for Brain Research. The study human research
AC power line frequency (60 Hz), and a 26 Hz vibration artifact
protocol was approved by the Western Institutional Review
frequency (Mayeli et al., 2016). Then the cardioballistic artifact
Board. All participants provided written informed consent
was corrected using optimal basis sets subtraction (Niazy et al.,
and received financial compensation for participation. As
2005), which requires the timing of the artifact cycle. In order
described in details in Victor et al. (2018), the study participants
to achieve a robust artifact cycle determination, the script
were screened on the basis of a treatment-seeking history
determined the artifact cycle using the cardioballistic component
and dimensional psychopathology scores: Patient Health
directly from the EEG-fMRI data (Wong et al., 2018), which
Questionnaire (PHQ-9) ≥ 10 and/or Overall Anxiety Severity
was extracted by independent component analysis (Bell and
and Impairment Scale (OASIS) ≥ 8, Drug Abuse Screening
Sejnowski, 1995) and was automatically identified (Wong et al.,
Test (DAST-10) score > 3, or Eating Disorder Screen (SCOFF)
2016).
score ≥ 2. Each participant underwent approximately 24 h
of testing over the course of 1 year including a standardized
EEG Feature Extraction
diagnostic assessment, self-report questionnaires, behavioral and
Feature extraction is a quintessential phase in any EEG analysis
physiological measurements indexing RDoC domains, magnetic
that depends on finding common features representation among
resonance imaging focusing on brain structure and reward-
EEG samples. The existing literature provides quite extensive
related processing, fear processing, cognitive control/inhibition,
span of features extraction using variety of signal processing
interoceptive processing, and blood/microbiome
approaches (Jenke et al., 2014). Choosing feature extraction
collection. Please refer to Figure S1 in Supplementary
method relies on the applications of the prediction and the
for the detailed information about the demographics of
compromisation between interpretation and performance. For
the dataset.
instance, advanced features extraction methods can be used at the
cost of interpretation, where such approaches have been shown to
EEG Recording outperform the typical approaches (Dimitriadis and Salis, 2017;
EEG signals were recorded simultaneously with fMRI using a 32- Al Zoubi et al., 2018). In our case, BrainAGE emphasizes on the
channel MR-compatible EEG system arranged according to the interpretation and understanding of the predictors since the goal
international 10–20 system from Brain Products GmbH. ECG is to find those features that influence BrainAGE modeling. Thus,
signal was recorded using an electrode on the subject’s back. In we adopted the similar set of features used by Toole and Boylan
order to synchronize the EEG system clock with the 10 MHz MRI (2017), which extracts a wide range of commonly used features
scanner clock, a Brain Products’ SyncBox device was utilized. from EEG. However, our work takes an extensive approach to
The EEG acquisition temporal resolution, and measurement survey all features from all channels and bands without reducing
resolution were 0.2 ms (i.e., 16-bit 5 kS/s sampling) and 0.1 µV features by averaging as done in Toole and Boylan (2017). That is,
respectively. A hardware filtering throughout the acquisition in all features from all possible channels, bands and across different
a frequency band between 0.016 and 250 Hz was applied to EEG types of features were extracted from EEG. In addition, the
signals. types of features used here are commonly used in literature to

Frontiers in Aging Neuroscience | www.frontiersin.org 3 July 2018 | Volume 10 | Article 184

