2634 IEEE SENSORS JOURNAL, VOL. 19, NO.

7, APRIL 1, 2019

EEG-Based Age and Gender Prediction Using Deep

BLSTM-LSTM Network Model
Pallavi Kaushik , Anmol Gupta , Partha Pratim Roy, Member, IEEE, and Debi Prosad Dogra , Member, IEEE

Abstract— With the rapid development of brain–computer three EEG is less expensive and complex as per the recording
interfaces (BCI), the number of applications that use BCI mechanism. Therefore, it is extensively used for developing
technology is increasingly thick and fast. Prediction of age and BCI applications. EEG recording devices can measures the
gender of a person through EEG analysis is a new application of
BCI that has been proposed in this paper. An industry standard electrical signals of brain using electrodes attached to the
EEG recording device has been used to record cerebral activities scalp. It is available in both invasive and non invasive forms,
of 60 subjects (both male and female) in relaxed position with though the non-invasive devices are more popular due to
closed eyes. Deep BLSTM-LSTM network has been used to affordable pricing and ease of use.
construct a hybrid learning framework for the aforementioned Ageing refers to the process of inevitably growing older that
analysis. Accuracy of 93.7% and 97.5% have been recorded for
age and gender classification problems respectively. These values is genetically determined and evolutionary modulated. While
are better than the state-of-the-art methods. Our analysis also growing older, there are several changes to the Central Nervous
reveals that the beta band frequencies are better in predicting System (CNS) which may lead to diseases like Alzheimer,
the age and gender as compared to other frequency bands of Dementia, Parkinson etc. The changes in the Central Nervous
the EEG signals. The proposed method has several applications, System (CNS) can be diagnosed with the help of EEG signal
including biometric, health-care, entertainment, and targeted
advertisements. analysis [32]–[34]. As far as gender is concerned, male and
female brains differ functionally, anatomically [35]–[37], and
Index Terms— Age detection, BCI, BLSTM, deep learning, in the chance of development of the neuropsychiatric diseases
EEG, gender detection.
and how they react to the treatment [38].
I. I NTRODUCTION There are some studies which have used traditional machine
learning approaches like LDA, SVM, Random Forest etc. to
B RAIN Computer Interface or BCI is an emerging field
which allows humans to communicate with comput-
ing devices without using conventional input systems. This
analyse the brain signals measured using EEG for various BCI
applications [1]–[4], [41]. The problem with the traditional
communication in turn allows several applications which approaches is that, some assumptions are made in order to
assist humans to perform varying tasks.The applications extract the features that are relevant to that particular appli-
can be broadly categorized into medical and non-medical cation. These assumptions are based on the theories that may
types. For example, a person with hearing impairment, vision not be true across all applications. As a result, some important
issues or limited physical movements can perform various features may get excluded during traditional feature extraction
tasks with the help of neuroprosthetics. The non-medical appli- process. However, deep learning model tries to learn features
cations include authentication [6], [7], controlling appliances, that may not be possible to extract using traditional feature
playing games [8], typing [9], etc. All such applications can be extraction techniques. In this paper, we propose a new method
assisted using brain signals. Human brain works with the help to predict the age and gender of a person through EEG signals
of neurons that are connected with other neurons using axons recorded using a portable 14-channel EEG recording setup. Bi-
and dendrites. The activity in our brain is a by-product of these directional LSTM has been used to predict the age and gender
neurons firing when a certain threshold of electrical energy is of the person, which is possibly the first of approach of this
achieved. Some of the electrical signals when travelling inside kind and the framework is called as “Deep BLSTM-LSTM”.
the brain escape the protective layer of fatty white substance The remainder of this paper is organized as follows.
called the myelin, and can be recorded. These signals can Section II reviews the related work. Section III gives the
be interpreted for understanding. Cerebral activities can be methodology used for the dataset collection, pre-processing
recorded using EEG, ECoG, or fMRI devices. Out of these and architecture of the proposed Deep BLSTM-LSTM model.
Section IV discusses the results for different models and
Manuscript received September 23, 2018; revised November 21, 2018; compare them. Finally, we make a conclusion in Section V.
accepted November 22, 2018. Date of publication December 7, 2018; date
of current version March 7, 2019. The associate editor coordinating the The main contributions of the paper include,
review of this paper and approving it for publication was Dr. Ravibabu 1) Beta band contains more important information than
Mulaveesala. (Pallavi Kaushik and Anmol Gupta contributed equally to this
work.) (Corresponding author: Pallavi Kaushik.) other bands for prediction of age and gender.
P. Kaushik, A. Gupta, and P. P. Roy are with the Department of Computer 2) Construction of a deep hybrid framework using
Science and Engineering, Indian Institute of Technology, Roorkee 247667, Bi-directional LSTM and LSTM to map human cerebral
India (e-mail: [email protected]).
D. P. Dogra is with the School of Electrical Sciences, IIT Bhubaneswar, activities with age and gender features.
Bhubaneswar 752050, India. 3) Prediction of age and gender of human participants with
Digital Object Identifier 10.1109/JSEN.2018.2885582 the help of above deep learning framework.
1558-1748 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: K.L. University(F.K.A. Koneru Lakshmaiah College of Engineering). Downloaded on February 16,2022 at 05:55:01 UTC from IEEE Xplore. Restrictions apply.
 KAUSHIK et al.: EEG-BASED AGE AND GENDER PREDICTION USING DEEP BLSTM-LSTM NETWORK MODEL 2635

