Machine Learning in ADHD and Depression Mental Health Diagnosis A Survey

Received 11 July 2023, accepted 31 July 2023, date of publication 10 August 2023, date of current version 17 August 2023.
Digital Object Identifier 10.1109/ACCESS.2023.3304236
Machine Learning in ADHD and Depression

Mental Health Diagnosis: A Survey
CHRISTIAN NASH 1 , (Student Member, IEEE), RAJESH NAIR2 ,
AND SYED MOHSEN NAQVI 1 , (Senior Member, IEEE)
1 IntelligentSensing and Communications Group, School of Engineering, Newcastle University, NE1 7RU Newcastle upon Tyne, U.K.
2 Adult ADHD Services, Cumbria, Northumberland, Tyne and Wear, NHS Foundation Trust, NE3 3XT Newcastle upon Tyne, U.K.
Corresponding authors: Christian Nash ([email protected]) and Syed Mohsen Naqvi ([email protected])
ABSTRACT This paper explores the current machine learning based methods used to identify Attention
Deficit Hyperactivity Disorder (ADHD) and depression in humans. Prevalence of mental ADHD and
depression is increasing worldwide, partly due to the devastating impact of the COVID-19 pandemic for
the latter but also because of the increasing demand placed on the mental health services. It is known that
depression is the most common mental health condition, affecting an estimated 19.7% of people aged over
16. ADHD is also a very prevalent mental health condition, affecting approximately 7.2% of all age groups,
with this being conceived as a conservative estimate. We explore the use of machine learning to identify
ADHD and depression using different wearable and non-wearable sensors/modalities for training and testing.
These modalities include functional Magnetic Resonance Imagery (fMRI), Electroencephalography (EEG),
Medical Notes, Video and Speech. With mental health awareness on the rise, it is necessary to survey the
existing literature on ADHD and depression for a machine learning based reliable Artificial Intelligence
(AI). With access to in-person clinics limited and a paradigm shift to remote consultations, there is a need
for AI-based technology to support the healthcare bodies, particularly in developed countries.
INDEX TERMS Artificial intelligence, attention deficit hyperactivity disorder, depression, machine learn-
ing, mental health.
I. INTRODUCTION due to overlapping symptoms and the potential side effects of

There are a multitude of mental health conditions that can ADHD medications. Saying this, differences exist in mood,
affect individuals, with various explanations accounting for motivation and sleep patterns between the two conditions [2],
their occurrence. There is no single definitive answer that has [3]. Both ADHD and depression are very broad topics,
been identified. Conditions like depression and schizophrenia so to specialise our paper we focus only on wearable/non-
have been associated with hereditary factors and chemical wearable sensing and machine learning. Due to the link in
imbalances in the human body [1]. However, this research symptoms, if one machine learning model can accurately
mainly focuses on ADHD and depression, the two most detect one of the disorders, there is a chance that the model
prevalent mental disorders in humans. Both conditions often can be generalised in identifying the other. These connec-
co-occur, with people diagnosed with one being more likely tions and the high prevalence rate is what motivated this
to be diagnosed with the other. In fact, adults with ADHD paper.
are three times more likely to have depression, and individ- ADHD is a global concern affecting both children and
uals with depression have a 30-40% prevalence of ADHD. adults. A 2015 meta-study found the worldwide prevalence of
There are also links between ADHD and increased suicidal ADHD among children aged 18 and under to be 7.2% with a
ideation. Distinguishing between the two can be challenging 95% confidence level [4]. Notably, cases of persistent ADHD,
where symptoms that begin in childhood continue into adult-
The associate editor coordinating the review of this manuscript and hood, have a lower prevalence of 2.58% [5]. This discrepancy
approving it for publication was Chih-Yu Hsu . is believed to stem from limited access to diagnosis during
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 11, 2023 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 86297
C. Nash et al.: Machine Learning in ADHD and Depression Mental Health Diagnosis: A Survey
youth, suggesting that the real prevalence of ADHD in adults process, a concern that further fuels the motivation for this
could be higher. work. Further complicating matters is the fact that 70-75% of
ADHD diagnosis is influenced by gender, with a male-to- people with diagnosable mental illnesses do not receive any
female ratio of 2.28:1 observed in a sample of 858 ADHD treatment [17], [18]. The consequences of this treatment gap
diagnosed participants [6]. It’s worth noting, however, that are significant, as evident in workplace-related mental health
this ratio varies across different studies and regions, with issues. From 2018 to 2019, stress and depression accounted
a consistent trend of higher prevalence in males. An inves- for 44% of all work-related illness cases. Furthermore, it is
tigation into underdiagnosis in London found an undiag- estimated that up to 55% of all lost working days were
nosed ADHD rate of approximately 12% among 226 partici- due to mental health conditions [19]. These lost workdays
pants [7]. This underdiagnosis is often due to symptoms being bear a substantial economic cost, estimated between £74-99
misinterpreted as simple misbehaviour by parents and teach- billion [20].
ers. The British Broadcasting Company (BBC) suggests that The situation took a turn for the worse with the advent
the issue of undiagnosed ADHD is widespread. It estimates of the COVID-19 pandemic in 2020. A UK government
that around 1.5 million adults in the UK have ADHD, but report using the General Health Questionnaire 12 (GHQ-
only 120,000 are officially diagnosed [8]. In addition, those 12) measure showed that average mental distress in April
seeking a diagnosis may face substantial wait times of up to 2020 rose by 8.1% compared to the 2017 to 2019 aver-
seven years [8]. age [21]. A study involving 1,300 healthcare providers and
The ramifications of ADHD extend beyond the individuals 6,200 non-healthcare providers showed that caregivers exhib-
directly affected. It impacts families, with studies suggesting ited higher rates of depression, likely due to the harsh impact
that an ADHD diagnosis can lead to higher divorce rates [9]. of the virus on UK care homes [22]. The pandemic’s toll was
A longitudinal study by the University of Pittsburgh recorded also felt in mental health services. Data from South London
a 22.7% divorce rate among families with ADHD, compared services revealed that between March and June 2020, there
to 12.6% in non-ADHD families [10]. Moreover, ADHD were 1,109 additional deaths among their patients compared
carries significant economic implications. In the US, the to previous years, with 64% of these fatalities attributed to
annual cost of ADHD was estimated to range from 143−266 COVID-19 [23]. Moreover, studies suggest that adults with
billion, with productivity-related adult income losses being mental health conditions were more likely to be hospitalised
the primary cost factor, accounting for 87−138 billion [11]. and even succumb to COVID-19 [24]. Given the gravity of
Additionally, a meta-analysis revealed a strong link between these findings, the development and deployment of viable AI
ADHD and criminal behavior, with individuals diagnosed solutions for mental health detection are more urgent than
with ADHD in childhood being two to three times more likely ever.
to be arrested, convicted, or incarcerated as adults [12]. There’s a recognized strong interrelation between ADHD
The process of diagnosing ADHD can be intricate and and depression, although the underlying causes remain elu-
lengthy, involving comprehensive history collection of an sive. Some theories propose that adults with ADHD are at an
individual’s behaviour across home and school environ- increased risk of experiencing adverse life events, which may
ments [13]. However, several challenges can limit the effec- contribute to the relationship between these two conditions.
tiveness of this process, including variability in the subjective This hypothesis was tested in a study of 230 adults diag-
judgments made by assessors, inaccuracy or incompleteness nosed with ADHD [25]. The data was processed using linear
of assessment questionnaires, and cultural considerations in and logistic regression models, which revealed that individ-
the standardization of ADHD tests such as Conners-3 [14]. uals who had experienced adverse life events had a higher
With these limitations, researchers are exploring alternative tendency towards depression. Further research supports the
methods for diagnosis, including machine learning. Current suggestion that an ADHD diagnosis may predispose individ-
techniques in ADHD diagnosis research often involve analyz- uals to develop depression in later life. This notion is backed
ing an individual’s brain activity during specific tasks using by a longitudinal study that examined the data of 8310 chil-
fMRI and EEG. Such objective measurements, compared dren with ADHD and found an increased risk of recurrent
between ADHD patients and healthy controls, could offer depression in young adulthood. Furthermore, Mendelian ran-
significant insights. Moreover, further exploration of longi- domization (MR) analyses have indicated a possible causal
tudinal studies could provide valuable knowledge about the effect of ADHD genetic liability on major depression later in
cause and progression of ADHD. life [26]. The findings from these studies underscore the com-
Mental illnesses, with depression being the most prevalent, plex interplay between ADHD and depression, suggesting an
significantly impact the lives of those affected. As of 2014, urgent need for more focused research in this area.
it was estimated that nearly 19.7% of individuals aged 16 and Existing work in ADHD and depression analysis using
above experienced symptoms of depression [15]. However, machine learning methods has so far exploited either
the World Health Organization reported that the diagnosed non-wearable data or wearable data. The most popular data
depression rates in the UK were a mere 4.5% as of 2015 [16]. sources to analyse for recognition of both ADHD and
This discrepancy suggests potential issues with the diagnostic Depression are EEG signal data and MRI imaging data.
86298 VOLUME 11, 2023

A vast range of machine learning methods have also been nose ADHD and depression, respectively. Lastly, conclusions
employed. Saying this, the most popular classification tech- and future direction has been discussed in Section VII.
niques are Support Vector Machines (SVM) and Neural
Networks. II. LITERATURE SELECTION CRITERIA
SVMs are a powerful supervised machine learning model Before the paper compilation, research questions were pro-
primarily used for classification or regression tasks. SVMs posed to allow for concise conclusions and efficient searches.
work by identifying an optimal hyperplane that maximally
separates different classes of data in a multi-dimensional A. RESEARCH QUESTIONS (RQs)
space, effectively finding the decision boundary that has the The following research questions were finalized to focus the
largest margin between classes. Advantages of SVMs include scope of the survey:
their robustness in high-dimensional spaces, effectiveness • What wearable and non-wearable sensing have been
when the number of dimensions exceeds the number of used in datasets for mental health ML-based ADHD and
samples, and flexibility through the use of different kernel depression detection research?
functions to capture complex decision boundaries. However, • What are the advantages and disadvantages with indi-
SVMs can be computationally intensive, especially for large vidual modalities when trying to diagnose ADHD or
datasets, they’re less effective with noisy data where classes depression using machine learning?
overlap, and they require proper tuning and selection of the • What is the most popular classification algorithm
kernel function and regularization parameter to perform opti- applied?
mally. The lack of a probabilistic interpretation of the results • What is the standard of the classification method used?
could also be seen as a disadvantage. Neural networks are a
class of machine learning models inspired by the biological B. SEARCH STRATEGY
structure of the brain. They consist of interconnected layers Compiling the papers was achieved through a keyword
of nodes or ‘‘neurons’’ that can learn to represent and manip- string query over several literature databases. The keywords
ulate data. Neural networks are particularly advantageous in the string were chosen to produce results that fit the
for their capacity to learn complex, non-linear relationships RQs. The keyword string query is as follows: (‘‘Classifica-
directly from raw data, making them useful for tasks like tion’’ OR ‘‘Neural Networks’’ OR ‘‘Machine Learning’’ OR
image recognition, natural language processing, and more. ‘‘Deep Learning’’ OR ‘‘Supervised Learning’’ OR ‘‘Unsu-
They can handle high-dimensional data and are highly scal- pervised Learning’’) AND (‘‘Depression’’ OR ‘‘ADHD’’)
able. However, neural networks also have some notable AND (‘‘Diagnosis’’). The following query was used on the
disadvantages. They require large amounts of labelled data following literature databases: IEEE Xplore, Science Direct,
for training, and they are often computationally expensive, PubMed and Web of Science.
both in terms of memory and processing power. The training
process can also be challenging due to issues like overfitting, C. CRITERIA FOR IDENTIFICATION OF STUDIES
vanishing or exploding gradients. Lastly, the ‘‘black box’’
The inclusion criteria:
nature of neural networks can make the interpretation of their
• Publication in English.
internal workings and decision processes difficult, posing
• Inclusion of data containing an individual with a formal
challenges for transparency and trust.
Part of this work has been published at the International diagnosis of depression or ADHD.
• Articles involving the diagnosis of a mental health con-
Conference on Information Fusion 2022 [27]. This is the
complete version of the survey, extensively covering the vast dition by using Machine Learning.
• Investigating the diagnosis of ADHD or depression in
majority of work completed in the area with broad explana-
tions of engineering and medical techniques. Not all literature humans using Machine Learning.
• Publication in a peer-reviewed journal.
can be included due to page limitations. The authors would
• Publication within the last 11 years (2011-2022).
like to acknowledge the existence of surveys into detecting
Mental Health using Machine Learning [28], [29], [30], [31], The exclusion criteria:
[32]. Saying this, they are different to this survey paper in • Publication in a non-peer-reviewed journal.
several ways, with the absolute focus of our paper being • Publication in conference proceedings, book chapters
ADHD and depression. and dissertations.
The rest of this paper is organised as follows. In Section II,
the selection of literature is provided, with the parameters III. TESTING FOR A MENTAL HEALTH CONDITION
for acceptance being discussed. In Section III, testing for The DSM, currently in its fifth edition, serves as a widely uti-
mental health conditions is presented for both ADHD and lized handbook for clinicians and psychiatrists in the United
depression. Section IV provides insight into the publicly States [13]. The DSM-V encompasses the majority of men-
available datasets that are used in some of the studies analysed tal health disorders and undergoes continuous professional
throughout this survey. Sections V and VI discuss the existing revision. Over four hundred experts from thirteen different
literature where machine learning has been exploited to diag- countries contributed to its development, representing fields
VOLUME 11, 2023 86299