Al Zoubi et al. Predicting Age From Brain EEG

analyze EEG data. That is, the interpretation and replication of Connectivity Domain Features
such features are less challenging than using uncommon features. We calculated the brain symmetry index (BSI) as the mean of
However, our approach results in a relatively large number of PSD difference between the left and right hemispheres for each
features from EEG. Therefore, a feature selection and suitable frequency band (K = δ, θ, α, β, γ).
ML algorithms are needed to deduce the important predictors. Let ai and bi be the lower and upper frequency limit of band i,
All features were extracted from each subject independently and the BSI for band i is:
arranged in one row/sample.  
bi 
X 
i 1  Pleft [K] − Pright [K] 
CBSI =   (1)
General Configuration (bi − ai )  Pleft [K] + Pright [K] 
k=ai
EEG bands of interest are [δ = 0.5–4; θ = 4–7; α = 7–13; β = 13–
30; W = 0.5–30] Hz using the bipolar montage of the EEG, with With
W denotes the whole frequency range of EEG. We denoted EEG
Pn/2 PM
time series as xi [n] with frequency bands of i = α, β, θ , γ , W Pm [K]
m=1 Pm [K] m= M
2 +1
and n is channel’s index (the total number of channels is N = 31). Pleft [K] = and Pright [K] = (2)
We selected five types of features as follows: amplitude, range, n/2 n/2
spectral, connectivity, and fractal dimension. We divided the
Also, we calculated the median and lag of maximum correlation
EEG recordings from each subject into 60 s and 50% overlap
coefficient of the Spearman correlation between envelopes of
among epochs (14 epochs). Figure 1 elaborates on the features
hemisphere-paired channels and coherence between channel
extraction process. For each channel, we divided the signal into m
pairs.
epochs, then we filtered each epoch into corresponding frequency
bands. A specific feature extraction was applied to each sub- Fractal Dimension Domain Features
segment yielding m values. Finally, we estimated the channel- Fractal dimension for time series is a value that estimates to what
level feature for the corresponding frequency band as the average extent the fractal pattern changes with respect to the scale at
across all epochs. The process is slightly different for Fractal which it embeds. We applied Higuchi method with k = 6 for
Dimension (FD) features, since we estimate the features without each EEG channel to estimate the FD.
filtering into the frequency bands. Table 2 summarizes the extracted set of features from EEG
data.
Amplitude Domain Features
The amplitude features characterize the statistical properties of Feature Reduction
the signal power Aipower and the signal envelope Emean i . We After feature extraction, we eliminated features that are either
calculated: (i) the mean, (ii) standard deviation, (iii) skewness, low in variation among subjects or highly correlated with other
and (iv) kurtosis for each channel across frequency bands. The features using the “findCorrelation” function in the “caret”
i
Emean is calculated using mean of the envelop e[n]i , which is package (Kuhn, 2008), version “6.0-78.” The “findCorrelation”
identified in complex notation as: ei [n] = |xi [n] + jH {xi [n]} |2 , evaluates the pair-wise correlation of features. Then, it finds
with which is the Hilbert transformation. the highest absolute pair-wise correlation, if two features have
a high correlation (r ≥ 0.9 Pearson’s correlation), it eliminates
the feature with the highest mean absolute correlation. It should
Range Domain Features (rEEG)
be noted that other feature selection methods could be used
Range features account for the peak-to-peak voltages changes
to select the best features using NCV approach. However, the
and characterize changes in the signal over the time. To achieve
interpretation of such approach could be challenging i.e., the
that, we segmented each epoch into short-time portions each with
selected features from the inner loop of the NCV may vary across
a window size of w = 2 s and overlap = 50%. Then, for each
folds. In addition, using other feature selections should be applied
segment, we calculated the corresponding range of peak-to-peak.
within each loop of NCV, which increases the computational
This produced samples from each epoch to estimate the mean,
overhead. Thus, removing correlated features provides a better
median, 5th and 95th percentiles, standard deviation, coefficient
way to select features in this case. Figures S2, S3 in Supplementary
of variation and the measure of symmetry.
shows the correlation matrices before and after removing the
correlated features.
Spectral Domain Features
Spectral features have been the most commonly used features Machine Learning Methods
for EEG. To extract the spectral features, we applied Welch Selecting appropriate ML algorithms is a critical step to achieve
periodgram to estimate the power spectral density (PSD) and robust BrainAGE estimation. Having represented each subject’s
the hamming window with a length of 2 s and overlap of 50%. features in one row, the final dataset dimension is x = n × m,
The following spectral features have been extracted: (1) power, with n = 468 and m = 863. We used R package “caret”
(2) relative power, (3) entropy (using Wiener and Shannon to perform a set of regression algorithms: Elastic Net (ENET),
methods), (4) edge frequency (the cut-off frequency at which Support Vector Regression (SVR), Random Forest (RF), extreme
encompasses 95% of spectral power), and (5) differences between gradient boosting tree (XgbTree), and Gaussian Process with
consecutive short-time spectral estimations. Polynomial Kernel (gaussprPoly). The aim is to test different

Frontiers in Aging Neuroscience | www.frontiersin.org 4 July 2018 | Volume 10 | Article 184

Al Zoubi et al. Predicting Age From Brain EEG

FIGURE 1 | Feature extraction procedure. Each channel is divided into m epoch. From there, we filtered each epoch into α, β, θ, γ, and W frequency bands. Then, for
each filtered epoch, we applied the desired feature. This resulted in m feature value from all epochs, which are then averaged to estimate the channel-level feature. In
the figure, we represent each feature using three indices: f(channel, epoch, band) with channel = [1..N], epoch = [1..m], and band = [α, β, θ, γ, W]. The final out is a
channel-level feature and represented with two indices f(channel, band).