4) Comparison of different LSTM, Bi-directional LSTM (LRP). The algorithm tries to understand the functioning of
architectures and GRU. deep learning by investigating which features are actually
responsible for age and gender prediction. The study
II. L ITERATURE R EVIEW shows that appropriate model initialisation via pretraining
This section summarises various studies which are presently counteracts overfitting, leads to a holistic perception of the
available to interpret the relation between EEG wave signals input that essentially leads to better performance. Majority of
and age/gender. The difference in frequency of brain waves the existing research works compare the results on the latest
as a consequence of age was first studied as early as 1939. benchmark datasets such as Adience [19] and lfw dataset.1
Lindsey [5] found that, there is an increase in the emission of The state-of-the-art accuracy in gender recognition is 92.8%
the waves with higher frequencies and reduction in the waves and 62.8% for age detection.
with lower frequencies in various regions of the brain. Since In [20]–[24] the authors have applied long short term
then, several research works [28]–[31] have studied the effects memory (LSTM) [25] on the EEG signals in applications
of age and gender at different frequency bands of the EEG such as emotion detection, lapse in attention, seizure detection,
signals. The results of these studies suggest that, in terms brain activity etc. Ni et al. [22] present a study on the brain
of age, delta and theta activity are possibly the prominent fog, i.e., the state when the brain is confused that results in low
bands in terms of brain activity until the age of four, while performance while doing tasks which require clear thinking.
alpha and beta activities increase throughout the childhood. Detecting a confused state can lead to several applications
Therefore, change in cerebral activity is highly dependent on in online education, offline education, driver fatigue detection
age and hence there is a need to study such effects for various etc. They have applied Bidirectional LSTM recurrent neural
applications. Some of these studies [35]–[38] also suggest that network on the EEG data to classify students’ confusion while
gender is responsible for the maturity of brain signals. It has watching online videos. They achieve the accuracy of 73.3%.
also been reported that girls maturity typically lag in the EEG Gupta et. al. [44] detected emotions using EEG to enhance
as compared to boys [43]. the text users write on social media with 74.9% accuracy with
However, the use of machine learning for studying data of 25 participants. Kumar et. al. [48] have used EEG
the relationship between various statistical parameters and in sentiment analysis. Khurana et. al. [40] have used the EEG
EEG signals as a function of age, has been first studied signals for predicting how much a user is familiar with certain
by Paiva et al. [4]. This study establishes the relationship English words. Yadava et. al. [27] have used the EEG signals
between the EEG signals and age. EEG signals of 59 subjects to predict whether a user likes or dislikes certain products that
was recorded which were divided into 7 groups, ageing are sold on an e-commerce website.
between 20 to 86 years from both gender. They used Linear Putten et al. [26] have applied CNN on EEG data to predict
Discriminant Analysis(LDA) for measuring the correlation the gender of a person. It shows that fast beta activity (2025Hz)
between different age groups and found that indeed there were is the main distinctive feature for classifying gender and
significant differences among the age groups. achieves accuracy as high as 80%. The authors have collected
Nguyen et al. [1] have proposed a framework which uses data of 1308 subjects using 26-channel EEG with 47% male
PARAFAC [2] and SVM [3]. They have used EEG features subjects and 53% female subjects. As far as we know, this is
in addition to speech-based features as inputs to the parallel the only paper that has used deep learning on EEG data to
factor analysis. The features are, namely spectral power, predict gender.
relative power, and discrete cosine transform. They further use Our paper emphasises on applying LSTM architecture in
multilinear partial least squares (N-PLS) and SVM to improve bidirectional modes to predict the age and gender which has
the classification accuracy and recorded 92.5% and 93.8% for not been done earlier.
age and two class gender classification. Kaur et al. [41] have III. M ETHODOLOGY
used random forest for predicting age and gender. EEG signals
of 60 healthy subjects have been recorded using 14 channel This section describes the dataset, preprocessing and the
Emotiv Epoch Plus and classification accuracy of 88.33% have proposed architecture of the Deep BLSTM-LSTM model
been recorded for age and an accuracy of 96.66% has been which comprises of the BLSTM, LSTM and dense layers, for
achieved for gender classification. classifying the gender and age of an individual by analysing
Franke et al. [10] have proposed a framework for predict- EEG signals. A comparison of different LSTM architectures,
ing age of 650 healthy subjects aged between 19-86 years, namely single LSTM, Deep LSTM and BLSTM has also been
from MRI brain scans and used relevance vector machine provided in all five bands. The results have been compared
(RVM) [11] regression for the prediction which yielded a mean with raw EEG data. The overall framework is presented
absolute error of 5 years. in Fig. 1.
In [12]–[15], authors have used face images to predict age
and gender using deep learning, specifically different versions A. Description of the Dataset
of convolution neural network like GoogleNet citeb16, This work uses the dataset proposed in [41]. The dataset
CaffeNet [17], VGG-16 [18] or hybrid models. comprises of the EEG signals of 60 individuals with 35 males
Lapuschkin et al. [13] provide a comparative study on deep and 25 females ranging from 6 to 55 years. The individuals
convolution neural networks with their proposed method of
initialisation, refereed to as Layerwise Relevance Propagation 1 http://vis-www.cs.umass.edu/lfw/