such as epidemiology, neurology, paediatrics, primary care,

psychiatry, psychology, and research methodology.
The DSM-V includes descriptions, symptoms, and other
relevant criteria for specific mental health disorders to aid
in diagnosis. Moreover, it provides diagnostic criteria for
both children and adults. As a result, the majority of studies
referenced in this paper employ the DSM-V to accurately
identify individuals with ADHD or depression.
The International Classification of Diseases (ICD-11) was
created by the World Health Organisation at a similar time
to the DSM-V [33]. Similarly, it provides a broad range of
knowledge on the extent, causes and consequences of human
diseases (both medical and mental). The ICD-11 allows for FIGURE 1. An example of the SST [107]. It can be done on a laptop or
tablet, with the latter being recommended. The SST task takes roughly
systematic recording, interpretation and therefore analysis of 15 minutes to complete. It is normal for both controls and ADHD subjects
mortality and morbidity data that is collected globally. to be wrong 50% of the time.
As both the DSM-V and ICD-11 are very similar in nature,

there is a push to harmonise both together. To make this
happen, in new iterations, the main focus will be to have 2) Making a decision in the short term without con-
the greatest clinical impact. Achieving this means increasing sidering the effects of the long term.
their international uniformity, with the enhancement of cul- 3) Difficulty with self control.
tural compatibility being the primary goal.
1) QUESTIONNAIRES
A. ADHD Numerous questionnaires, including Conners-3 [14], can aid
Regarding ADHD specifically, the DSM-V asserts that to be in the diagnosis of ADHD. These questionnaires can be
diagnosed with Attention Deficit Disorder (ADD), an indi- completed by clinicians, patients, primary caregivers, or sec-
vidual must exhibit five or more symptoms of inattention ondary caregivers. When completed by someone other than
persisting for over six months. Additionally, to be diagnosed the clinician, the questionnaire offers valuable insight into
with ADHD, five or more symptoms of both hyperactivity the individual’s behavioural history. However, the subjective
and impulsivity, along with inattention symptoms, must be nature of the responses may lead to inaccuracies and false
present for more than six months. positives in individuals pursuing a diagnosis, even if they do
Symptoms are classified into three major components: not genuinely have ADHD.
Inattention, Hyperactivity and Impulsivity [13], [34]:
• Inattention: 2) STOP SIGNAL TASK (SST)
1) Forgetfulness in daily tasks/work. Fig. 1 illustrates the screen a participant would encounter
2) Making careless mistakes in work or tasks. when taking the SST. This test represents a unique version
3) Difficulty sustaining attention in tasks. of a classic method for measuring response inhibition (i.e.,
4) Fails to complete tasks. impulse control). Participants respond to an arrow stimulus by
5) Doesn’t listen when spoken directly to. selecting one of two options based on the arrow’s direction.
6) Reluctance in joining tasks that require sustained The test comprises two parts:
attention. • First, the participant is introduced to the test and
7) Often loses things necessary for tasks. instructed to press the left-hand button when they see a
8) Easily distracted by external stimuli. left-pointing arrow and the right-hand button when they
9) Often forgetful in daily tasks. see a right-pointing arrow. The participant practices this
• Hyperactivity: task in 16 trials.
1) An individual constantly moving around, even dur- • Next, the participant is asked to continue selecting but-
ing inappropriate times such as in a cinema. tons corresponding to the arrow directions. However,
2) Fidgeting with their hands excessively. if an auditory signal (such as a beep) occurs, they should
3) Tapping surfaces with their fingers or tapping their refrain from responding and not press the button.
feet on the ground.
4) Excessive talking. 3) CONTINUOUS PERFORMANCE TEST (CPT)
5) Difficulty engaging in leisure activities quietly. The CPT is a task-oriented, computerized assessment, that
6) Always on the go. evaluates attention-related issues in individuals aged 8 years
• Impulsivity: and older. It measures the participant’s performance in areas
1) Interrupting conversations or answering before the such as attentiveness, impulsivity, sustained attention, and
question has been asked in full. vigilance. The CPT supplements information obtained from
86300 VOLUME 11, 2023

rating scales like Conners-3 [14], offering insights into an

individual’s performance in attention tasks.
B. DEPRESSION
The DSM-V presents depression as persistent feelings of
sadness and hopelessness while showing lack of interest in
activities that were once enjoyed [13]. It mentions that indi-
viduals could experience additional physical symptoms such
as chronic pain or digestive issues. The DSM-V states that the
subject must be experiencing five or more of the following
symptoms during the same 2 week period:
1) Depressed mood most of the day, experienced nearly
every day. FIGURE 2. A typical MRI Scanner [108]. MRI scanners are expensive
2) Noticeable diminished interest or pleasure in all (or pieces of equipment so they are constantly in use at hospitals for
multiple needs. The bed moves in and out of the main scanner depending
almost all) activities. on what area of the human body is being imaged.
3) Experiencing significant weight loss or gain with a
decrease or increase in appetite.
4) A reduction of physical movement and thoughts slow-
each respective question has a list of four statements that are
ing down.
arranged in increasing severity. Each question is focused on a
5) Fatigue or loss of energy nearly every day.
particular symptom of depression. Its second revision aligned
6) Feelings of worthlessness nearly every day.
its questions with the DSM-IV criteria by having the answers
7) Diminished ability to think or concentrate nearly every
focused on the last 2 weeks upon taking the test.
day.
8) Recurrent thoughts of death.
C. MAGNETIC RESONANCE IMAGING (MRI)
At least one of the symptoms should be either a depressed Magnetic Resonance Imaging (MRI) is a type of scan that
mood or loss of interest or please. It should be noted that to uses powerful magnetic fields and radio waves to provide
receive a diagnosis of depression, the symptoms must cause highly detailed images of the inside of the body. A scan can
the subject clinically significant distress and impairment to last between fifteen and ninety minutes, depending on the size
everyday life. of the area being scanned. The main advantage of MRI scan-
ners is that they are harmless to the subject. A main downside
1) PATIENT HEALTH QUESTIONNAIRE-9 (PHQ-9) to them is that they can be claustrophobic. There are several
The PHQ-9 is a self-administered diagnostic tool used types of MRI scanner measurements with the focus of this
for criteria-based diagnosis of depression, as established survey being functional MRI (f-MRI) and resting-state MRI
by Kroenke, Spitzer and Williams [35]. The initial study (rs-MRI). Initially developed to showcase regional/localized,
involved 6,000 patients from various clinics. Criterion valid- time-varying changes in brain metabolism, f-MRI scanners
ity, which is predictive of outcomes, and construct validity, have gained popularity due to their versatility with invasive
which assesses how well a test measures its intended subject, and non-invasive techniques, good spatial resolution, and rel-
were determined against a mental health professional-led atively low cost [38]. In the context of the studies mentioned
interview and the 20-item Short-Form General Health Survey throughout this paper, f-MRI’s primary use is to observe
respectively [36]. increased neural activity by having a subject perform a task
The PHQ-9, comprising of only nine questions, is based while in an MRI scanner, as shown in Fig. 2.
on the actual nine criteria for DSM-V depressive disorders Two techniques exist for tracking neural activity: invasive
diagnosis and can also indicate depressive symptom severity. and non-invasive sensing. Invasive sensing involves injecting
Subjects respond based on their feelings and thoughts over the participant with a contrast agent to identify increased
the past 2 weeks as presented in Table 1. Clinicians interpret local cerebral blood flow and changes in oxygen concen-
answers and scores to determine the presence and severity tration. In contrast, non-invasive sensing employs Arterial
of depression (Table 2). Four or more ticks in the bold area Spin Labeling (ASL), which suffers from reduced sensitivity,
suggest a depressive disorder. increased acquisition time, and higher sensitivity to body
motion.
2) BECK DEPRESSION INVENTORY (BDI-II) Typically, the subject in the f-MRI performs a task using
The BDI-II is in its second iteration and is one of the most visual, auditory, or other stimuli to induce cognitive states.
widely used instruments for detecting depression [37]. It is A two-condition design is commonly used to identify acti-
similar to the PHQ-9 with respect to it being a self-report vation. In the studies discussed in this paper, condition one
questionnaire with it being designed to measure the severity is a resting state where the individual is instructed to sleep,
of a subject’s depression. It consists of 21 questions where providing an unstressed brain activity baseline known as
VOLUME 11, 2023 86301

TABLE 1. The 9 questions a subject will answer to determine whether they suffer from depression and the level of depression [35].
TABLE 2. Interpretation of the total scores when evaluating depression

severity using the PHQ-9 [35].
FIGURE 4. An advanced wearable EEG head cap [109]. The headcaps

normally have a minimum of 8 electrodes with some advanced caps
containing 256 electrodes. 32 electrodes have been found to be optimal
for research purposes.
FIGURE 3. fMRI Scan Output from the ADHD-200 Dataset [39]. trical impulses, and when thousands of neurons fire in syn-
chrony (due to localized brain area activation), an electrical
field strong enough to penetrate tissue and the skull is pro-
rs-MRI. Condition two involves the subject undertaking the duced. This enables humans to measure brain activity using
designated activity. The images are combined to create a specialized equipment, as shown in Fig. 4. The cap contains
contrast map, and through image processing, an activation multiple electrodes that touch the individual’s scalp to record
map is generated. Fig. 3 displays the output MRI image for a brain activity, ensuring identical data collection between par-
participant’s resting state in the ADHD-200 dataset [39]. ticipants as the electrodes remain in fixed positions.
EEG offers a significant advantage as a non-invasive
D. ELECTROENCEPHALOGRAPHY (EEG) technique utilizing wearable sensors, providing highly accu-
The brain consists of densely packed neurons interconnected rate time-resolution data. This precise resolution enables
through synapses, which serve as gateways for inhibitory or researchers to detect activity within cortical areas on a sub-
excitatory activity. Synaptic activity generates subtle elec- second timescale. As the voltage fluctuations measured at the
86302 VOLUME 11, 2023