ML techniques in order to provide a better estimation for Fourth, XgbTree utilizes a combination of ensemble learning,
age. First, ENEST is a linear regression technique that uses optimization and regularization to build generalized model from
L1 and L2 regularization to prevent overfitting. Second, SVR training data. Finally, gaussprPoly is a probabilistic approach
uses optimization to build the regression model, but in high to build a regression model by learning the distribution of the
dimensional version of the training data. In our case, we used training data given the response (age). Similar to the kernel
a kernel with radial basis function to project the data into function in SVR, gaussprPoly adopts a polynomial kernel to
high dimension space. Third, RF is one of the most common project data into high dimension space.
ensemble techniques, where it performs subsampling for the To provide un-biased prediction for age, the nested cross
feature space of training data to build multi weak learners. validation was adopted in building age prediction models
Thus, different models from the training data are produced (Varma and Simon, 2006). Figure 2 depicts the NCV procedure
and then averaged to minimize the variance across models. consisting of two main loops: the inner and outer loops. The

Frontiers in Aging Neuroscience | www.frontiersin.org 5 July 2018 | Volume 10 | Article 184

Al Zoubi et al. Predicting Age From Brain EEG

TABLE 2 | The extracted features from EEG data.

Feature group Subset of features Across Across Number of

bands channels features

Amplitude Total power, mean, standard deviation, skewness, kurtosis, envelope mean, and standard Yes Yes 6 × 4 × 31
deviation
Peak-to-peak Mean, median, 5th and 95th percentiles, standard deviation, the coefficient of variation and the Yes Yes 7 × 4 × 31
measure of symmetry
Spectral power Spectral power and relative power, spectral entropy (using Wiener and Shannon methods), Yes Yes 6 × 4 × 31
spectral edge frequency (the cut-off frequency at which encompasses 95% of spectral power)
and spectral differences between consecutive short-time spectral estimations
Connectivity Brain symmetry index, correlation, mean and maximum of frequency at which the maximum Yes No 5×4
coherence is achieved
Fractal dimension Fractal dimension No Yes 31

FIGURE 2 | The nested-cross-validation procedure for predicting age. The example here demonstrates the first fold of the outer loop. The procedure consists of an
inner loop (blue color) and outer loop. The inner loop is used to find the best models to predict the age. The outer loop uses those models to predict the age on the
testing set. The process is repeated for all folds of the outer loop, which results in building a prediction of age from all samples.

inner loop is used to find the best parameters from training set, is the method index of the corresponding method Mi (i = 1..r)
while the outer loop is used to evaluate the best parameters on and l refers to the fold l from the outer loop. The symbol “P”
the testing set. To elaborate on the NCV, let the subscript refers refers to the prediction process associated with each method.
to data and models from inner loop of NCV, while the superscript Stacking ensemble helps to improve the stability of prediction
represents the ones from outer loop. In our run, we used 10- by combining the prediction from other models; i.e., predictions
fold cross-validation (KI = 10) for the inner and 10-fold cross- from the five methods were combined by learning weights via a
validation for outer loop (KO = 10). The inner loop was used general linear model (GLM). In details, the GLM was trained on
to estimate the best parameters on training data (Tr1 ) using a the resampled predicted age from the inner loop (yTril ). Then,
grid search and the one-standard error rule. Each inner loop the GLM was used to provide one weighted-average prediction in
consists of 5-repeat (R = 5) for each method. The outer loop 10-fold cross-validation (KEns = 10). From there, the best stack-
uses the best obtained models to build a stack-ensemble model. ensemble model (θEnsl ) was used to predict age for the testing

The best models are represented by its best parameters θ li , with i d l

set (YTs ). That is, the prediction of age is calculated for the

Frontiers in Aging Neuroscience | www.frontiersin.org 6 July 2018 | Volume 10 | Article 184

Al Zoubi et al. Predicting Age From Brain EEG

individual methods yTrli = Pi (Trl , θi ), and then the weighted notice that “spectral flatness of beta band from channel TP9”
average is estimated for fold l. is the most important predictor of age with r = 0.34. Please
refer to Figure S4 in Supplementary for detailed graphing for the
dl = P ([yTrl , yTrl , . . . , yTrl ], θ l )
YTs relationship between top predictors and age.
Ens 1 2 r Ens
The relationship between chronological age and the top
After iterating over all folds from the outer loop, a prediction for features was studied by the Partial Dependence Plot (PDP)
the age for the entire dataset can be built. In addition, the variable (Friedman et al., 2001). For each training model, the consistency
importance of predictors from the stacking ensemble models was across folds was examined by overlaying the PDP curves. One
estimated across the outer loop of NCV. Finally, the predicted wants the same feature to behave similarly among the folds of
age and age values were used to estimate the BrainAGE for the the outer loop of NCV. Figure 8 shows the PDP for the top
dataset. Figure 3 shows the overall framework to estimate the feature. As can be seen, the PDP for each fold (thin lines) have
BrainAGE. consistent behavior among all folds. Figure S6 in Supplementary
shows PDPs for the top features.
RESULTS
The NCV R-Squared performance for Stack-Ensemble and
underlay methods is shown in Figure 4. The individual
performance for each ML method was calculated before the
stack-ensemble phase. The results showed that SVR with
radial kernel achieved the best accuracy R2 = 0.34 (0.056),
MAE = 7.01(0.68) years and RMSE = 8.7(0.63) years. On
the other hand, the stack-ensemble improved the overall
performance with R2 = 0.37 (0.064), MAE = 6.87(0.69) years,
and RMSE = 8.46 (0.59) years.
The correlation between predicted age and age is shown in
Figure 5, while the BrainAGE variable is plotted in Figure 6.
The importance of features was estimated such that the total
summation of features importance is 100 from each fold of the
outer loop of NCV. Then, the importance scores were averaged
across folds. In our case, we report the results as the mean across
all folds. Figure 7 shows the top 15 important predictors of age.
FIGURE 4 | Models performance using NCV. Error bar represents the
The color of the bars represents the Pearson’s correlation values standard deviation of performances across the outer loop of NCV.
between each predictor and the age. From the graph, we can