Authorized licensed use limited to: K.L. University(F.K.A. Koneru Lakshmaiah College of Engineering). Downloaded on February 16,2022 at 05:55:01 UTC from IEEE Xplore. Restrictions apply.
 2636 IEEE SENSORS JOURNAL, VOL. 19, NO. 7, APRIL 1, 2019

TABLE I
A S UMMARY OF THE E XISTING R ESEARCH W ORK IN T HIS D OMAIN

Fig. 1. A High-level representation of the functional block diagram of the proposed age and gender recognition system.

TABLE II
D ATASET D ESCRIPTION

are divided into 6 age groups. A more detailed overview of Fig. 2. Position of the electrodes according to the 10-20 international system.
the dataset can be found in the Table II. The EEG signals
were collected for 10 seconds and individuals were asked B. Preprocessing Using Discrete Wavelet Transform (DWT)
to sit in a relaxed position with their eyes closed. Emotive
Epoch plus device was used to capture the EEG signals using The EEG signals comprises of waves of various frequencies.
14-electrodes namely, AF3, AF4, F3, F4, F7, F8, FC5,FC6, Neuroscientists have broken them into frequency bands with
P7, P8, T7, T8, O1, O2 and 2 references (CMS and DRL) each frequency band responsible for a particular activity of
as per the International 10-20 system, shown in Fig. 2. The the brain. Various frequency bands are also referred to as the
brain maps showing the difference in the EEG signals for both brain waves. The various brainwaves along with the activity
genders and various age groups, are presented in [41]. responsible for them are as follows:

Authorized licensed use limited to: K.L. University(F.K.A. Koneru Lakshmaiah College of Engineering). Downloaded on February 16,2022 at 05:55:01 UTC from IEEE Xplore. Restrictions apply.
 KAUSHIK et al.: EEG-BASED AGE AND GENDER PREDICTION USING DEEP BLSTM-LSTM NETWORK MODEL 2637

Fig. 3. Diagram showing the raw data and various brainwaves, namely delta, beta, theta gamma and alpha obtained after DWT for a sample user taken from
the dataset.

• Delta- These are the waves that lie in the frequency range n = 0, 1, . . . , 2m − 1, T is the length of the signal, a0 and b0
of 0.5-4 Hz and correspond to deep meditation and deep values are set to 2 and 1, respectively.
and dreamless sleep.  ∞
• Theta- These waves fall into the frequency range ψ(t) dt = 0 (1)
of 4-8 Hz and correspond to day dreaming, drowsi- −∞
−m/2
ness or sleep. ψm,n (t) = a0 ψ(a0m t − nb0 ) (2)
• Alpha- These brainwaves lie between 8-12 Hz and cor-
respond to the resting state of the brain i.e calmness and The scaling and wavelet functions required for the evaluation
A and D coefficients given in (3) and (4), respectively.
ideal relaxation.
• Beta- These waves lie within 12-30 Hz and corresponds φ p,q (t) = 2 p/2 h(2 p t − q) (3)
to active thinking, consistent focus, memory recall, high p/2 p
ω p,q (t) = 2 g(2 t − q) (4)
alert and anxiety.
• Gamma- These brain waves are comprised of the fre- The Ai and Di coefficients at the i t h level are evaluated using
quency >30 Hz and they correspond to information (5) and (6), respectively.
processing, cognition, learning and matching of recog-
1 
nized objects, shapes, sounds, etc. Ai = √ x(t)φ p,q (t) (5)
T t
As various activities of the brain correspond to a brain wave, 1 
it makes sense to work with the individual brain waves rather Di = √ x(t)ω p,q (t) (6)
than the whole EEG signal depending on the type of activity T t
one is trying to analyze. The Discrete Wavelet Transformation The Db-8 decomposition gives remaining four wavelet coef-
breaks the input signals into various constituting small range ficients that correspond to noise and five wavelet coefficients
frequency bands via decomposition of the input wave at corresponding to alpha, beta, delta, gamma and theta brain
various levels and it also helps in analyzing the nonstationarity waves. The raw EEG data from an individual is shown in Fig. 3
in the signals. along with corresponding DB-8 wavelet decomposition out-
The DWT is done by obtaining the Approximation (A) puts.
and Detail (D) coefficients by performing decomposition at
various levels. It is done through the use of low and high pass
filters at various levels. The low pass filter (L) ignores the C. Adopted Deep Learning Architecture
high frequency fluctuations and helps in preserving the low Experiments have been conducted using varying setups of
frequency trends. Similarly, the high pass filter (H) helps in LSTM and Deep BLSTM+LSTM models. We present two
keeping the high frequency fluctuations and ignoring the slow specific architecture of LSTM and BLSTM that have worked
trends in the signal. The outcomes from the low pass filters better than the other models in the present context.
help in forming the approximation coefficients and those from 1) Deep Long Short Term Memory (LSTM): LSTMs [25]
the high pass filters help in forming the detailed coefficients. are a form of Recurrent Neural Network (RNN) which over-
The Daubechies (Db)-8 wavelet transform has been used in come the issue of exploding and vanishing gradient in the
this work, hence the levels of decomposition are 8. RNNs. The building blocks of LSTM consists of a cell,
The wavelet function at time t is shown in (1) and (2), an input gate, an output gate and a forget gate. The cell is
where m = 0, 1, . . . , M −1, t = 0, 1, . . . , T −1, M = log2 (t), responsible for handling the long term dependency while the