TABLE 3. Different frequency bands and their purposes in an EEG a lavalier microphone, is attached to the participant being
machine.
recorded. While a secondary microphone, or microphone
array is placed in the room to record the environment noises.
To achieve the best quality audio recording, it is recom-
mended to record in the highest sampling rate that the chosen
microphone has to offer. Pair the highest sampling rate with
a 24-bit rate to increase the quality of the recordings while
increasing the level of detail. When designing video data
recording, it is vital to consider the stability of the cameras.
Correct tripods and mounts are essential as you don’t want
any additional blurriness or motion captured. Depending on
what is being captured, the resolution of the camera and frame
rate can differ due to there being a trade-off between res-
olution and performance with machine learning algorithms.
Provided the camera is stable, recording at a resolution of
1920 × 1080 (High-Definition) at a frame-per-second (fps)
of 30 is suitable. If a budget allows for it, there is also the
option to record in 4K (4096 × 2160) at 30 fps.
IV. DATASETS
There are not a lot of publicly available datasets for Mental
Health challenges due to the highly sensitive nature of the
data. The main concern is the protection of the participant’s
privacy i.e, identity and health information. Therefore in
some cases, it is safer to not release the data publicly. In the
cases where data has been made public, mainly the video
modality, it is processed into features that can not be reverse
transformed into their original format. This pre-processing of
the data can impact the algorithms used to experiment with
such data.
It is known that a common problem with datasets involving
medical information are usually small in size. This is due to
complications with preserving the participant’s identity while
FIGURE 5. Example of the raw EEG signal with the extracted common
also facing challenges in finding enough individuals with
frequency bands [110]. the condition being researched. If a researcher has access to
the original video, the small dataset size could be increased
using data augmentation techniques. The choice is with the
electrodes are small, the data is amplified to be displayed as a researcher but as a few examples, blurring can be applied to
sequence of voltage values. Moreover, since the electrodes the video or frame mirroring could be applied [40].
analyze brain activity across various regions, data can be
examined in specific brain cortices. Different cortices within A. ADHD
the brain are responsible for distinct activities, allowing cer- 1) ADHD-200
tain tasks to isolate particular brain areas. The EEG identifies The ADHD-200 Machine Learning competition invited the
specific frequency bands, as illustrated in Table 3: neuroimaging and data mining communities to develop a
Fig. 5 is showing the output of wearing the EEG hat. It is pattern classification method that could distinguish brain
showing the outputs in the form of the main frequency bands activity differences between a control and an individual
highlighted in Table 3. with ADHD [39]. The dataset comprises a combination of
structural MRIs (s-MRIs) and resting state functional MRIs
E. AUDIO/VISUAL (rs-fMRI). The ADHD-200 dataset features pre-processed rs-
Before recording a dataset that will be evaluated using fMRI data from 973 participants, categorized as Typically
machine learning, consideration of the equipment is vital. Developing (TD), ADHD-Impulsive (ADHD-I), and ADHD-
For instance, speech signals can be severely corrupted by Combined (ADHD-C). Table 4 displays the breakdown of the
background noise, depending on the recording environment. sample population. The training set released for the developed
Therefore, it is recommended that a dual microphone con- models contained 776 participants data. For testing the mod-
figuration is implemented. Ideally, a microphone, such as els, a further 197 data entries were released.
VOLUME 11, 2023 86303

TABLE 4. The complete breakdown of training and testing data combined TABLE 5. Showing the baseline Mean Absolute Error (MAE) and Root
for the ADHD-200 dataset. Where PU is Peking University, UPitt is the Mean Square Error (RMSE) for the AVEC2013 [42] and AVEC2014 [43]
University of Pittsburgh, NYU is the New York University Child Study challenges.
Center, BrownU is the Bradley Hospital at Brown University, NI is
NeuroIMAGE, OHSU is the Oregon Health and Science University, KKI is
the Kennedy Krieger Institute and WashU is Washington University in St.
Louis. There are a total of 973 participants in the dataset.
for depression evaluate their machine learning models on

the AVEC2013 and AVEC2014 datasets [42], [43]. Their
baseline scores are shown in Table 5.
1) DAIC-WOZ
The Distress Assessment Interview Corpus Wizard of Oz
(DAIC-WOZ) was first introduced by Scherer et al [44].
Wizard of Oz interviews are conducted by having an animated
The dataset contains more controls than ADHD patients
interviewer (Ellie) who is controlled by a human interviewer
because the competition’s primary goal was to accurately
in another room.
identify controls, while ADHD individual identification was
The DAIC-WOZ dataset contains clinical interviews
a secondary objective. For each participant, the resting state
designed to support the diagnosis of psychological distress
fMRI data was processed through respective pipelines based
conditions. It is composed of recordings and transcripts
on the Statistical Parametric Mapping 8 (SPM8) fMRI anal-
taken from 142 subjects that went through interviews with
ysis package. The processing steps included:
a computer agent. The computer agent, named Ellie, is the
1) Six parameter rigid body motion correction. fundamental feature of a Wizard-Of-Oz style interview. Ellie
2) Non-linear spatial warping (involves estimating and is controlled by a human interviewer in another room with
interpolating) of each participant’s anatomical volume her function being to eradicate interviewer effects. The inter-
to the MNMI T1 template space at a 1 × 1 × 1 mm viewer effect is the influence of the characteristics of an
resolution. interviewer on the responses of the interviewee. To choose the
3) Interpolation of fMRI volumes into the T1 template subjects, recording took place over two sites, a United States
space at a 3 × 3 × 3 mm spatial resolution. (US) Veteran centre and the University of South California
4) Eight millimetre full width at half maximum (FWHM) (USC) Institute for Creative Technologies. The interviews
Gaussian spatial filtering of fMRI volumes. were conducted by one of two female interviewers, both
5) Truncation of resting state fMRI data to length 185 sec- having basic clinical experience.
onds and temporal linear interpolation of all scans into The dataset is ever expanding but to begin with, there were
a sampling rate of 2Hz, 142 subjects with data being collected for: PHQ-8 score, gaze
As a result of the pre-processing, all participants have direction, pose, facial expressivity and acoustic indications
resting state fMRI data that: [45]. The PHQ-8 score is a measure of the severity of the
• is aligned in the MNI T1 Template space. subject’s depression. Along with the PHQ-8 score is a binary
• have the same spatial dimensions of 57×67×50 voxels. label for whether the subject has depression. Furthermore,
• have the same spatial resolution of 3 × 3 × 3 mm voxel the dataset is split into a training set consisting of 107 sub-
size. jects where 30 are depressed and 77 are controls. There is
• have the same temporal dimensions of 370 time-points then a validation set consisting of 35 subjects where 12 are
with a 0.5s volume time. depressed and 23 are controls.
where a voxel is a 3-Dimensional unit of the image with a
single value. V. MACHINE LEARNING AND COMPUTER VISION IN
ADHD DETECTION
B. DEPRESSION A. NON-WEARABLE TECHNIQUES
When the Audio/Visual Emotion Challenge and Workshop 1) IMAGING
(AVEC) is held, there is often a state-of-mind detection Table 6 shows that numerous studies have analyzed imag-
challenge or a detecting depression challenge. The data is ing data using various techniques. Exploiting an SVM is a
released to authors and a competition begins where authors popular approach for classification, applied to both imaging
can improve on the baselines and provide a state-of-the-art and EEG data. This popularity could be attributed to SVM’s
system. Normally datasets used are extensions of the previous ability to capture complex relationships in the data or its
challenges datasets. The last AVEC challenge was its 9th resilience to overfitting. With increasing mental health aware-
proceedings in 2019 [41]. The vast majority of the work ness, it is anticipated that more advanced analysis methods,
86304 VOLUME 11, 2023

TABLE 6. Summary of the research conducted using the ADHD-200 dataset. Where ‘-’ denotes the authors have not specified the result or if validation
has not been performed. To the best of the authors’ knowledge, the official training and testing data split was used throughout the papers discussed in
the table.
VOLUME 11, 2023 86305

such as deep learning, will become widespread, as summa- (ASD) using machine learning. They found that classifica-
rized in Table 6. However, deep learning requires a large tion performance fluctuated significantly depending on the
amount of data, and despite the ADHD-200 dataset being dataset used, and the model that incorporated all data sources
extensive, it is argued that without data augmentation, a deep and utilized repeated cross-validation achieved an AUC of
learning algorithm cannot reliably train and predict outcomes. 0.89 ± 0.01. This study underlined the importance of data
Researchers often used only partial amounts of the ADHD- source selection and rigorous validation techniques, high-
200 dataset in their experiments to reduce bias within the lighting that these tools could serve as a promising new
models, as the dataset contains more controls than patients, avenue for ADHD detection.
causing models to learn control characteristics more effec- In summary, these studies indicate that machine learning
tively than those of ADHD patients. holds significant promise in the field of ADHD diagnosis and
Tang et al. achieved the best-performing study, obtaining a differentiation from other mental health conditions. However,
97.6% testing accuracy with high sensitivity and specificity the robustness and generalizability of the findings are crucial,
[46]. The selected ADHD features were brain Functional and careful attention needs to be paid to the dataset size, qual-
Connectivity (FC), processed through an L2,1 -Norm Linear ity, and the potential for overfitting in model development.
Discrimination Analysis model. The model’s output was then It also suggests that advanced algorithms like LightGBM and
processed through a binary hypothesis testing framework hybrid AI models can handle complex multi-class problems
to make a decision. The authors suggest that the binary more effectively.
hypothesis testing framework can alleviate some issues when
testing with a smaller dataset. Among the studies incorpo- B. WEARABLE TECHNIQUES
rating an SVM, Sartipi et al. achieved the highest accuracy Convolutional Neural Networks (CNNs) have been exten-
at 96.68% [47]. As fMRI data in different brain cortices sively applied in these studies. Amado-Caballero et al. [53]
change rapidly, a robust time-frequency transform (Stockwell achieved a striking classification accuracy of 98.57%, as well
Transformation) was applied. The time-frequency domains as high sensitivity and specificity, demonstrating the poten-
were then partitioned into sub-matrices for calculating their tial of CNNs in interpreting EEG data. Chen et al. [54] and
fuzzy entropies. Moghaddari, Lighvan, and Danishvar [55] further validated
this approach, with the latter achieving an impressive average
testing accuracy of 98.48%. Ahmadi et al. [56] developed a
2) MEDICAL NOTES sophisticated deep CNN model, extracting both spatial and
Bledsoe et al. [48] utilized a SVM to achieve impressive frequency band features and yielding a nearly perfect classi-
classification accuracy, sensitivity, and specificity, each scor- fication accuracy of 99.46% for ADHD subtypes.
ing a perfect 100%. However, it is crucial to consider the Simultaneously, studies using SVMs demonstrate that
limitations of this study. The dataset size was quite small, effective pre-processing techniques and SVM parameter
and there is a substantial risk of overfitting. These results, selection can also yield high classification accuracies.
while promising, may not generalize well to larger, more Chang et al. [57] utilized a decorrelation method with an
heterogeneous populations. independent 2-sample t-test to achieve 80% accuracy, while
Subsequently, Chen et al. [49] and Tachmazidis et al. [50] Chen et al. [58] leveraged Mutual Information analysis to
both used a larger dataset made available by a National Health reach 85.7% accuracy. De-Dea et al. [59] and Rezaeezadeh,
Service specialist mental health provider. While the former Shamekhi, and Shamsi [60] further substantiate the SVM’s
used a Decision Tree and achieved an accuracy of 85.51% and utility, with the latter achieving a remarkable 99.58% classi-
an AUC of 0.87, the latter used a hybrid AI model, achieving fication accuracy.
an accuracy of 95.7% for a three-way classification problem Additionally, wavelet transform techniques, when coupled
(ADHD/No-ADHD/Expert). These studies showcased the with diverse classifiers, have shown promise in ADHD clas-
potential of advanced machine learning techniques in ADHD sification tasks. Notably, Tor et al. [61] achieved a 97.88%
diagnosis while also indicating the importance of large, high- classification accuracy using a kNN classifier, demonstrating
quality datasets in training effective models. the potential efficacy of this technique.
Further extending this approach, Christiansen et al. [51] Tenev et al. [62] took a unique approach, applying multiple
applied LightGBM, a more advanced machine learning classifiers to EEG data under various conditions to categorize
algorithm, to a multi-class classification problem. This prob- ADHD subtypes and controls. While the accuracy fluctuated
lem involved differentiating not just between ADHD and across different conditions (69.2%-82.3%), this approach
control groups but also between subjects with obesity and underlines the importance of considering the context in which
problematic gambling. The algorithm achieved a global accu- EEG data is collected and its impact on model performance.
racy of 80%, with precision and recall varying between Lastly, Poil et al. [63] emphasized the need to consider age
different classes, hinting at the potential complexity when and frequency effects on ADHD-related EEG signal alter-
more conditions are included. ations, pointing out an important factor for future research.
Duda et al. [52] took a different approach by trying to In conclusion, the analysis of EEG data using machine
differentiate between ADHD and Autism Spectrum Disorder learning, particularly CNN and SVM models, has proved to
86306 VOLUME 11, 2023

