STRESS ASSESSMENT

WITH DEEP LEARNING

INDEX

 What is Stress and Depression

 How is Stress Assessed
 Conventional Clinical Assessment 2

 Modern days Assessment.

 AI based techniques for Stress
Assessment.
 Machine Learning Approaches
 Deep Learning Approaches
 Medical and Wearable Devices
 Deep Learning with EEG
 What is stress?

Chronic stress has been identified as a

factor for neuropsychiatric disorder.
3

When the brain experiences high

demands of activities in certain
external or internal pressure, which
can be psychological or social.
 What is stress?

Brain tries to utilize more resources

which may lead to initial abnormal
activities. 4

Continuous utilization of resources

leads to initial abnormal activities
and functions by other organs.
What is depression

Major Depressive disorder

(MDD) is commonly termed as
depression and, it is a mental illness. 5

Psychotic depression is more severe

form of depression and involves more
factors (causes) than non-psychotic
depression.
 Effects of
stress and
depression
Affect the immune system, making it
harder for the body to fight
infection. Vaccinations are even less 6

effective
Heart disease and increased
the risk for substance
abuse.
Insomnia or hypersomnia nearly
every day Significant change in
appetite or weight
 Treatment

Psychological treatments for

depression 7
🢝 Cognitive behavior therapy (CBT)
🢝 Interpersonal therapy (IPT)
🢝 Behavior therapy
🢝 Mindfulness-based cognitive therapy
(MBCT)
Medical treatments for
depression
🢝 Antidepressant
Neuro-Feedback based treatment
How stress is
accessed
Conventional Clinical
Assessment 8

Modern days
Assessment
 Conventional clinical
assessment

These assessment techniques are

mostly questionnaire based. Following
are the most popular techniques and 9

approaches to assess stress in a

person.
▹ Perceived Stress Scale (PSS)
▹ Stress Response Inventory (SRI)
▹ Life Events and Coping Inventory
(LECI)
1
0
1
1
 Modern assessments

Modern day approaches involves less

human involvement with better
generalization ability. These systems 1
2
which help in diagnosis are called
CADs (Computer Aided Diagnosis
Systems).

CAD systems uses data recorded

from modern day medical equipment.
Modern
assessments
The different modalities aids the
evaluator in
1
making decisions regarding the 3

diagnosis.
These techniques provide in-depth
insight to the physicians.
Following are different modalities
being
▹ MRI used ▹ EEG
for stress
▹ fMRI assessment.
▹ MEG
▹ PET ▹ fNIR
▹ MEG
What is EEG?

The modalities discussed can be

categorized in two different classes or
approaches, invasive and non-invasive
nature of these modalities. 1
4

Invasive techniques are those which can

or may be harmful to the patient in
short or long period on time.

Non-Invasive are those which does not

have any harmful effect of the patient.

EEG is one of the Non-Invasive

techniques with high ability to record
temporal information of the brain.
EEG
(ELECTROENCEPHALOGRAPHY)

1
5
 Challenges

The treatment management of stress and depression similar

disorders is a big
challenge. 1
6
The diagnosis is such disorders are currently not efficient enough.
After prescription of an antidepressant, the psychiatrist or
clinician reassesses after 2 to 4 weeks.
• If the treatment is ineffective
• Precious time is wasted
• High cost incurred
▹ Risk of treatment withdrawal
▹ Potential risk of harming himself or others
Requirement

• A scientific system that can

• Predict the treatment outcome for
depression.
1
7

• Select the type of medication with high

level of
confidence about the efficacy.
ITMS
Patien
t
Histor
y
Subjectiv
e
Experienc
e
EEG

1
8
C OMPARISON WI T H CONVENTIONAL PR
OC(IETDM
URSE
) Time wasted if Medicine not
effective
Conventional Trial
Medication for MDD
4 Weeks waiting
Time

ITSS
Medication Selection No need to wait
with Scientific Promises Patient
Evidence Commitment for
treatment

WEEK 0 WEEK WEEK WEEK

1 2 3
Stress assessment
with EEG

Recording Stress in real-life routine is

very difficult. In order to observe
similar effects, experiments are 2
0
designed for participants.
These experiments involves tasks
for the participant and these tasks
acts as a stimuli for the brain to
trigger stress.

EEG recordings are observed in

normal (relax) state and stressed
state of the participant.
 Artifacts

EEG artifact

Raw EEG signals are

contaminated with artifacts.
These artifacts contains eye
blinks, muscle movements etc.
2
2

In order to use machine

learning or any AI technique,
dataset needs to be cleaned
and pre-processed.

In case of EEG for stress

assessment we can filter out
the required frequency band
and eliminate artifacts either
with signals processing or
AAS Alpha Asymmetry

AAS has been associated with

negative emotions and often
linked with stress conditions in brain. 2
5

Alpha waves in left−prefrontal

are responsible for good memories
and thoughts. While alpha brain-
activity in right prefrontal is
responsible for bad memories

AAS is used to label each short

window for training.
Stress Assessment with EEG

Contr Stresse Non-

ol d Stressed 2
6
Medical
application and
AI
Still with modern medical
equipment, the process of diagnosis
remained dependent of expert 2
physicians. 7

With lack of physician all around the

world the requirement is impossible to
meet without the aid of expert
computer systems for assistance.