Authorized licensed use limited to: K.L. University(F.K.A. Koneru Lakshmaiah College of Engineering). Downloaded on February 16,2022 at 05:55:01 UTC from IEEE Xplore. Restrictions apply.
 2638 IEEE SENSORS JOURNAL, VOL. 19, NO. 7, APRIL 1, 2019

Fig. 5. Results obtained using the brainwaves and the deep LSTM model.

Fig. 4. Architecture of the proposed Deep BLSTM-LSTM model, with one

BLSTM and two LSTM layers followed by dropout, batch normalization and hidden neurons. Another batch normalization layer is applied
dense layers.
then which is further followed by another LSTM layer with
64 hidden neurons and another batch normalization layer.
three gates regulates the flow of values between the different A dense layer of 32 neurons follows these layers and finally it
layers of the LSTM network. is connected to another dense layer comprising of 6 neurons.
Experiments were done and the LSTM model that gave the This is used for classifying the EEG signals for each age
best results for age and gender prediction is being described group. For gender classification the same architecture is used
in this Subsection. The model consists of a LSTM layer with the only difference of 1 neuron in the last dense layer.
comprising of 128 hidden neurons. Next, a batch normalization The total number of parameters in the described model are
layer is applied which is further followed by another LSTM 939, 750 with 938, 342 being trainable. Through experiments,
layer with 64 hidden neurons and another batch normalization it has been found that the beta waves give better classification
layer. A dense layer of 32 neurons follows these layers and accuracy for determining both the age and gender of indi-
finally it is connected to another dense layer comprising of 6 viduals. A complete overview of the architecture is presented
neurons (age prediction) and 1 neuron for gender prediction. in Fig. 4. The model has been trained for 50 epochs.
2) BLSTM: Various BLSTM models were experimented
with but this Subsection only describes the model that gave the IV. R ESULTS
best results. The model consists of a BLSTM layer followed The training, testing and validation sets are composed of
by a dropout of 0.2 followed by batch normalization. A dense 60%, 30% and 10% of the whole data. The Deep BLSTM-
layer of 32 neurons follows that is followed by another LSTM, BLSTM and LSTM models described above have
dense layer having 6 neurons for different age groups for age been used to experiment with each of alpha, beta, theta, delta
classification and a single neuron for gender classification as and gamma brain waves. Beta brainwave gives the highest
it is a binary classification. accuracy of 97.52% for gender as well as maximum accuracy
3) Proposed Deep BLSTM-LSTM: The Deep BLSTM- of 93.69% for predicting the age of individuals using the
LSTM architecture consists of three stacked LSTM layers stacked Deep BLSTM-LSTM.
with 256, 128 and 64 units, respectively. The first layer is Experiments have been conducted on the raw dataset in
subjected to a dropout of 0.2 followed by batch normalization, order to draw the parallel between with DWT and without
while the rest of the two LSTM layers are accompanied with DWT. Categorical cross-entropy and binary cross-entropy have
batch normalization only. A dense layer of 32 neurons follows been used as the loss functions for age and gender prediction,
that is followed by another dense layer having 6 neurons for respectively. The accuracies of various brainwaves obtained
different age groups in case of age classification. It has a by using the deep LSTM, BLSTM and Deep BLSTM-LSTM
single neuron in case of gender classification as it has only models are as shown in the Fig. 5, Fig. 6 and Fig. 7. It can
two classes, male and female. The model has been trained be seen that the alpha and beta waves produce better results
for 50 epochs with 10-fold cross validation. The number of than the rest of the brainwaves and also that the prediction
trainable parameters has been found to be 528, 230 for age results obtained from the raw data can be highly improved
and 527, 041 for gender prediction. The results obtained with using the Discrete Wavelet Transform from 83.27% and 84.6%
the architecture are discussed in the next section. to 93.69% and 97.52% for age and gender classification,
After obtaining various brainwaves from the DWT trans- respectively.
form, each brainwave is analyzed in order to know which brain The confusion matrix for the age prediction using the alpha
wave is more responsible for determining the age and gender brain wave for the Deep BLSTM-LSTM model is shown
of an individual. in Fig. 8. It can be seen from the Fig. 8 that the age
The EEG signals are first fed into a BLSTM layer of group of 18 − 23 corresponding to class 2 has the least
256 neurons and then into a dropout layer with dropout misclassifications, where only 2 individuals out of 3642 are
rate of 0.2. Next, batch normalization is performed and the categorised into the age group of 25 − 29 (Class 3). The
results are then fed into a LSTM layer comprising of 128 maximum misclassifications can be seen in the age group of