be a promising tool for ADHD diagnosis and differentiation sion analysis. Using spectral representations processed by
from other conditions. The studies reviewed indicate that CNNs and Artificial Neural Networks (ANN), their model
careful pre-processing of EEG data and appropriate selection achieved a competitive MAE/RMSE score on the AVEC2013
of model parameters are essential to achieve high classifica- test set.
tion accuracy. Furthermore, they highlight the importance of Yang et al. [74] introduced a multi-modal framework that
accounting for factors such as age, ADHD subtypes, and con- exploited video, audio and text data to estimate PHQ-8 scores
text in future research to ensure generalizability and clinical and infer the mental condition of the subject. The combi-
relevance of these models. nation of a deep CNN, DNN, Paragraph Vector, SVM, and
random forest methods showed the value of integrating dif-
ferent types of data in depression detection.
VI. MACHINE LEARNING AND COMPUTER VISION IN Lastly, De Melo, Granger and López [75] tackled the
DEPRESSION DETECTION cost-effectiveness of 3D-CNNs by proposing a deep learn-
A. NON-WEARABLE TECHNIQUES ing architecture that operates without 3D convolutions.
1) VIDEO Despite the reduction in trainable parameters, their MDN
Research by Zhu et al. [64] and Meshram and Rambola [65] still improved performance when compared to existing 3D
use CNNs to extract facial features, either static or dynamic, ResNet models, illustrating the importance of resource opti-
to estimate the BDI-II depression severity score. The model mization without compromising performance.
by Zhu et al. notably improved upon results from the AVEC- In conclusion, the research reviewed supports the notion
2013 dataset, while Meshram and Rambola’s model achieved that a multi-modal, multi-feature approach combining var-
a high classification average of 92.56% on the AVEC-2016 ious machine learning techniques can effectively predict
challenge dataset, suggesting the effectiveness of deep learn- depression severity scores. The findings are congruent in that
ing techniques in detecting depression. they all indicate the potential of machine learning and feature
Similar trends can be observed in the work of extraction in improving depression detection and analysis.
He et al. [66], who also employed the extraction of dynamic However, they also highlight the complexity and the need for
facial features, specifically using the MRLBP-TOP tech- further research to enhance accuracy and achieve effective
nique. The promising RMSE value of 8.90 achieved by this real-time applications.
framework compared to the AVEC-2013 baseline (10.72)
demonstrates the potential of advanced feature extraction
techniques in this field.
Li et al. [67] took a slightly different approach by focus- 2) AUDIO
ing on eye movement as a predictor of depression. Despite The detection and diagnosis of depression have been signif-
this novel approach, the model still achieved a classification icantly advanced by various machine learning approaches,
accuracy of 80.1%, indicating that diverse biological signals particularly with the use of speech analysis. A selection of
can potentially be valuable in depression detection. studies provides an intriguing narrative of the progress and
In the research conducted by Hong et al. [68], the focus development in this area, highlighting the power of speech as
was on distinguishing between bipolar and unipolar disor- a rich source of information for depression analysis.
ders, as well as identifying healthy controls. Their method- Arevian et al. [76] pioneered the use of lexical content,
ology combined action unit descriptors and motion vectors complexity, and vocal expression in tracking depression.
processed by machine learning techniques, yielding a reason- They also leveraged latent semantic analysis (LSA), taking
able classification accuracy of 72.2%. into account semantic coherence, which was an innovative
Zhou et al. [69] introduced the MR-DepressNet, a deep step in expanding the feature pool. Their use of the SVM
regression network that used visual features to estimate classifier established a fundamental framework for future
depression severity. This model improved on the AVEC-2013 research.
baseline RMSE value (13.61) by achieving 8.28, demonstrat- Building upon the traditional feature extraction methods,
ing the potential of visual feature exploitation in improving Chen et al. [77] delved deeper into the hidden structures
depression estimation. within the feature set. Their introduction of a sparse stacked
The research of Tadalagi and Joshi [70], Shang et al. [71], autoencoder to learn higher-quality deep features marked an
and Uddin, Joolee and Lee [72] all took innovative important milestone in enhancing the efficacy of depression
approaches by combining various techniques and method- detection. The significant accuracy of 89% attained by this
ologies. Tadalagi and Joshi implemented their model on a study, using an SVM for classification, underscored the need
real-time system, while Shang et al. introduced a quaternion- for high-quality features.
based method. Uddin, Joolee and Lee used a two-stream deep Meanwhile, Cummins et al. [78] innovated in a different
spatiotemporal network. All of these methods were evaluated direction, examining how depression impacts acoustic mod-
on AVEC datasets and achieved competitive results. els of spectral features. By using Monte Carlo sampling, they
Song et al. [73] focused on the extraction of multi-scale introduced a robust feature estimation method that uniquely
video-level features, providing a novel perspective on depres- connected depression to the Acoustic Volume.
VOLUME 11, 2023 86307

Taking deep learning to the forefront, He and Cao [79] by using nonparametric permutation tests for group com-
used CNNs to extract deep-learned features from raw speech parisons. The effective use of topological metrics as inputs
waveforms and spectrograms. This shift towards automated to classifiers, particularly the SVM-RBF, yielded an impres-
depression analysis tools that could generate complex fea- sive classification accuracy of 83.0%. Guo et al. [86] further
tures signaled a significant step towards more sophisticated refined their approach by constructing an automatic classifier
models. based on a high-order minimum spanning tree functional
Conversely, Jiang et al. [80] proposed an ensemble logis- brain network. Their multi-kernel SVM, after intricate feature
tic regression model, offering separate models for males extraction and selection, achieved an exceptional accuracy of
and females. This gender-specific approach, combined with 97.54%.
diverse feature extraction, added a new dimension to the field, In parallel, Li et al. [87] used voxel-based morphometry
recognizing the potential differences in depression expression (VBM) and regional homogeneity (ReHo) analyses to extract
across genders. key features. By adopting the LASSO approach to isolate
Meanwhile, Li et al. [81] introduced the Multiscale Audio the most informative brain regions, they achieved a validated
Data Normalization (MADN) algorithm, marking another classification accuracy of 86.4% with an SVM classifier.
significant advancement in feature extraction. Their approach Their emphasis on feature extraction highlighted the impor-
further emphasized the potential for innovative methods to tance of selecting relevant regions in the brain for effective
improve upon existing models. depression detection.
Pushing the boundaries of feature analysis, Muzammel Additionally, Rosa et al. [88] proposed a sparse frame-
et al. [82] focused on the acoustic features of vowel and work for depression classification. Their utilization of sparse
consonant spaces. Their method of augmenting data and seg- inverse covariance models to estimate functional connectiv-
menting speech brought attention to the nuanced elements of ity, coupled with an L1-norm SVM, resulted in an accuracy of
speech and their potential role in depression detection. 85%. This approach once again demonstrated the importance
A notable turning point occurs with Zhao et al. [83], who of functional connectivity in the detection of depression.
introduced a comprehensive approach combining unsuper- Conversely Li et al. [89], who leveraged independent com-
vised learning, hierarchical attention, and knowledge transfer. ponent analysis to define the triple network model. Their
Their impressive results underscored the potential of com- integration of effective connectivity features, dynamic func-
plex deep learning architectures in determining depression tional connectivity features, and rigorous statistical testing
severity. led to an accuracy of 90.91% with an SVM classifier.
In summary, this collection of studies forms an engaging Simultaneously, Sen et al. [90] focused on dynamic and
narrative that underscores the evolution of speech analysis static connectivity measures, extracted from rs-fMRI data,
in depression detection. From basic lexical analysis to deep as a basis for feature extraction. Their use of Pearson’s cor-
learning, and from gender-specific models to nuanced vowel relation and entropy measures resulted in a combination of
and consonant analysis, it shows a consistent progression static and dynamic features, yielding a classification result of
towards more complex and sophisticated models. This nar- 82% with an RBF-SVM.
rative serves as a testament to the ongoing advancement Finally, Wang et al. [91] distinguished themselves by using
of machine learning techniques in mental health research, functional near-infrared spectroscopy (fNIR) instead of
particularly in understanding and addressing depression. fMRI. By utilizing the unique properties of near-infrared light
and the absorptive characteristics of blood, they were able to
extract crucial features. With an AlexNet structured network,
3) IMAGING they achieved an impressive accuracy of 90%.
Depression detection research has expanded beyond speech Together, these studies form a compelling narrative that
analysis to leverage advanced neuroimaging technologies, showcases the interplay between neuroimaging technologies
and machine learning has remained a constant ally. Various and machine learning in depression detection. From func-
studies have illustrated how different feature extraction and tional connectivity and brain network analysis to innovative
classification methods can yield significant results in identi- uses of light in fNIR, the research direction showcases a
fying depression at an individual level. continued progression toward individual-level analysis and
In a novel approach, Cao et al. [84] made strides by focus- a growing emphasis on sophisticated feature extraction and
ing on the individual rather than group dynamics, using selection methods. The consistent use of SVM classifiers
probability density functions (PDFs) to target functional con- across most studies points to their effectiveness in this con-
nectivity. They integrated a t-test for primitive selection, text, further highlighting the importance of machine learning
followed by Kernel density estimation for PDFs, resulting in in mental health research.
a significant classification accuracy of 84.21% using an SVM
classifier. This shift towards an individual-oriented approach
marked a significant milestone in depression detection. 4) MEDICAL NOTES
Building upon functional connectivity, Guo et al. [85] fur- De-Souza et al. [92] sought to develop a machine learning
ther demonstrated the potential of brain network analysis tool to detect depression using an amalgamation of clinical,
86308 VOLUME 11, 2023