FIGURE 3 | The complete framework for estimating the BrainAGE form EEG. The framework uses the nested-cross-validation method (Figure 2) to build estimation
for the age. Then, those estimations are used to calculate the BrainAGE from the entire dataset.

Frontiers in Aging Neuroscience | www.frontiersin.org 7 July 2018 | Volume 10 | Article 184

Al Zoubi et al. Predicting Age From Brain EEG

FIGURE 5 | Predicted age vs. age constructed from the outer loop of NCV.

FIGURE 7 | The top 15 important features to predict age sorted from most
important (bottom) to top. Ventricle axis shows the scoring values from
stack-ensemble model predictor, while the color indicates the correlation
values between that feature and age.

FIGURE 6 | The BrainAGE variable as a function of the chronological age.

To show the spatial distribution of feature importance, MNE

software (Gramfort et al., 2013) was used. More specifically, the
feature importance scores obtained from the NCV were averaged
based on the feature type and categorized based on the frequency
bands. The resultant mapping for the feature importance scores
is shown in Figure 9. Figure S5 in Supplementary presents the
PDPs for the top features.
Finally, we consider the effect of number of samples on the
performance of predicting age. We tested our framework on
different number of samples. Figure 10 graphs the R2 of NCV
as a function of the number of samples in our dataset.

FIGURE 8 | PDP for the top feature from NCV from Stack-Ensemble model.

DISCUSSION
In the discussion part, we address the results, our research goals Age-Related Changes Are Affecting Brain
and elaborate on different implementation details. In addition, EEG Signals
we compare our results with related work and point out various Results suggest that indeed the aging affects human brain
aspects of differences. EEG signals. We have also determined that, a comprehensive

Frontiers in Aging Neuroscience | www.frontiersin.org 8 July 2018 | Volume 10 | Article 184

Al Zoubi et al. Predicting Age From Brain EEG

FIGURE 9 | Mean feature importance scores sorted by bands and channels for predicting Age. The darker the color, the more important is the feature.

that utilized EEG predictors feature different mechanisms of

influence by age and/or disease. In addition for features
extraction, selecting the best features is important to improve
the performance and reduce the complexity of the model. We
eliminated the correlated features to select the best features,
which improve the overall R2 . Our selection for correlated
features preserves the consistency among NCV folds and
importantly eases the interpretation of the results. The age-
related changes in EEG are strongly supported by the literature
(Benninger et al., 1984; Gasser et al., 1988; Marciani et al., 1994;
Widagdo et al., 1998; Kikuchi et al., 2000; Babiloni et al., 2006;
Hashemi et al., 2016) and by our results as well, where the
correlation between top four features and age was relative high
with r = 0.34, 0.3, 0.26, and 0.24, respectively.

FIGURE 10 | The effect of the number of samples on the age prediction. Can Age Be Predicted From EEG Signals?
Using unbiased prediction of age, NCV, we were able to provide
a reasonable accuracy for predicting age. The best results were
obtained by SVR (R2 = 0.34) and were slightly improved by the
feature extraction is required from EEG signals to capture the Stack-ensemble approach (R2 = 0.37). The correlation between
relationship between chronological ge and the age predictors. predicted age and age (r = 0.6), which shows the ability of our
This suggests that the aging is reflected broadly on the EEG model to predict the age. The overall feature importance scores
signals without selected predominate feature and also suggests were extracted for each fold in the outer loop of NCV and then

Frontiers in Aging Neuroscience | www.frontiersin.org 9 July 2018 | Volume 10 | Article 184