Authorized licensed use limited to: K.L. University(F.K.A. Koneru Lakshmaiah College of Engineering). Downloaded on February 16,2022 at 05:55:01 UTC from IEEE Xplore. Restrictions apply.
 KAUSHIK et al.: EEG-BASED AGE AND GENDER PREDICTION USING DEEP BLSTM-LSTM NETWORK MODEL 2639

TABLE III
C OMPARATIVE A NALYSIS W ITH THE R ECENT S TATE - OF - THE -A RT T ECHNIQUES

Fig. 6. Results obtained using the brainwaves via the BLSTM model.

Fig. 8. Confusion matrix for age classification on the beta brainwave using
the Deep BLSTM-LSTM model.

Fig. 7. Results obtained using the brainwaves via the Deep BLSTM-LSTM
model.

25 −29, where 89 out of 2905 individuals have been classified

into the age group of 12 − 15 (Class 1). The confusion
matrix for categorising the gender using the beta brainwave
is shown in Fig. 9, where the model misclassifies 38 male
members as females of the total 1469 males and it misclassifies Fig. 9. Confusion matrix for gender classification on the beta brainwave
119 females of the total 3612 female individuals. using the Deep BLSTM-LSTM model.
The Deep BLSTM-LSTM model also gives better accuracy
than the state-of-the-art models for predicting the age and
gender of various individuals. Kaur et al. [41] have used the using the Savitzky-Golay filter of degree 2 and extracted
random forest classifier on the same dataset and reported the various features like mean, root mean square and energy
accuracies of 88.33% and 96.66% for age and gender classi- from the brainwaves obtained after the application of DWT.
fication, respectively. They have smoothened the EEG signals Levi and Hassner [43] have proposed a Convolutional Neural

Authorized licensed use limited to: K.L. University(F.K.A. Koneru Lakshmaiah College of Engineering). Downloaded on February 16,2022 at 05:55:01 UTC from IEEE Xplore. Restrictions apply.
 2640 IEEE SENSORS JOURNAL, VOL. 19, NO. 7, APRIL 1, 2019