laboratory, and sociodemographic data. The Random Forest patients. Their system, MTERMS, consistently outperformed
(RF) algorithm emerged as the best performer, achieving a standard classifiers, reaffirming the strength of machine
robust classification accuracy of 89% and an AUC of 0.87. learning in interpreting unstructured text data.
This study solidifies the foundational premise that machine Together, these studies craft a compelling tale of how
learning can effectively discern depression from a combina- machine learning has been applied to diverse data types and
tion of diverse data types. challenges in depression detection, consistently achieving
Next, Liu et al. [93] introduced EarlyDetect (ED), a com- impressive results. It illustrates a trend towards increasingly
posite screening application utilizing machine learning to complex and real-world applicable models, with promising
incorporate a wide spectrum of variables. From family his- indications for the future of machine learning in mental health
tory of mental illness to suicide ideation, ED exemplifies diagnostics.
a comprehensive approach. Using the ElasticNet algorithm,
it achieved a balanced accuracy of 72% with an AUC of B. WEARABLE TECHNIQUES
0.781. This underlines the potential for machine learning A vast amount of research has been conducted into detect-
to be effective in complex, real-world settings, synthesizing ing depression using a wearable sensors that produce EEG
multiple factors into predictive models. signals. Table 7 is summarising the best performing studies
Adding to the narrative, Ma et al. [94] strived to create a with the most popular classification method being CNNs and
machine learning framework that could expedite the Affec- SVMs.
tive Disorder Evaluation scale. The resulting Bipolar Diag- Following on, different classification methods have been
nosis Checklist in Chinese (BDCC), which used the RF used with EEG signals. Firstly, Cai et al. [101] introduced a
algorithm to rank feature importance, achieved an outstand- multimodal model fusing different EEG data sources, gath-
ing classification accuracy of 99.6%. The success of BDCC ered under a range of emotional conditions. Feature weight-
emphasises the role machine learning can play in simplifying ing was performed through a genetic algorithm on linear
and accelerating mental health evaluations. and non-linear features, with this unique approach leading
Simultaneously, Mato-Abad et al. [95] leveraged Arti- to a robust classification accuracy of 86.98%. This method
ficial Neural Networks (ANNs) to identify a subtype of emphasizes the potential of multimodal models in detecting
mild cognitive impairment (MCI) associated with depression. depression and accentuates the strength of genetic algorithms
The ANN’s success, with an 86% accuracy, highlights the in feature weighting.
potential of machine learning in elucidating the nuanced Akbari et al. used k-Nearest Neighbours (k-NN) with
intersections between different mental health conditions. geometric features extracted from the EEG signals’ Self-
On a parallel track, Meng et al. [96] introduced a temporal Organising Decision Process (SODP) [102]. The Binary Par-
deep learning model performing bi-directional representation ticle Swarm Optimisation (BPSO) algorithm was utilised for
learning on Electronic Health Record (EHR) sequences. The feature selection, culminating in impressive results: 98.79%
model’s AUC ranged from 0.73 - 0.85, based on the prediction classification accuracy, 97.72% sensitivity, and 99.86%
window timeframe. This exploration of temporal modelling specificity. This work underscores the value of geometric fea-
in EHR data showcases machine learning’s ability to draw tures and the effectiveness of the BPSO algorithm in feature
insights from longitudinal health data. selection for depression detection.
In another initiative, Meng et al. [97] devised a model Moreover, Li et al. [103] ventured to extract multiple lin-
incorporating temporal Hierarchical Clinical Embeddings ear and non-linear features from EEG signals. A rigorous
with Topic Modelling (HCET), addressing data sparsity comparison of five different feature selection methods was
issues. The improvement in AUCs further emphasized the carried out, with significant discriminant features being iden-
potential of machine learning in handling complex, sparse tified using Bonferroni correction t-tests. The outcome was a
datasets. commendable average classification accuracy of 95%, high-
Adding another dimension, Parker et al. [98] sought to lighting the importance of meticulous feature selection in
discriminate between bipolar and unipolar subjects. They achieving high classification accuracy.
achieved a promising classification accuracy of 96%, demon- Simultaneously, Saeedi et al. applied sample and approx-
strating machine learning’s capacity to distinguish between imate entropy to wavelet packets, with significant features
different mental health disorders, even within the challenging selected using a Genetic Algorithm (GA) [104]. This method
context of unbalanced datasets. achieved a classification accuracy, sensitivity, and specificity
Sharma and Verbeke [99] employed the Extreme Gradi- of 98.44%, 97.10%, and 100% respectively. The use of GA
ent Boosting algorithm on a biomarker dataset, achieving once again demonstrates its potency in feature selection,
a balanced classification accuracy of 94.42% despite ini- enhancing the classification performance.
tial dataset biases. This finding underscores machine learn- Furthermore, research has also delved into the utilization
ing’s robustness and adaptability in the face of imbalanced of MLPs in classifying EEG data. Ahmadlou et al. employed
data. a wavelet-chaos methodology, using Katz’s and Higuchi’s
Lastly, Zhou et al. [100] leveraged natural language pro- fractal dimensions as measures of nonlinearity and com-
cessing in analysing discharge summaries of depressed plexity [105]. The resulting 91.3% classification accuracy
VOLUME 11, 2023 86309

86310
TABLE 7. The list of studies where EEG signals have been analysed to detect depression. ‘-’ denotes that the author has not specified a value and ‘*’ denotes an average value. Different datasets have been
used throughout the studies, but the table concisely summarises the effectiveness of EEG signals for classification of Depression.
VOLUME 11, 2023

TABLE 7. (Continued.) The list of studies where EEG signals have been analysed to detect depression. ‘-’ denotes that the author has not specified a value and ‘*’ denotes an average value. Different
datasets have been used throughout the studies, but the table concisely summarises the effectiveness of EEG signals for classification of Depression.
VOLUME 11, 2023

86311
illuminates the potential of fractal dimensions and chaos (EULA) would have to be signed to protect the organisers and
theory in depression detection. subjects’ data that is involved within the dataset.
In a parallel effort, Cukic et al. examined Higuchi’s Machine learning is transforming the landscape of ADHD
Fractal Dimension and Sample Entropy as non-linear mea- and depression detection and classification through innova-
sures in discriminating between depressed patients and con- tive data collection and analysis methods. These encompass
trols [106]. By leveraging Principal Component Analysis imaging techniques, processing of medical notes, and wear-
(PCA) for feature dimensionality reduction, they achieved an able technology, reflecting ADHD’s complex nature and
average classification accuracy of 97.56%. This reinforces showcasing machine learnings potential in diagnosis and
the idea that non-linear measures can be highly discrimina- treatment.
tive and the role of dimensionality reduction techniques in ADHD diagnosis has seen successful employment of imag-
boosting classification performance. ing techniques, leveraging SVM and deep learning models.
Overall, these studies show the exploration into the detec- Despite needing large data sets and often dealing with unbal-
tion of depression using EEG signals. They illustrate the anced ADHD-200 datasets, these challenges are overcome
evolution of methodologies, from the use of different feature using data augmentation and hypothesis testing frameworks.
extraction techniques to the application of various machine High classification accuracies from multiple studies rein-
learning algorithms. The consistently high classification force the value of imaging data in ADHD detection. ML has
accuracies across studies reinforce the potential of these also proven successful in extracting rich clinical informa-
approaches in advancing depression detection. tion from medical notes, with Decision Trees, SVMs, and
hybrid AI models delivering impressive classification accu-
racy. While there are issues like overfitting and data het-
VII. CONCLUSION AND FUTURE WORK erogeneity, these applications highlight the role of AI in
This survey has gone into detail about machine learning appli- clinical decision-making. Incorporating wearable technology
cations in mental health detection. It can be observed that the provides a non-invasive means of collecting EEG signals for
most popular methods for automatic detection of depression ADHD classification. Techniques such as CNNs and SVMs
and ADHD is by exploiting imaging data and EEG data. The have been effective in analyzing this data. However, ensur-
non-intrusive nature of the EEG provides an argument that ing the models’ applicability to new patients and real-world
it is the preferred choice. This is due to the vast amount of conditions remains a challenge.
methods that can be applied to analysis, while causing no For depression detection, machine learning has simi-
harm to the subject. larly demonstrated remarkable adaptability and effective-
The biggest drawback about research involving mental ness. Brain imaging data, clinical notes, sociodemographic
health conditions is the size of the dataset. Due to the nature data, laboratory data, wearable sensor data, and electronic
of the conditions, for both ADHD and depression it is dif- health records have all been effectively utilized. Algorithms
ficult to get enough subjects to participate in the research. such as SVMs, Random Forest, ElasticNet, Extreme Gradi-
Furthermore there are possible implications with protecting ent Boosting, and Artificial Neural Networks have yielded
the privacy of all subjects due to it being very sensitive data. high accuracy rates across diverse data sources. Moreover,
When subjects have agreed to have their data used, there is machine learning’s success in discerning between differ-
also the issue of whether the data can be publicly shared or ent depressive disorders could revolutionize personalized
whether it remains private. Lastly, with regards to ADHD and treatment.
depression, the spectrum of behaviour is vast, meaning some However, the quality of machine learning models is con-
behaviour is very rigid or too excitatory. Therefore, training tingent on the quality of data they’re trained on. Continued
a classifier to detect these behaviours can be even harder as efforts are essential to ensure the robustness and applicability
there is not enough data to cover such a vast spectrum. of these models across various populations and settings. The
Following on, there is more research being conducted Intelligent Sensing Group at Newcastle University is con-
into depression. This could be due to the awareness of ducting their own Intelligent Sensing ADHD trial (ISAT) that
the mental health condition being bigger or because of the involves audio-visual data of controls and ADHD subjects.
available datasets. We suggest that for both ADHD and The aim is for this data to be publicly available once correctly
Depression respectively, there is a collective movement for a processed.
joint database containing multimodal data for the respective In conclusion, machine learning offers substantial potential
mental health conditions. Within these databases, there would for improving ADHD and depression diagnostics. Despite
be an established method for protecting the participants pri- challenges related to data quality, overfitting, and algorithm
vacy such as converting their identity to a number/letter and interpretability, machine learnings ability to identify patterns
processing the video/image data using techniques such as in complex datasets makes it a valuable tool in mental health
the Histogram of Gradients. The file types would be made research. Future efforts should focus on creating reliable
consistent so that all users would know what to expect and models, protecting patient data, and ensuring models can be
baseline scores would be achieved to provide state-of-the-art generalized to different populations. Effective collaboration
comparisons. Lastly, for use in research, an End User Licence between clinicians, data scientists, and patients will be key
86312 VOLUME 11, 2023