Al Zoubi et al. Predicting Age From Brain EEG

averaged across all folds. The feature importance showed that Predicting age from functional brain imaging is probably
the important predictors are spread out across different features more challenging than structural imaging. One can notice from
types and bands. In addition, we used PDP to examine the Table 1 that fMRI yields generally a lower performance than MRI
consistency of features across the outer loops of NCV, where we data. The best results was reported by Cole et al. (2017) with
showed that top features have a similar behavior across the folds. r = 0.96 from structural imaging of healthy subjects. EEG and
The effect of the number of samples on prediction accuracy fMRI are both functional imaging for the brain and thus it’s
is shown in Figure 10. The graph indicates a potential more subjective to compare EEG results with fMRI results. Our
improvement may be achievable adding more samples. When method’s performance is relatively lower than those from fMRI
testing on 50 samples, the overall accuracy was R2 ∼ = 0.26, works reported in Dosenbach et al. (2010) with R2 = 0.55 and
which shows that the features are informative for predicting age Qin et al. (2015) with MAE = 4.6 years. Without a subjective
even from small number of samples. It should be noted that comparison between EEG and fMRI from the same dataset, it’s
our samples size is relatively smaller than other works, especially hard to draw conclusions about amount of information that
those ones used MRI. each domain embeds. Although fMRI/MRI imaging may yield a
We found no differences in age prediction across female and higher accuracy, but it comes at extra cost and less portability as
male groups. Both groups have a relatively matched average compared to EEG.
chronological age: female group = 34.47 (10.65) and male The contribution of some features in BrainAGE is in line
group = 35.29(10.47). The average predicted age resulted in 34.78 with previous works (Chiang et al., 2011; Zappasodi et al., 2015).
(6.87) for female and male 35.11 (6.04). The MAE was 6.99 (5.10) For instance, our findings show the negative correlation between
and 6.66(4.77) years for female and male groups, respectively. age and alpha power spectra in healthy groups reported in
Mapping of the spatial distribution of feature importance Chiang et al. (2011). This correlation trend could be observed
scores revealed that age predictors are not uniquely in other frequency bands, especially Delta and Theta bands. FD
corresponding to specific channels, frequency band nor to is positively correlated with age for Healthy subjects, which is
a specific feature domain. That is, different features types capture consistent with finding in Zappasodi et al. (2015). However,
some characteristics of EEG, but not the whole relationship. For Zappasodi et al. (2015) showed that FD increases for ages from
example, Figure 7 showed that among the top 15 important 20 to 50 years and then decreases. Since our age limit is 58,
features, the spectral features are positively correlated with age, the pattern is increasing overall for ranges from 18 to 58 years.
while rEEG features are negatively correlated. That is, one type of Figures S6, S7 in Supplementary provide a spatial mapping of
features captures a specific aspect of the relationship between that the correlation values between the spectral and FD features
feature type and the age. Thus, providing heterogeneous features and age.
can improve the predictability of age. This is also supported by
Figure 9, where the spatial distribution of feature importance
scores does not exhibit a uniform representation. Our analysis
CONCLUSIONS
shows that relative contribution of features importance is We have introduced the rigorous framework for BrainAGE
46, 31, 18, 3, and 2% for spectral, rEEG, amplitude, FD, and estimation based on EEG brain signals. Proof-of-concept analysis
connectivity, respectively. It should be noted that the number of showed that, it is possible to build a robust BrainAge estimation
features among different domains are not the same especially that by harnessing both extensive EEG feature representation and
is the case for FD and connectivity features. Similarly, features suitable ML algorithms. ML and NCV play a significant role
contributions are also spread out across bands as follows: 31, 21, in identifying informative features and studying the spatial
27, and 18% for theta, delta, alpha, beta, and theta, respectively. distribution of significant predictors, and providing unbiased
prediction. In addition, we showed how to evaluate and interpret
Comparison With Other Works the results using the feature importance scores and partial
Predicting age from EEG features was also studied in Dimitriadis dependence plots. The introduced framework can be extended
and Salis (2017). Compared with the current study, they reported to test association with and predict other physiological relevant
relatively higher prediction accuracy, 0.6, compared with 0.4 measures based on EEG brain signals.
here. There are a number of differences which may contribute to
this disparity. Perhaps the most significant one is that they seem
to have done feature selection using the response variable and AUTHOR CONTRIBUTIONS
the entire dataset, which will generally lead to more optimistic
evaluations than doing feature selection within a nested cross All authors contributed significantly to work regarding
validation framework, as done here. Additionally, we report R2 conception and design; acquisition, analysis; drafting the work
as 1-SSresid/SStotal (SSresid is the squared residuals from the and revised critically; approval of the version to be published,
regression and SStotal is the total sum of squares of differences and carrying the responsibility for achieving the accuracy or
from the mean) taken from the model prediction, while they integrity of any part of the work.
seem to have reported the R2 of a line fit through Age vs.
Predicted Age. Other differences include the feature sets used FUNDING
and the fact that our data were collected during fMRI, which
may leave some residual artifact. Furthermore, we use here an The study was supported by the Laureate Institute for Brain
interpretation-friendly features. Research and William K. Warren foundation, and in part by

Frontiers in Aging Neuroscience | www.frontiersin.org 10 July 2018 | Volume 10 | Article 184

Al Zoubi et al. Predicting Age From Brain EEG

the W81XWH-12-1-0697 grant from the U.S. Department of scanning. We would like to thank Dr. Patrick Britz, Dr. Robert
Defense, and the P20 GM121312 award from National Institute Störmer, Dr. Mario Bartolo, and Dr. Brett Bays of Brain Products,
of General Medical Sciences, National Institutes of Health. GmbH for their help and technical support.