value at time t, recurrent models such as LSTM and BLSTM

help in preserving time dependencies and thus perform better
than the vanilla neural network or traditional machine learning
approaches. Thus, we have developed a architecture called
Deep BLSTM-LSTM which uses one layer of BLSTM fol-
lowed by two layers of LSTM which in turn are followed by
two fully connected dense layers. The recorded accuracies are
as high as 93.69% and 97.5% for age and gender classification,
respectively. Moreover, no smoothing has been performed that
reduces computational burden and other overhead. However,
raw data is segregated into different frequency bands to analyse
how brain activity is affected with the change in age of a
person and whether there is a difference in the brain activity
Fig. 10. Accuracies obtained for age classification using label encoding.
among genders. It is found that, the results are in accordance
with the previous results mentioned in the literature. In our
Network for the prediction of age and gender of individuals results, beta band waves provide a better prediction accuracy
by using a large dataset comprising of 26, 000 facial images. for gender and age classification.
They reported the accuracy of 50.7% and 86.6% for age and In future, more data can be collected for the purpose of the
gender classification, respectively. experiment. The age groups created can be expanded to 5-80
We have also computed the classification accuracies using years so that the elderly people can also use the application.
GRU (Gated Recurrent Unit) in order to establish the effec- Moreover, the technology can be mixed with several existing
tiveness of LSTMs to GRU. The architecture used is analogous technologies to create new applications with the help of EEG
to the Deep BLSTM-LSTM model, which gave the best devices. For example, we can use the brain segmentation
classification accuracies amongst the LSTM based models. technology [45] to further improve the results and we can also
Using GRU, beta brainwave gave an accuracy of 92.14% and leverage the cloud computing facilities [49] for the purpose of
95.14% for age and gender classification respectively. sentiment analysis [46] and for better visualisation [47].
We have also experimented with Convolutional Neural
Networks (CNNs) and found the accuracies less than 30% and
R EFERENCES
less than 50% for age and gender classification respectively.
As CNNs are good at learning the features on thier own, [1] P. Nguyen, D. Tran, T. Vo, X. Huang, W. Ma, and D. Phung, “EEG-
we performed the experiments on the raw EEG signals. CNNs based age and gender recognition using tensor decomposition and speech
tend to work best on images and RNN models for temporal features,” in Proc. Int. Conf. Neural Inf. Process., 2013, pp. 632–639.
[2] R. A. Harshman and M. E. Lundy, “PARAFAC: Parallel factor analysis,”
data, hence the LSTM, BLSTM, GRU and Deep LSTM- Comput. Statist. Data Anal., vol. 18, no. 1, pp. 39–72, 1994.
BLSTM perform much better than the CNN model for the [3] M. A. Hearst, S. T. Dumais, E. Osuna, J. Platt, and B. Scholkopf,
raw EEG signals. “Support vector machines,” IEEE Intell. Syst. Appl., vol. 13, no. 4,
pp. 18–28, Jul./Aug. 1998.
We have also tried to predict the age of the individuals [4] L. R. M. D. Paiva, A. A. Pereira, M. F. S. D. Almeida, G. L. Cavalheiro,
by using label encoding. The same Deep BLSTM-LSTM and S. T. Milagre, and A. D. O. Andrade, “Analysis of the relationship
LSTM architectures described in the previous section have between EEG signal and aging through linear discriminant analysis
(LDA),” Revista Brasileira de Engenharia Biomédica, vol. 28, no. 2,
been used to experiment. However, only the alpha and beta pp. 155–168, 2012.
waves have been found to be giving better predictions for [5] D. B. Lindsley, “A longitudinal study of the occipital alpha rhythm
the age and gender prediction. It has been found that the in normal children: Frequency and amplitude standards,” Pedagogical
Seminary J. Genetic Psychol., vol. 55, no. 1, pp. 197–213, 1939.
alpha brainwave gives the maximum accuracy of 91.31%. [6] P. Kumar, R. Saini, P. P. Roy, and D. P. Dogra, “A bio-signal based
The results obtained for beta and alpha brain waves on framework to secure mobile devices,” J. Netw. Comput. Appl., vol. 89,
both the LSTM and Deep BLSTM-LSTM models are shown pp. 62–71, Jul. 2017.
in Fig. 10. [7] S. Marcel and J. D. R. Millán, “Person authentication using brainwaves
(EEG) and maximum a posteriori model adaptation,” IEEE Trans.
Pattern Anal. Mach. Intell., vol. 29, no. 4, pp. 743–752, Apr. 2007.
V. C ONCLUSION AND F UTURE W ORK [8] L. Wang, X. Zhang, X. Zhong, and Y. Zhang, “Analysis and classification
of speech imagery EEG for BCI,” Biomed. Signal Process. Control,
Predicting the age and gender of an individual can be vol. 8, no. 6, pp. 901–908, 2013.
highly useful for various applications such as health care, [9] L.-D. Liao et al., “Gaming control using a wearable and wireless
EEG-based brain-computer interface device with novel dry foam-based
knowing consumer preferences for online retailers, biometric, sensors,” J. Neuroeng. Rehabil., vol. 9, no. 1, pp. 1–11, 2012.
etc. As EEG signals are unique for each person they carry a [10] K. Franke, G. Ziegler, S. Klöppel, C. Gaser, and A. D. N. Initiative,
lot of meaningful information. We attempt to build a robust “Estimating the age of healthy subjects from T1-weighted MRI scans
using kernel methods: Exploring the influence of various parameters,”
system for predicting the age and gender of an individual by Neuroimage, vol. 50, no. 3, pp. 883–892, 2010.
analysing the EEG signals. An EEG dataset of 60 individuals [11] M. Tipping, “United states patent and trademark office,”
including male and female was recorded while they sat with U.S. Patent 6 633 857, 2003.
[12] G. Levi and T. Hassner, “Age and gender classification using convo-
their eyes closed in a resting position. As EEG signals are lutional neural networks,” in Proc. IEEE Conf. Comput. Vis. Pattern
time series data where the value at time t + 1 depends on the Recognit. (CVPR), Boston, MA, USA, Jun. 2015, pp. 34–42.

Authorized licensed use limited to: K.L. University(F.K.A. Koneru Lakshmaiah College of Engineering). Downloaded on February 16,2022 at 05:55:01 UTC from IEEE Xplore. Restrictions apply.
 KAUSHIK et al.: EEG-BASED AGE AND GENDER PREDICTION USING DEEP BLSTM-LSTM NETWORK MODEL 2641