to maximizing machine learning’s potential in mental health [20] D. Stevenson and P. Farmer, ‘‘Thriving at work: A review of mental
diagnosis and treatment. health and employers,’’ Dept. Health Social Care UK, London, U.K.,
Tech. Rep., Oct. 2017. [Online]. Available: https://assets.publishing.
service.gov.uk/government/uploads/system/uploads/attachment_data/file/
REFERENCES 658145/thriving-at-work-stevenson-farmer-review.pdf
[21] J. Banks and X. Xu, The Mental Health Effects of the First Two Months
[1] J. W. Kanter, A. M. Busch, C. E. Weeks, and S. J. Landes, ‘‘The nature
of Lockdown and Social Distancing During the COVID-19 Pandemic in
of clinical depression: Symptoms, syndromes, and behavior analysis,’’
the UK. Linköping, Sweden: IFS, Jun. 2020.
Behav. Anal., vol. 31, no. 1, pp. 1–21, Apr. 2008.
[22] S. Gallagher and M. A. Wetherell, ‘‘Risk of depression in family care-
[2] K. Low. Understanding Hypersensitivity in ADHD. Accessed:
givers: Unintended consequence of COVID-19,’’ BJPsych Open, vol. 6,
Jan. 6, 2023. [Online]. Available: https://www.verywellmind.com/
no. 6, pp. 1–5, 2020, Art. no. e119.
sensitivities-and-adhd-20473
[23] R. Stewart, A. Jewell, M. Broadbent, I. Bakolis, and J. Das-Munshi,
[3] World Health Organisation. Depressive Disorder (Depression).
‘‘Causes of death in mental health service users during the first
Accessed: Jan. 6, 2023. [Online]. Available: https://www.who.int/news-
wave of the COVID-19 pandemic: South London and Maudsley data
room/fact-sheets/detail/depression
from March to June 2020, compared with 2015–2019,’’ MedRxiv, doi:
[4] R. Thomas, S. Sanders, J. Doust, E. Beller, and P. Glasziou, ‘‘Preva- 10.1101/2020.10.25.20219071.
lence of attention-deficit/hyperactivity disorder: A systematic review [24] H. Yang, W. Chen, Y. Hu, Y. Chen, Y. Zeng, Y. Sun, Z. Ying,
and meta-analysis,’’ Pediatrics, vol. 135, no. 4, pp. e994–e1001, J. He, Y. Qu, D. Lu, F. Fang, U. Valdimarsdóttir, and H. Song, ‘‘Pre-
Apr. 2015. pandemic psychiatric disorders and risk of COVID-19: A cohort analysis
[5] P. Song, M. Zha, Q. Yang, Y. Zhang, X. Li, and I. Rudan, ‘‘The in the U.K. biobank,’’ Lancet Healthy Longevity, Tech. Rep., 2020,
prevalence of adult attention-deficit hyperactivity disorder: A global sys- vol. 1, no. 2. [Online]. Available: https://www.thelancet.com/journals/
tematic review and meta-analysis,’’ J. Global Health, vol. 11, Feb. 2021, lanhl/article/PIIS2666-7568(20)30013-1/fulltext
Art. no. 04009. [25] J. S. Evert, C. C. Hannie, J. J. S. Kooij, M. Marieke, T. F. B. Aartjan,
[6] U. P. Ramtekkar, A. M. Reiersen, A. A. Todorov, and R. D. Todd, ‘‘Sex and J. H. D. Dorly, ‘‘The role of adverse life events on depression in
and age differences in attention-deficit/hyperactivity disorder symptoms older adults with ADHD,’’ J. Affect. Disorders, vol. 174, pp. 574–579,
and diagnoses: Implications for DSM-V and ICD-11,’’ J. Amer. Acad. Mar. 2015.
Child Adolescent Psychiatry, vol. 49, no. 3, pp. 217–283, 2010. [26] A. J. Lundervold, S. P. Hinshaw, L. Sørensen, and M.-B. Posserud, ‘‘Co-
[7] Z. Huntley, S. Maltezos, C. Williams, A. Morinan, A. Hammon, occurring symptoms of attention deficit hyperactivity disorder (ADHD)
D. Ball, E. J. Marshall, F. Keaney, S. Young, P. Bolton, K. Glaser, in a population-based sample of adolescents screened for depression,’’
R. Howe-Forbes, J. Kuntsi, K. Xenitidis, D. Murphy, and P. J Asherson, BMC Psychiatry, vol. 16, no. 1, pp. 1–10, Dec. 2016.
‘‘Rates of undiagnosed attention deficit hyperactivity disorder in London [27] C. Nash, R. Nair, and S. M. Naqvi, ‘‘Machine learning and ADHD mental
drug and alcohol detoxification units,’’ BMC Psychiatry, vol. 12, p. 223, health detection—A short survey,’’ in Proc. 25th Int. Conf. Inf. Fusion
Dec. 2012. (FUSION), Jul. 2022, pp. 1–8.
[8] BBC. ADHD Diagnosis for Adults ’Can Take Seven Years. Accessed: [28] D. S. B. A. Hamid, S. B. Goyal, and P. Bedi, ‘‘Integration of deep learning
Jan. 3, 2022. [Online]. Available: https://www.bbc.co.uk/news/ for improved diagnosis of depression using EEG and facial features,’’
uk-england-44956540#:~:text=About%201.5%20million%20adults% Mater. Today, Proc., vol. 80, pp. 1965–1969, Jan. 2023.
20in,only%20120%2C000%20are%20formally%20diagnosed [29] A. B. R. Shatte, D. M. Hutchinson, and S. J. Teague, ‘‘Machine learning in
[9] B. T. Wymbs, W. E. Pelham, B. S. G. Molina, E. M. Gnagy, T. K. Wilson, mental health: A scoping review of methods and applications,’’ Psychol.
and J. B. Greenhouse, ‘‘Rate and predictors of divorce among parents Med., vol. 49, no. 9, pp. 1426–1448, Jul. 2019.
of youth with ADHD,’’ J. Consulting Clin. Psychol., vol. 76, no. 5, [30] S. Graham, C. Depp, E. E. Lee, C. Nebeker, X. Tu, H.-C. Kim, and
pp. 735–744, Oct. 2008. D. V. Jeste, ‘‘Artificial intelligence for mental health and mental ill-
[10] K. Doheny, ‘‘Divorce more likely in ADHD families?’’ WebMD, nesses: An overview,’’ Current Psychiatry Rep., vol. 21, no. 11, p. 116,
Tech. Rep., 2008. [Online]. Available: https://www.cbsnews.com/ Nov. 2019.
news/divorce-more-likely-in-adhd-families/ [31] H. W. Loh, C. P. Ooi, P. D. Barua, E. E. Palmer, F. Molinari,
[11] J. A. Doshi, P. Hodgkins, J. Kahle, V. Sikirica, M. J. Cangelosi, and U. R. Acharya, ‘‘Automated detection of ADHD: Current trends
J. Setyawan, M. H. Erder, and P. J. Neumann, ‘‘Economic impact of and future perspective,’’ Comput. Biol. Med., vol. 146, Jul. 2022,
childhood and adult attention-deficit/hyperactivity disorder in the United Art. no. 105525.
States,’’ J. Amer. Acad. Child Adolescent Psychiatry, vol. 51, no. 10, [32] N. Sethu and R. Vyas, ‘‘Overview of machine learning methods in
pp. 990–1002, Oct. 2012. ADHD prediction,’’ in Advances in Bioengineering. Springer, May 2020,
[12] C. Mohr-Jensen and H.-C. Steinhausen, ‘‘A meta-analysis and systematic pp. 51–71.
review of the risks associated with childhood attention-deficit hyperactiv- [33] The ICD-10 Classification of Mental and Behavioural Disorders: Diag-
ity disorder on long-term outcome of arrests, convictions, and incarcera- nostic Criteria for Research, World Health Organization, Geneva,
tions,’’ Clin. Psychol. Rev., vol. 48, pp. 32–42, Aug. 2016. Switzerland, 1993.
[13] Diagnostic and Statistical Manual of Mental Disorders (DSM-5-TR), [34] Attention Deficit Hyperactivity Disorder, National Institute of Mental
5th ed., Amer. Psychiatric Assoc., Arlington, VA, USA, 2013. Health, Bethesda, MD, USA, 2022.
[14] C. K. Conners, J. Pitkanen, and S. R. Rzepa, Conners (Conners 3; Con- [35] K. Kroenke, R. L. Spitzer, and J. B. W. Williams, ‘‘The PHQ-9: Validity
ners 2008), 3rd ed. New York, NY, USA: Springer, 2011, pp. 675–678. of a brief depression severity measure,’’ J. Gen. Internal Med., vol. 16,
[15] I. Macrory, Measuring National Well-Being: Life in the UK. South Wales, no. 9, pp. 606–613, Sep. 2001.
Wales: Office for UK National Statistics, 2016. [36] J. Ware and C. D. Sherbourne, ‘‘The MOS 36-item short-form health
[16] W. H. Organisation, ‘‘Depression and other common mental disor- survey (SF-36). I. Conceptual framework and item selection,’’ Med. Care,
ders: Global health estimates,’’ World Health Organisation, Geneva, vol. 30, no. 6, pp. 473–483, 1992.
Switzerland, Tech. Rep. WHO/MSD/MER/2017.2, 2017. [Online]. Avail- [37] J. Upton, Beck Depression Inventory (BDI). New York, NY, USA:
able: https://apps.who.int/iris/handle/10665/254610 Springer, 2013, pp. 178–179.
[17] J. Alonso et al., ‘‘Treatment gap for anxiety disorders is global: Results [38] G. H. Glover, ‘‘Overview of functional magnetic resonance imaging,’’
of the world mental health surveys in 21 countries,’’ Depression Anxiety, Neurosurg. Clinics North Amer., vol. 22, no. 2, p. 133–139, 2011.
vol. 35, no. 3, pp. 195–208, Mar. 2018. [39] M. R. G. Brown, G. S. Sidhu, R. Greiner, N. Asgarian, M. Bastani,
[18] T. Beiwinkel, S. Kindermann, A. Maier, C. Kerl, J. Moock, G. Bar- P. H. Silverstone, A. J. Greenshaw, and S. M. Dursun, ‘‘ADHD-200
bian, and W. Rössler, ‘‘Using smartphones to monitor bipolar disorder global competition: Diagnosing ADHD using personal characteristic data
symptoms: A pilot study,’’ JMIR Mental Health, vol. 3, no. 1, p. e2, can outperform resting state fMRI measurements,’’ Frontiers Syst. Neu-
Jan. 2016. rosci., vol. 6, p. 69, Sep. 2012.
[19] U.K.-Government. Labour Force Survey: Self Reported Work-Related Ill [40] N. Cauli and D. Reforgiato Recupero, ‘‘Survey on videos data augmen-
Health Workplace Injuries. Accessed: Oct. 5, 2021. [Online]. Available: tation for deep learning models,’’ Future Internet, vol. 14, no. 3, p. 93,
https://www.hse.gov.uk/statistics/lfs/index.html Mar. 2022.
VOLUME 11, 2023 86313