ACKNOWLEDGMENTS SUPPLEMENTARY MATERIAL

We are particularly grateful to Dr. Vadim Zotev for continuous The Supplementary Material for this article can be found
help with EEG hardware, Julie Owen, Bill Alden, Julie DiCarlo, online at: https://www.frontiersin.org/articles/10.3389/fnagi.
and Greg Hammond for helping with MRI and EEG-fMRI 2018.00184/full#supplementary-material

REFERENCES Friedman, J., Hastie, T., and Tibshirani, R. (2001). The Elements of Statistical
Learning. New York, NY: Springer Series in Statistics.
Allen, P. J., Josephs, O., and Turner, R. (2000). A method for removing imaging Gasser, T., Verleger, R., Bächer, P., and Sroka, L. (1988). Development
artifact from continuous EEG recorded during functional MRI. Neuroimage 12, of the EEG of school-age children and adolescents. I. Analysis
230–239. doi: 10.1006/nimg.2000.0599 of band power. Electroencephalogr. Clin. Neurophysiol. 69, 91–99.
Al Zoubi, O., Awad, M., and Kasabov, N. K. (2018). Anytime multipurpose doi: 10.1016/0013-4694(88)90204-0
emotion recognition from EEG data using a liquid state machine based Gramfort, A., Luessi, M., Larson, E., Engemann, D. A., Strohmeier, D., Brodbeck,
framework. Artif. Intell. Med. 86, 1–8. doi: 10.1016/j.artmed.2018.01.001 C., et al. (2013). MEG and EEG data analysis with MNE-Python. Front.
Ashburner, J. (2007). A fast diffeomorphic image registration algorithm. Neurosci. 7:267. doi: 10.3389/fnins.2013.00267
Neuroimage 38, 95–113. doi: 10.1016/j.neuroimage.2007.07.007 Harmony, T., Marosi, E., De León, A. E. D., Becker, J., and Fernández, T.
Babiloni, C., Binetti, G., Cassarino, A., Dal Forno, G., Del Percio, C., (1990). Effect of sex, psychosocial disadvantages and biological risk factors
Ferreri, F., et al. (2006). Sources of cortical rhythms in adults during on EEG maturation. Electroencephalogr. Clin. Neurophysiol. 75, 482–491.
physiological aging: a multicentric EEG study. Hum. Brain Mapp. 27, 162–172. doi: 10.1016/0013-4694(90)90135-7
doi: 10.1002/hbm.20175 Hashemi, A., Pino, L. J., Moffat, G., Mathewson, K. J., Aimone, C., Bennett,
Babiloni, C., Triggiani, A. I., Lizio, R., Cordone, S., Tattoli, G., Bevilacqua, V., P. J., et al. (2016). Characterizing population EEG dynamics throughout
et al. (2016). Classification of single normal and Alzheimer’s disease individuals adulthood. eNeuro 3:ENEURO. 0275-0216.2016. doi: 10.1523/ENEURO.0275-
from cortical sources of resting state EEG rhythms. Front. Neurosci. 10:47. 16.2016
doi: 10.3389/fnins.2016.00047 Jenke, R., Peer, A., and Buss, M. (2014). Feature extraction and selection for
Bell, A. J., and Sejnowski, T. J. (1995). An information-maximization approach emotion recognition from EEG. IEEE Trans. Affect. Comput. 5, 327–339.
to blind separation and blind deconvolution. Neural Comput. 7, 1129–1159. doi: 10.1109/TAFFC.2014.2339834
doi: 10.1162/neco.1995.7.6.1129 Kikuchi, M., Wada, Y., Koshino, Y., Nanbu, Y., and Hashimoto, T. (2000).
Benninger, C., Matthis, P., and Scheffner, D. (1984). EEG development of healthy Effect of normal aging upon interhemispheric EEG coherence: analysis
boys and girls. Results of a longitudinal study. Electroencephalogr. Clin. during rest and photic stimulation. Clin. Electroencephalogr. 31, 170–174.
Neurophysiol. 57, 1–12. doi: 10.1016/0013-4694(84)90002-6 doi: 10.1177/155005940003100404
Càmara, E., Bodammer, N., Rodríguez-Fornells, A., and Tempelmann, C. (2007). Kuhn, M. (2008). Caret Package. J. Stat. Softw. 28, 1–26. doi: 10.18637/jss.v028.i05
Age-related water diffusion changes in human brain: a voxel-based approach. Lao, Z., Shen, D., Xue, Z., Karacali, B., Resnick, S. M., and Davatzikos, C.