[13] W. Samek, A. Binder, S. Lapuschkin, and K. R. Müller, “Understanding [37] D. Joel et al., “Sex beyond the genitalia: The human brain mosaic,”
and comparing deep neural networks for age and gender classification,” Proc. Nat. Acad. Sci. USA, vol. 112, no. 50, pp. 15468–15473, 2015.
in Proc. IEEE Int. Conf. Comput. Vis. Workshops (ICCVW), Oct. 2017, [38] M. Arns et al., “EEG alpha asymmetry as a gender-specific predictor
pp. 1629–1638. of outcome to acute treatment with different antidepressant medications
[14] J. Mansanet, A. Albiol, and R. Paredes, “Local deep neural networks in the randomized iSPOT-D study,” Clin. Neurophysiol., vol. 127, no. 1,
for gender recognition,” Pattern Recognit. Lett., vol. 70, pp. 80–86, pp. 509–519, 2016.
Jan. 2016. [39] Recurrent Neural Network. Accessed: Sep. 18, 2018. [Online].
[15] X. Liu, J. Li, C. Hu, and J.-S. Pan, “Deep convolutional neural networks- Available: https://en.wikipedia.org/wiki/Bidirectional_recurrent_neural_
based age and gender classification with facial images,” in Proc. 1st Int. networks
Conf. Electron. Instrum. Inf. Syst. (EIIS), Jun. 2017, pp. 1–4. [40] V. Khurana, P. Kumar, R. Saini, and P. P. Roy, “EEG based word
[16] C. Szegedy et al., “Going deeper with convolutions,” in Proc. IEEE familiarity using features and frequency bands combination,” Cognit.
Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2015, pp. 1–9. Syst. Res., vol. 49, pp. 33–48, Jun. 2018.
[17] Y. Jia et al., “Caffe: Convolutional architecture for fast feature embed- [41] B. Kaur, D. Singh, and P. P. Roy, “Age and gender classification using
ding,” in Proc. 22nd ACM Int. Conf. Multimedia, 2014, pp. 675–678. brain–computer interface,” Neural Comput. Appl., pp. 1–14, Mar. 2018.
[18] K. Simonyan and A. Zisserman, “Very deep convolutional networks for [42] A. N. V. Ruigrok et al., “A meta-analysis of sex differences in human
large-scale image recognition,” in Proc. Int. Conf. Learn. Represent., brain structure,” Neurosci. Biobehavioral Rev., vol. 39, pp. 34–50,
2015, pp. 1–14. Feb. 2014.
[19] Face Image Project. Accessed: Sep. 17, 2018. [Online]. Available: [43] A. R. Clarke, R. J. Barry, R. McCarthy, and M. Selikowitz “Age
http://www.openu.ac.il/home/hassner/Adience/data.html and sex effects in the EEG: Development of the normal child,” Clin.
[20] M. Soleymani, S. Asghari-Esfeden, Y. Fu, and M. Pantic, “Analysis of Neurophysiol., vol. 112, no. 5, pp. 806–814, 2001.
EEG signals and facial expressions for continuous emotion detection,” [44] A. Gupta, H. Sahu, N. Nanecha, P. Kumar, P. P. Roy, and V. Chang,
IEEE Trans. Affect. Comput., vol. 7, no. 1, pp. 17–28, Jan./Mar. 2016, “Enhancing text using emotion detected from EEG signals,” J. Grid
doi: 10.1109/TAFFC.2015.2436926. Comput., pp. 1–16, Aug. 2018.
[21] P. R. Davidson, R. D. Jones, and M. T. R. Peiris, “Detecting behavioral [45] V. Chang, “Cloud Computing for brain segmentation—A perspective
microsleeps using EEG and LSTM recurrent neural networks,” in Proc. from the technology and evaluations,” Int. J. Big Data Intell., vol. 1,
IEEE Eng. Med. Biol. 27th Annu. Conf., Sep. 200, pp. 5754–5757. no. 4, pp. 192–204, 2014.
[22] Z. Ni, A. C. Yuksel, X. Ni, M. I. Mandel, and L. Xie, “Confused [46] C. Karyotis, F. Doctor, R. Iqbal, A. James, and V. Chang, “A fuzzy
or not confused?: Disentangling brain activity from EEG data using computational model of emotion for cloud based sentiment analysis,”
bidirectional LSTM recurrent neural networks,” in Proc. 8th ACM Int. Inf. Sci., vols. 433–434, pp. 448–463, Apr. 2018.
Conf. Bioinf., Comput. Biol., Health Inform., 2017, pp. 241–246. [47] V. Chang, “An overview, examples, and impacts offered by emerging
[23] M. Golmohammadi, S. Ziyabari, V. Shah, S. L. de Diego, I. Obeid, and services and analytics in cloud computing virtual reality,” Neural Com-
J. Picone. (2017). “Deep architectures for automated seizure detection put. Appl., vol. 29, no. 5, pp. 1243–1256, 2018.
in scalp EEGs.” [Online]. Available: https://arxiv.org/abs/1712.09776 [48] S. Kumar, M. Yadava, and P. P. Roy, “Fusion of EEG response and
[24] M. M. Hasib, T. Nayak, and Y. Huang, “A hierarchical LSTM model sentiment analysis of products review to predict customer satisfaction,”
with attention for modeling EEG non-stationarity for human decision Inf. Fusion, vol. 52, pp. 41–52, Dec. 2019.