[41] AVEC 19: Proceedings of the 9th International on Audio/Visual Emo- [58] H. Chen, J. Yan, Y. Gu, Y. Song, and X. Li, ‘‘Mutual information analysis
tion Challenge and Workshop, Association for Computing Machinery, of EEG of children with attention-deficit hyperactivity disorder,’’ in Proc.
New York, NY, USA, 2019. Chin. Autom. Congr. (CAC), Oct. 2017, pp. 2342–2347.
[42] M. Valstar, B. Schuller, K. Smith, F. Eyben, B. Jiang, S. Bilakhia, [59] F. D. Dea, M. Ajcevic, M. Stecca, C. Zanus, M. Carrozzi, A. Cuzzocrea,
S. Schnieder, R. Cowie, and M. Pantic, ‘‘AVEC 2013: The continuous and A. Accardo, ‘‘A big-data-analytics framework for supporting classi-
audio/visual emotion and depression recognition challenge,’’ in Proc. 3rd fication of ADHD and healthy children via principal component analysis
ACM Int. Workshop Audio/Vis. Emotion Challenge, Barcelona, Spain, of EEG sleep spindles power spectra,’’ Proc. Comput. Sci., vol. 159,
2013. [Online]. Available: https://www.researchgate.net/publication/ pp. 1584–1590, 2019. [Online]. Available: https://www.researchgate.
262157517_AVEC_2013_-_The_continuous_AudioVisual_Emotion_ net/publication/336537750_A_Big-Data-Analytics_Framework_for_
and_depression_recognition_challenge Supporting_Classification_of_ADHD_and_Healthy_Children_via_
[43] M. Valstar, B. Schuller, K. Smith, T. Almaev, F. Eyben, J. Krajewski, Principal_Component_Analysis_of_EEG_Sleep_Spindles_Power_
R. Cowie, and M. Pantic, ‘‘AVEC 2014: 3D dimensional affect Spectra, doi: 10.1016/j.procs.2019.09.329.
and depression recognition challenge,’’ in Proc. 4th Int. Workshop [60] M. Rezaeezadeh, S. Shamekhi, and M. Shamsi, ‘‘Attention deficit hyper-
Audio/Visual Emotion Challenge (AVEC), Orlando, FL, USA, activity disorder diagnosis using non-linear univariate and multivariate
2014. [Online]. Available: https://www.researchgate.net/publication/ EEG measurements: A preliminary study,’’ Phys. Eng. Sci. Med., vol. 43,
287702683_AVEC_2014_-_3D_dimensional_affect_and_depression_ no. 2, pp. 577–592, Jun. 2020.
recognition_challenge#fullTextFileContent [61] H. T. Tor, C. P. Ooi, N. S. Lim-Ashworth, J. K. E. Wei, V. Jahmunah,
[44] C. Uyulan, S. de la Salle, T. T. Erguzel, E. Lynn, P. Blier, V. Knott, S. L. Oh, U. R. Acharya, and D. S. S. Fung, ‘‘Automated detection of
M. M. Adamson, M. Zelka, and N. Tarhan, ‘‘Depression diagnosis mod- conduct disorder and attention deficit hyperactivity disorder using decom-
eling with advanced computational methods: Frequency-domain eMVAR position and nonlinear techniques with EEG signals,’’ Comput. Methods
and deep learning,’’ Clin. EEG Neurosci., vol. 53, no. 1, pp. 24–36, Programs Biomed., vol. 200, Mar. 2021, Art. no. 105941.
Jan. 2022.
[62] A. Tenev, S. Markovska-Simoska, L. Kocarev, J. Pop-Jordanov,
[45] B. Ay, O. Yildirim, M. Talo, U. B. Baloglu, G. Aydin, S. D. Puthankattil,
A. Müller, and G. Candrian, ‘‘Machine learning approach for classifica-
and U. R. Acharya, ‘‘Automated depression detection using deep repre-
tion of ADHD adults,’’ Int. J. Psychophysiol., vol. 93, no. 1, pp. 162–166,
sentation and sequence learning with EEG signals,’’ J. Med. Syst., vol. 43,
Jul. 2014.
no. 7, p. 205, Jul. 2019.
[63] S. S. Poil, S. Bollmann, C. Ghisleni, R. L. O’Gorman, P. Klaver, J. Ball,
[46] Y. Tang, X. Li, Y. Chen, Y. Zhong, A. Jiang, and C. Wang, ‘‘High-accuracy
D. Eich-Höchli, D. Brandeis, and L. Michels, ‘‘Age dependent elec-
classification of attention deficit hyperactivity disorder with l2,1 -norm
troencephalographic changes in attention-deficit hyperactivity disorder
linear discriminant analysis and binary hypothesis testing,’’ IEEE Access,
(ADHD),’’ Clin Neurophysiol, vol. 125, no. 8, pp. 1626–1638, 2014.
vol. 8, pp. 56228–56237, 2020.
[47] S. Sartipi, H. Kalbkhani, P. Ghasemzadeh, and M. G. Shayesteh, ‘‘Stock- [64] Y. Zhu, Y. Shang, Z. Shao, and G. Guo, ‘‘Automated depression diagnosis
well transform of time-series of fMRI data for diagnoses of attention- based on deep networks to encode facial appearance and dynamics,’’
deficit hyperactivity disorder,’’ Appl. Soft Comput., vol. 86, Jan. 2020, IEEE Trans. Affect. Comput., vol. 9, no. 4, pp. 578–584, Oct. 2018.
Art. no. 105905. [65] P. Meshram and R. K. Rambola, ‘‘Diagnosis of depression level using
[48] J. C. Bledsoe, C. Xiao, A. Chaovalitwongse, S. Mehta, T. J. Grabowski, multimodal approaches using deep learning techniques with multiple
M. Semrud-Clikeman, S. Pliszka, and D. Breiger, ‘‘Diagnostic classifica- selective features,’’ Expert Syst., vol. 40, no. 4, May 2023.
tion of ADHD versus control: Support vector machine classification using [66] L. He, D. Jiang, and H. Sahli, ‘‘Automatic depression analysis using
brief neuropsychological assessment,’’ J. Attention Disorders, vol. 24, dynamic facial appearance descriptor and Dirichlet process Fisher encod-
no. 11, pp. 1547–1556, 2020. ing,’’ IEEE Trans. Multimedia, vol. 21, no. 6, pp. 1476–1486, Jun. 2019.
[49] T. Chen, G. Antoniou, M. Adamou, I. Tachmazidis, and P. Su, [67] X. Li, T. Cao, S. Sun, B. Hu, and M. Ratcliffe, ‘‘Classification
‘‘Automatic diagnosis of attention deficit hyperactivity disorder using study on eye movement data: Towards a new approach in depression
machine learning,’’ Appl. Artif. Intell., vol. 35, no. 9, pp. 657–669, detection,’’ in Proc. IEEE Congr. Evol. Comput. (CEC), Jul. 2016,
Jul. 2021. pp. 1227–1232.
[50] I. Tachmazidis, T. H. Chen, M. Adamou, and G. Antoniou, ‘‘A hybrid [68] Q.-B. Hong, C.-H. Wu, M.-H. Su, and C.-C. Chang, ‘‘Exploring macro-
AI approach for supporting clinical diagnosis of attention-deficit hyper- scopic and microscopic fluctuations of elicited facial expressions for
activity disorder (ADHD) in adults,’’ Health Inf. Sci. Syst., vol. 9, no. 1, mood disorder classification,’’ IEEE Trans. Affect. Comput., vol. 12, no. 4,
pp. 1–8, 2021. pp. 989–1001, Oct. 2021.
[51] H. Christiansen, M.-L. Chavanon, O. Hirsch, M. H. Schmidt, C. Meyer, [69] X. Zhou, K. Jin, Y. Shang, and G. Guo, ‘‘Visually interpretable repre-
A. Müller, H.-J. Rumpf, I. Grigorev, and A. Hoffmann, ‘‘Use of machine sentation learning for depression recognition from facial images,’’ IEEE
learning to classify adult ADHD and other conditions based on the Trans. Affect. Comput., vol. 11, no. 3, pp. 542–552, Jul. 2020.
Conners’ adult ADHD rating scales,’’ Sci. Rep., vol. 10, Nov. 2020, [70] M. Tadalagi and A. M. Joshi, ‘‘AutoDep: Automatic depression detection
Art. no. 18871. using facial expressions based on linear binary pattern descriptor,’’ Med.
[52] M. Duda, N. Haber, J. Daniels, and D. P. Wall, ‘‘Crowdsourced validation Biol. Eng. Comput., vol. 59, no. 6, pp. 1339–1354, Jun. 2021.
of a machine-learning classification system for autism and ADHD,’’
[71] Y. Shang, Y. Pan, X. Jiang, Z. Shao, G. Guo, T. Liu, and H. Ding,
Transl. Psychiatry, vol. 7, May 2017, Art. no. e1133.
‘‘LQGDNet: A local quaternion and global deep network for facial
[53] P. Amado-Caballero, P. Casaseca-de-la-Higuera, S. Alberola-Lopez,
depression recognition,’’ IEEE Trans. Affect. Comput., early access,
J. M. Andres-de-Llano, J. A. L. Villalobos, J. R. Garmendia-Leiza, and
Dec. 31, 2021, doi: 10.1109/TAFFC.2021.3139651.
C. Alberola-Lopez, ‘‘Objective ADHD diagnosis using convolutional
[72] M. A. Uddin, J. B. Joolee, and Y.-K. Lee, ‘‘Depression level predic-
neural networks over daily-life activity records,’’ IEEE J. Biomed. Health
tion using deep spatiotemporal features and multilayer bi-LTSM,’’ IEEE
Informat., vol. 24, no. 9, pp. 2690–2700, Sep. 2020.
[54] H. Chen, W. Chen, Y. Song, L. Sun, and X. Li, ‘‘EEG characteristics Trans. Affect. Comput., vol. 13, no. 2, pp. 864–870, Apr. 2022.
of children with attention-deficit hyperactivity disorder,’’ Neuroscience, [73] S. Song, S. Jaiswal, L. Shen, and M. Valstar, ‘‘Spectral representation
vol. 406, pp. 444–456, May 2019. of behaviour primitives for depression analysis,’’ IEEE Trans. Affect.
[55] M. Moghaddari, M. Z. Lighvan, and S. Danishvar, ‘‘Diagnose ADHD Comput., vol. 13, no. 2, pp. 829–844, Apr. 2022.
disorder in children using convolutional neural network based on contin- [74] L. Yang, D. Jiang, and H. Sahli, ‘‘Integrating deep and shallow models for
uous mental task EEG,’’ Comput. Methods Programs Biomed., vol. 197, multi-modal depression analysis—Hybrid architectures,’’ IEEE Trans.
Dec. 2020, Art. no. 105738. Affect. Comput., vol. 12, no. 1, pp. 239–253, Jan. 2021.
[56] A. Ahmadi, M. Kashefi, H. Shahrokhi, and M. A. Nazari, ‘‘Computer [75] W. C. de Melo, E. Granger, and M. B. López, ‘‘MDN: A deep
aided diagnosis system using deep convolutional neural networks for maximization-differentiation network for spatio-temporal depression
ADHD subtypes,’’ Biomed. Signal Process. Control, vol. 63, Jan. 2021, detection,’’ IEEE Trans. Affect. Comput., vol. 14, no. 1, pp. 578–590,
Art. no. 102227. Jan. 2023.
[57] M.-Y. Chang, C.-S. Ouyang, C.-T. Chiang, R.-C. Yang, R.-C. Wu, [76] A. C. Arevian, D. Bone, N. Malandrakis, V. R. Martinez, K. B. Wells,
H.-C. Wu, and L.-C. Lin, ‘‘A new method of diagnosing attention-deficit D. J. Miklowitz, and S. Narayanan, ‘‘Clinical state tracking in serious
hyperactivity disorder in male patients by quantitative EEG analysis,’’ mental illness through computational analysis of speech,’’ PLoS ONE,
Clin. EEG Neurosci., vol. 50, no. 5, pp. 339–347, Sep. 2019. vol. 15, no. 1, Jan. 2020, Art. no. e0225695.
86314 VOLUME 11, 2023