Neuroimage 34, 1588–1599. doi: 10.1016/j.neuroimage.2006.09.045 (2004). Morphological classification of brains via high-dimensional shape
Chiang, A., Rennie, C., Robinson, P., Van Albada, S., and Kerr, C. (2011). Age transformations and machine learning methods. Neuroimage 21, 46–57.
trends and sex differences of alpha rhythms including split alpha peaks. Clin. doi: 10.1016/j.neuroimage.2003.09.027
Neurophysiol. 122, 1505–1517. doi: 10.1016/j.clinph.2011.01.040 Liem, F., Varoquaux, G., Kynast, J., Beyer, F., Masouleh, S. K., Huntenburg,
Clarke, A. R., Barry, R. J., McCarthy, R., and Selikowitz, M. (2001). Age and sex J. M., et al. (2017). Predicting brain-age from multimodal imaging
effects in the EEG: development of the normal child. Clin. Neurophysiol. 112, data captures cognitive impairment. Neuroimage 148, 179–188.
806–814. doi: 10.1016/S1388-2457(01)00488-6 doi: 10.1016/j.neuroimage.2016.11.005
Cole, J. H., Poudel, R. P., Tsagkrasoulis, D., Caan, M. W., Steves, C., Spector, T. Lindsley, D. B. (1939). A longitudinal study of the occipital alpha rhythm in normal
D., et al. (2017). Predicting brain age with deep learning from raw imaging children: frequency and amplitude standards. Pedagog. Semin. J. Genet. Psychol.
data results in a reliable and heritable biomarker. Neuroimage 163, 115–124. 55, 197–213. doi: 10.1080/08856559.1939.10533190
doi: 10.1016/j.neuroimage.2017.07.059 Lu, A. T., Hannon, E., Levine, M. E., Crimmins, E. M., Lunnon, K., Mill,
Cragg, L., Kovacevic, N., McIntosh, A. R., Poulsen, C., Martinu, K., J., et al. (2017). Genetic architecture of epigenetic and neuronal ageing
Leonard, G., et al. (2011). Maturation of EEG power spectra in rates in human brain regions. Nat. Commun. 8:15353. doi: 10.1038/ncomms
early adolescence: a longitudinal study. Dev. Sci. 14, 935–943. 15353
doi: 10.1111/j.1467-7687.2010.01031.x Marciani, M. G., Maschio, M., Spanedda, F., Caltagirone, C., Gigli, G., and
Delorme, A., and Makeig, S. (2004). EEGLAB: an open source toolbox for analysis Bernardi, G. (1994). Quantitative EEG evaluation in normal elderly subjects
of single-trial EEG dynamics including independent component analysis. J. during mental processes: age-related changes. Int. J. Neurosci. 76, 131–140.
Neurosci. Methods 134, 9–21. doi: 10.1016/j.jneumeth.2003.10.009 doi: 10.3109/00207459408985998
Dimitriadis, S. I., and Salis, C. I. (2017). Mining time-resolved functional brain Marshall, P. J., Bar-Haim, Y., and Fox, N. A. (2002). Development of the
graphs to an EEG-based chronnectomic brain aged index (CBAI). Front. Hum. EEG from 5 months to 4 years of age. Clin. Neurophysiol. 113, 1199–1208.
Neurosci. 11:423. doi: 10.3389/fnhum.2017.00423 doi: 10.1016/S1388-2457(02)00163-3
Dosenbach, N. U., Nardos, B., Cohen, A. L., Fair, D. A., Power, J. D., Church, J. A., Matthis, P., Scheffner, D., Benninger, C., Lipinski, C., and Stolzis, L.
et al. (2010). Prediction of individual brain maturity using fMRI. Science 329, (1980). Changes in the background activity of the electroencephalogram
1358–1361. doi: 10.1126/science.1194144 according to age. Electroencephalogr. Clin. Neurophysiol. 49, 626–635.
Franke, K., Ziegler, G., Klöppel, S., Gaser, C., and Initiative, A. S. D. N. (2010). doi: 10.1016/0013-4694(80)90403-4
Estimating the age of healthy subjects from T 1-weighted MRI scans using Mayeli, A., Zotev, V., Refai, H., and Bodurka, J. (2016). Real-Time EEG
kernel methods: exploring the influence of various parameters. Neuroimage 50, artifact correction during fMRI using ICA. J. Neurosci. Methods 274, 27–37.
883–892. doi: 10.1016/j.neuroimage.2010.01.005 doi: 10.1016/j.jneumeth.2016.09.012

Frontiers in Aging Neuroscience | www.frontiersin.org 11 July 2018 | Volume 10 | Article 184