prediction,” in Proc. IEEE EMBS Int. Conf. Biomed. Health Inform. [49] S. Sharma, V. Chang, U. S. Tim, J. Wong, and S. Gadia, “Cloud-based
(BHI), Mar. 2018, pp. 104–107. emerging services systems,” Int. J. Inf. Manage., pp. 1–19, Mar. 2016.
[25] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
Comput., vol. 9, no. 8, pp. 1735–1780, 1997.
[26] M. J. A. M. van Putten, S. Olbrich, and M. Arns, “Predicting sex
from brain rhythms with deep learning,” Sci. Rep., vol. 8, no. 1, 2018,
Art. no. 3069. Pallavi Kaushik received the B.Tech. degree from Kamla Nehru Institute of
[27] M. Yadava, P. Kumar, R. Saini, P. P. Roy, and D. P. Dogra, “Analysis of Technology, Sultanpur, in 2015, and the M.Tech. degree from the National
EEG signals and its application to neuromarketing,” Multimedia Tools Institute Technology, Hamirpur. She is currently pursuing the Ph.D. degree
Appl., vol. 76, no. 18, pp. 19087–19111, 2017. with the Department of Computer Science and Engineering, IIT Roorkee,
[28] p. Matoušek and I. Petersén, “Automatic evaluation of EEG background under the supervision of Dr. P. P. Roy. She is mainly interested in the areas
activity by means of age-dependent EEG quotients,” Electroencephalogr. of machine leaning and brain–computer interfaces.
Clin. Neurophysiol., vol. 35, no. 6, pp. 603–612, 1973.
[29] T. Gasser, R. Verleger, P. Bächer, and L. Sroka, “Development of the
EEG of school-age children and adolescents. I. Analysis of band power,”
Electroencephalogr. Clin. Neurophysiol., vol. 69, no. 2, pp. 91–99, 1988.
[30] T. Harmony, E. Marosi, A. E. D. de León, J. Becker, and T. Fernández, Anmol Gupta received the B.Tech. degree from the National Institute of
“Effect of sex, psychosocial disadvantages and biological risk factors Technology, Uttarkhand, in 2014, and the M.Tech. degree from the National
on EEG maturation,” Electroencephalogr. Clin. Neurophysiol., vol. 75, Institute Technology, Hamirpur, in 2016. He is currently pursuing the Ph.D.
no. 6, pp. 482–491, 1990. degree with the Department of Computer Science and Engineering, IIT
[31] A. E. D. De León, T. Harmony, E. Marosi, J. Becker, and A. Alvarez, Roorkee, working under the guidance of Dr. P. P. Roy. He is mainly interested
“Effect of different factors on EEG spectral parameters,” Int. J. Neu- in the areas of machine leaning and brain–computer interfaces.
rosci., vol. 43, nos.1–2, pp. 123–131, 1988.
[32] K. Bennys, G. Rondouin, C. Vergnes, and J. Touchon, “Diagnostic
value of quantitative EEG in Alzheimer’s disease,” Clin. Neurophysiol.,
vol. 31, no. 3, pp. 153–160, 2001.
[33] L. Bonanni, A. Thomas, P. Tiraboschi, B. Perfetti, S. Varanese, and Partha Pratim Roy is currently an Assistant Professor with the Department of
M. Onofrj, “EEG comparisons in early Alzheimer’s disease, dementia Computer Science and Engineering, Indian Institute of Technology, Roorkee,
India. His research interests encompass the areas of pattern recognition, hand-
with Lewy bodies and Parkinson’s disease with dementia patients with a
2-year follow-up,” Brain, A J. Neurol., vol. 131, pp. 690–705, Mar. 2008, written document analysis, multilingual text recognition, biometrics, computer
doi: 10.1093/brain/awm322. vision, image segmentation, brain signal analysis, EEG-based pattern analysis,
machine learning, and temporal data analysis.
[34] J. Dauwels et al., “Slowing and loss of complexity in Alzheimer’s EEG:
Two sides of the same coin?” Int. J. Alzheimer’s Disease, vol. 2011,
pp. 1–10, 2011, doi: 10.4061/2011/539621.
[35] L. Cahill, “Why sex matters for neuroscience,” Nature Rev. Neurosci.,
vol. 7, no. 6, pp. 477–484, 2006. Debi Prosad Dogra is currently an Assistant Professor with the School of
[36] T. Guadalupe et al., “Human subcortical brain asymmetries in Electrical Sciences, Indian Institute of Technology, Bhubaneswar. His research
15,847 people worldwide reveal effects of age and sex,” Brain interests encompass visual surveillance, intelligent transportation systems,
Imag. Behav., vol. 11, no. 5, pp. 1497–1514, 2017, doi: 10.1007/ augmented reality, human–computer interface, and sensor guided healthcare
s11682-016-9629-z. systems.

Authorized licensed use limited to: K.L. University(F.K.A. Koneru Lakshmaiah College of Engineering). Downloaded on February 16,2022 at 05:55:01 UTC from IEEE Xplore. Restrictions apply.