AI plays a pivotal role in reducing the

burden off the physicians and reducing
the error rate of wrongful diagnosis by
the physicians.
 Medical
application and
AI
Currently the AI techniques can be
divided
into two major portions. 2
8

▹ Machine Learning
▹ Deep Learning
For our problem of Stress Assessment
both of above are applicable and
provide better classification ability
between stress and non- stress
signals.
Medical
application and
AI
The foundation of ML (Machine Learning)
lies in the pre-processing and features
which are being extracted and then 2
selected from the available data. 9

Garbage IN Garbage OUT.

For Machine learning model to

perform efficiently it is crucial that data
is pre-processed properly. Most
appropriate and relevant features
are extracted and selected after
evaluation which can discriminate
data in required classes.
 Medical
application and
AI
For stress assessment convention
the features which al
are extracted involves 3
0
following: the
▹ Power ▹ Relative Energy
▹ Relative Power ▹ Relative Energy
▹ Relative Power Ratio
Ratio ▹ Coherence
▹ Energy ▹ Asymmetry
 Existing online
Assessment
All the assessment and analysis
perform using the previous
techniques i.e. clinical questionnaire
assessment, using different modalities
for data acquisition for machine 3
1
learning classification and prediction
restrict us to only offline assessment.
Whereas, the requirement of such
medical problems are only possible in
real-time procedure. The previously
discussed techniques are not suitable
for real-time application that involves
depression and stress like mental
disorders.
Need of online Assessment

In order to fulfill the requirement of such

mental disorders i.e. workplace mental
load, epilepsy, depression, and PTSD 3
etc. which needs online assessment which 2
response in real-time.
Process involve that are time expensive,
needs to be eliminated or reduced.

We need models which does not

rely of conventional feature
extraction techniques, process data in
real-time, generalize to eliminate
requirement of self-tuning and time
efficient.
Medical Application and
DL
Deep learning is continuously turning
tables for Machine Learning and
Medical Application.
DL provides better accuracies as compared
to ML techniques without the hectic process
3
3
of feature engineering and selection.

Its ability to extract features in unsupervised

manner provide the most efficient feature
maps for classification.

The generalization ability of DL approaches

makes them suitable for standardize stress
assessment technique.
EEG Representational
Learning

Machine and Deep Learning techniques are

being applied on EEG with various different
representation. 3
5

The most popular representation learning

found while using EEG signals data is the use
of Spectrograms or using filtered signals.

But both the approaches of representational

learning are depending merely on the
threshold of the signals, in other words
amplitudes.
Self-Tuning
 MEDICAL AND
WEARABLE DEVICES.

Nowadays, when everyone has the access to

smart phones and devices. It is now possible
to track body health condition anytime in 3
real-time. 6

Blood pressure, heart beat, sugar level,

oxygen level etc. are all available in wearable
devices, trackable of smart gadgets and
can be analyzed by using ML and DL
approaches for health care application in
real-time.

Hence, Stress assessment is also important

to be assessed in real-time with efficiency.
 MEDICAL AND
WEARABLE DEVICES.
For the purpose of real-time stress
assessment, machine learning techniques
may not be appropriate to use. As these
approaches use extensive pre-processing and
feature extraction. In this way the total time 3
expense will be much higher with respect to 7

the time needed to spend for a real-time

application.

This is why DL techniques provides better

efficiency in accuracy as well as with regards
to time consumption for stress assessment,
making it the most appropriate approach for
real-time stress assessment.
 DEEP LEARNING

• Supervise
• Unsupervis
d Feature
ed Feature
Extraction
Extraction
and
and
Selection
Selection. • 3
• Time 8
Time
Expensiv
Efficient
e
 DEEP LEARNING
Popular Deep learning algorithms are
CNN (Convolutional Neural Networks) and
LSTM (Long Short-Term Memory).

CNN are mostly used with image data

and are best known for extracting
spatial information. CNNs can also be LSTM 3

used with signal data in both formats cell

9

i.e. Images (Spectrograms) and Raw

Signals.

LSTMs are advance form of RNN

(Recurrent Neural Networks). LSTM units
are popular in application involving
temporal information i.e. speech
recognition and natural language
processing.
Convolution
EEG FREQUENCY BANDS

Each EEG signal for stress can be

divided into
certain specified frequency bands. 4
0

These bands are as

follows:
Alpha (8-12 Hz)
Beta (12-32 Hz)
Theta (4-7 Hz)
Gamma (30-100 Hz)
Delta (0-4 Hz)
 Unfiltered

EEG BANDPASS
FILTER
In order to remove the underlining noise
present in the signals, we can use a band-
pass filter.
4
1
Band-pass filter will eliminate those
frequencies which can be irrelevant to the
process of classification or are not under
examination.

Filtered
DATA SPLITTING

In order to make a better generalized

model which can learn patterns
efficiently, data should be split into 4
2
training, testing and validation
datasets, so that the model can
learn patterns and evaluated
properly.
 INPUT DIMENSION

In case of stress assessment using EEG

signals, we intend to develop a model which
can accept signals in real-time and classify
them into respect classes in prompt
response. 4
3
It is crucial to have a small size signal
window so that there should be a
synchronous procedure of data acquisition,
buffering and input to DL model, and this
whole process is needed to be fast as it can
be to support the concept of real-time
assessment.
 Input
Layer

DEEP LEARNING
ARCHITECTURE (CNN)

CNN Neur Networ

(Convolutional of alvarious k)
architecture separatepurpose Kerne Featur
comprises layers, important e
each withits and l
44
operation.
Generally, the architectures Convolution
contains the
following layers.
▹ Input layer ▹ Activation
▹ Batch function
normalizati ▹ Fully-
on layer Connected
▹
Convolution Layer
Poolin
al layer ▹ Softmax
Pooling g
▹ Layer