Al Zoubi et al. Predicting Age From Brain EEG

Nenadić, I., Dietzek, M., Langbein, K., Sauer, H., and Gaser, C. (2017). Victor, T. A., Khalsa, S. S., Simmons, W. K., Feinstein, J. S., Savitz, J., Aupperle,
BrainAGE score indicates accelerated brain aging in schizophrenia, R. L., et al. (2018). Tulsa 1000: a naturalistic study protocol for multilevel
but not bipolar disorder. Psychiatry Res. Neuroimaging 266, 86–89. assessment and outcome prediction in a large psychiatric sample. BMJ Open
doi: 10.1016/j.pscychresns.2017.05.006 8:e016620. doi: 10.1136/bmjopen-2017-016620
Niazy, R., Beckmann, C., Iannetti, G., Brady, J., and Smith, S. (2005). Removal Widagdo, M., Pierson, J., and Helme, R. (1998). Age-related changes
of FMRI environment artifacts from EEG data using optimal basis sets. in qEEG during cognitive tasks. Int. J. Neurosci. 95, 63–75.
Neuroimage 28, 720–737. doi: 10.1016/j.neuroimage.2005.06.067 doi: 10.3109/00207459809000650
Pardoe, H. R., Cole, J. H., Blackmon, K., Thesen, T., Kuzniecky, R., and Wong, C.-K., Luo, Q., Zotev, V., Phillips, R., Kam Wai Clifford, C., and Bodurka, J.
Investigators, H. E. P. (2017). Structural brain changes in medically refractory (2018). Automatic cardiac cycle determination directly from EEG-fMRI data
focal epilepsy resemble premature brain aging. Epilepsy Res. 133, 28–32. by multi-scale peak detection method. J. Neurosci. Methods. 304, 168–184.
doi: 10.1016/j.eplepsyres.2017.03.007 doi: 10.1016/j.jneumeth.2018.03.017
Qin, J., Chen, S.-G., Hu, D., Zeng, L.-L., Fan, Y.-M., Chen, X.-P., et al. (2015). Wong, C.-K., Zotev, V., Misaki, M., Phillips, R., Luo, Q., and Bodurka, J.
Predicting individual brain maturity using dynamic functional connectivity. (2016). Automatic EEG-assisted retrospective motion correction for fMRI
Front. Hum. Neurosci. 9:418. doi: 10.3389/fnhum.2015.00418 (aE-REMCOR). Neuroimage 129, 133–147. doi: 10.1016/j.neuroimage.2016.
Saletu, B., Paulus, E., Linzmayer, L., Anderer, P., Semlitsch, H. V., Grünberger, J., 01.042
et al. (1995). Nicergoline in senile dementia of Alzheimer type and multi-infarct Zappasodi, F., Marzetti, L., Olejarczyk, E., Tecchio, F., and Pizzella,
dementia: a double-blind, placebo-controlled, clinical and EEG/ERP mapping V. (2015). Age-related changes in electroencephalographic signal
study. Psychopharmacology 117, 385–395. doi: 10.1007/BF02246209 complexity. PLoS ONE 10:e0141995. doi: 10.1371/journal.pone.01
Toole, J. M., and Boylan, G. B. (2017). NEURAL: quantitative features for newborn 41995
EEG using Matlab. arXiv preprint arXiv:1704.05694
Triggiani, A. I., Bevilacqua, V., Brunetti, A., Lizio, R., Tattoli, G., Cassano, Conflict of Interest Statement: The authors declare that the research was
F., et al. (2017). Classification of healthy subjects and Alzheimer’s conducted in the absence of any commercial or financial relationships that could
disease patients with dementia from cortical sources of resting state EEG be construed as a potential conflict of interest.
rhythms: a study using artificial neural networks. Front. Neurosci. 10:604.
doi: 10.3389/fnins.2016.00604 Copyright © 2018 Al Zoubi, Ki Wong, Kuplicki, Yeh, Mayeli, Refai, Paulus and
Valizadeh, S., Hänggi, J., Mérillat, S., and Jäncke, L. (2017). Age prediction on Bodurka. This is an open-access article distributed under the terms of the Creative
the basis of brain anatomical measures. Hum. Brain Mapp. 38, 997–1008. Commons Attribution License (CC BY). The use, distribution or reproduction in
doi: 10.1002/hbm.23434 other forums is permitted, provided the original author(s) and the copyright owner(s)
Varma, S., and Simon, R. (2006). Bias in error estimation when using are credited and that the original publication in this journal is cited, in accordance
cross-validation for model selection. BMC Bioinformatics 7:91. with accepted academic practice. No use, distribution or reproduction is permitted
doi: 10.1186/1471-2105-7-91 which does not comply with these terms.

Frontiers in Aging Neuroscience | www.frontiersin.org 12 July 2018 | Volume 10 | Article 184