[77] H. Chen, Y. Lin, Y. Li, W. Wang, P. Wang, and Y. Lei, ‘‘Hybrid fea- [97] Y. Meng, W. Speier, M. Ong, and C. W. Arnold, ‘‘HCET: Hierarchical
ture embedded sparse stacked autoencoder and manifold dimensionality clinical embedding with topic modeling on electronic health records for
reduction ensemble for mental health speech recognition,’’ IEEE Access, predicting future depression,’’ IEEE J. Biomed. Health Informat., vol. 25,
vol. 9, pp. 28729–28741, 2021. no. 4, pp. 1265–1272, Apr. 2021.
[78] N. Cummins, V. Sethu, J. Epps, S. Schnieder, and J. Krajewski, ‘‘Analysis [98] G. Parker, M. J. Spoelma, G. Tavella, M. Alda, D. L. Dunner,
of acoustic space variability in speech affected by depression,’’ Speech C. O’Donovan, J. K. Rybakowski, A. Bayes, V. Sharma, P. Boyce, and
Commun., vol. 75, pp. 27–49, Dec. 2015. V. Manicavasagar, ‘‘A new machine learning-derived screening measure
[79] L. He and C. Cao, ‘‘Automated depression analysis using convolu- for differentiating bipolar from unipolar mood disorders,’’ J. Affect. Dis-
tional neural networks from speech,’’ J. Biomed. Informat., vol. 83, orders, vol. 299, pp. 513–516, Feb. 2022.
pp. 103–111, Jul. 2018. [99] A. Sharma and W. J. M. I. Verbeke, ‘‘Improving diagnosis of depression
[80] H. Jiang, B. Hu, Z. Liu, G. Wang, L. Zhang, X. Li, and H. Kang, with XGBOOST machine learning model and a large biomarkers Dutch
‘‘Detecting depression using an ensemble logistic regression model based dataset (n = 11,081),’’ Frontiers Big Data, vol. 3, p. 15, Apr. 2020.
on multiple speech features,’’ Comput. Math. Methods Med., vol. 2018, [100] L. Zhou, A. W. Baughman, V. J. Lei, K. H. Lai, A. S. Navathe,
Sep. 2018, Art. no. 6508319. F. Chang, M. Sordo, M. Topaz, F. R. Zhong, M. Murrali, S. Navathe, and
[81] J. Li, X. Fu, Z. Shao, and Y. Shang, ‘‘Improvement on speech depression
R. A. Rocha, ‘‘Identifying patients with depression using free-text clini-
recognition based on deep networks,’’ in Proc. Chin. Autom. Congr.
cal documents,’’ in Proc. MEDINFO eHealth-Enabled Health (Studies in
(CAC), Nov. 2018, pp. 2705–2709.
Health Technology and Informatics), vol. 216, 2015, pp. 629–633.
[82] M. Muzammel, H. Salam, Y. Hoffmann, M. Chetouani, and A. Othmani,
[101] H. S. Cai, J. S. Han, Y. F. Chen, X. C. Sha, Z. Y. Wang, B. Hu, J. Yang,
‘‘AudVowelConsNet: A phoneme-level based deep CNN architecture
L. Feng, Z. J. Ding, Y. Q. Chen, and J. Gutknecht, ‘‘A pervasive approach
for clinical depression diagnosis,’’ Mach. Learn. with Appl., vol. 2,
to EEG-based depression detection,’’ Complexity, vol. 2018, pp. 1–13,
Dec. 2020, Art. no. 100005.
[83] Z. Zhao, Z. Bao, Z. Zhang, J. Deng, N. Cummins, H. Wang, J. Tao, 2018.
and B. Schuller, ‘‘Automatic assessment of depression from speech via a [102] H. Akbari, M. T. Sadiq, M. Payan, S. S. Esmaili, H. Baghri,
hierarchical attention transfer network and attention autoencoders,’’ IEEE and H. Bagheri, ‘‘Depression detection based on geometrical features
J. Sel. Topics Signal Process., vol. 14, no. 2, pp. 423–434, Feb. 2020. extracted from SODP shape of EEG signals and binary PSO,’’ Traitement
[84] L. Cao, S. Guo, Z. Xue, Y. Hu, H. Liu, T. E. Mwansisya, W. Pu, B. Yang, du Signal, vol. 38, no. 1, pp. 13–26, Feb. 2021.
C. Liu, J. Feng, E. Y. Chen, and Z. Liu, ‘‘Aberrant functional connectivity [103] X. Li, B. Hu, S. Sun, and H. Cai, ‘‘EEG-based mild depressive detec-
for diagnosis of major depressive disorder: A discriminant analysis,’’ tion using feature selection methods and classifiers,’’ Comput. Methods
Psychiatry Clin. Neurosci., vol. 68, no. 2, pp. 110–119, 2014. Programs Biomed., vol. 136, pp. 151–161, Nov. 2016.
[85] H. Guo, X. Cao, Z. Liu, H. Li, J. Chen, and K. Zhang, ‘‘Machine learning [104] M. Saeedi, A. Saeedi, and A. Maghsoudi, ‘‘Major depressive disorder
classifier using abnormal brain network topological metrics in major assessment via enhanced k-nearest neighbor method and EEG signals,’’
depressive disorder,’’ Neuroreport, vol. 23, no. 17, pp. 1006–1011, 2012. Phys. Eng. Sci. Med., vol. 43, no. 3, pp. 1007–1018, Sep. 2020.
[86] H. Guo, M. Qin, J. Chen, Y. Xu, and J. Xiang, ‘‘Machine-learning clas- [105] M. Ahmadlou, H. Adeli, and A. Adeli, ‘‘Fractality analysis of frontal
sifier for patients with major depressive disorder: Multifeature approach brain in major depressive disorder,’’ Int. J. Psychophysiol., vol. 85, no. 2,
based on a high-order minimum spanning tree functional brain network,’’ pp. 206–211, 2012.
Comput. Math. Methods Med., vol. 2017, Dec. 2017, Art. no. 4820935. [106] M. Čukić, M. Stokić, S. Simić, and D. Pokrajac, ‘‘The successful dis-
[87] H. Li, L. Cui, L. Cao, Y. Zhang, Y. Liu, W. Deng, and W. Zhou, ‘‘Identifi- crimination of depression from EEG could be attributed to proper feature
cation of bipolar disorder using a combination of multimodality magnetic extraction and not to a particular classification method,’’ Cogn. Neuro-
resonance imaging and machine learning techniques,’’ BMC Psychiatry, dyn., vol. 14, no. 4, pp. 443–455, Aug. 2020.
vol. 20, no. 1, pp. 1–12, Dec. 2020. [107] C. Cognition. Stop Signal Task. Accessed: Jan. 3, 2022. [Online].
[88] M. J. Rosa, L. Portugal, T. Hahn, A. J. Fallgatter, M. I. Garrido, Available: https://www.cambridgecognition.com/cantab/cognitive-
J. Shawe-Taylor, and J. Mourao-Miranda, ‘‘Sparse network-based mod- tests/memory/stop-signal-task-sst
els for patient classification using fMRI,’’ NeuroImage, vol. 105, [108] S. Healthineers. 1.5T MRI Scanners. Accessed: Jan. 3, 2022. [Online].
pp. 493–506, Jan. 2015. Available: https://www.siemens-healthineers.com/en-uk/magnetic-
[89] Y. Li, X. Dai, H. Wu, and L. Wang, ‘‘Establishment of effective biomark- resonance-imaging/0-35-to-1-5t-mri-scanner
ers for depression diagnosis with fusion of multiple resting-state connec- [109] g.NAUTICULUS. g.NAUTICULUS Research. Accessed: Mar. 1, 2020.
tivity measures,’’ Frontiers Neurosci., vol. 15, Sep. 2021, Art. no. 729958. [Online]. Available: https://www.gtec.at/product/gnautilus-research/
[90] B. Sen, K. R. Cullen, and K. K. Parhi, ‘‘Classification of adolescent major
[110] N. Bajaj, ‘‘Wavelets for EEG analysis,’’ in Wavelet Theory. IntechOpen,
depressive disorder via static and dynamic connectivity,’’ IEEE J. Biomed.
ch. 5, Feb. 2020, doi:10.5772/intechopen.94398.
Health Informat., vol. 25, no. 7, pp. 2604–2614, Jul. 2021.
[111] S. De Silva, S. U. Dayarathna, G. Ariyarathne, D. Meedeniya, and
[91] R. Wang, Y. Hao, Q. Yu, M. Chen, I. Humar, and G. Fortino, ‘‘Depression
S. Jayarathna, ‘‘FMRI feature extraction model for ADHD classification
analysis and recognition based on functional near-infrared spectroscopy,’’
using convolutional neural network,’’ Int. J. E-Health Med. Commun.,
IEEE J. Biomed. Health Informat., vol. 25, no. 12, pp. 4289–4299,
vol. 12, no. 1, pp. 81–105, Jan. 2021.
Dec. 2021.
[92] E. M. de Souza Filho, H. C. V. Rey, R. M. Frajtag, D. M. A. Cook, [112] C. Dou, S. Zhang, H. Wang, L. Sun, Y. Huang, and W. Yue, ‘‘ADHD fMRI
L. N. D. de Carvalho, A. L. P. Ribeiro, and J. Amaral, ‘‘Can machine short-time analysis method for edge computing based on multi-instance
learning be useful as a screening tool for depression in primary care?’’ learning,’’ J. Syst. Archit., vol. 111, Dec. 2020, Art. no. 101834.
J. Psychiatric Res., vol. 132, pp. 1–6, Jan. 2021. [113] Z. Mao, Y. Su, G. Xu, X. Wang, Y. Huang, W. Yue, L. Sun, and N. Xiong,
[93] Y. Liu, J. Hankey, B. Cao, and P. Chokka, ‘‘Screening for major depressive ‘‘Spatio-temporal deep learning method for ADHD fMRI classification,’’
disorder in a tertiary mental health centre using EarlyDetect: A machine Inf. Sci., vol. 499, pp. 1–11, Oct. 2019.
learning-based pilot study,’’ J. Affect. Disorders Rep., vol. 3, Jan. 2021, [114] J. Peng, M. Debnath, and A. K. Biswas, ‘‘Efficacy of novel summation-
Art. no. 100062. based synergetic artificial neural network in ADHD diagnosis,’’ Mach.
[94] Y. Ma, J. Ji, Y. Huang, H. Gao, Z. Li, W. Dong, S. Zhou, Y. Zhu, Learn. With Appl., vol. 6, Dec. 2021, Art. no. 100120.
W. Dang, T. Zhou, H. Yu, B. Yu, Y. Long, L. Liu, G. Sachs, and X. Yu, [115] L. Zou, J. Zheng, C. Miao, M. J. Mckeown, and Z. J. Wang, ‘‘3D CNN
‘‘Implementing machine learning in bipolar diagnosis in China,’’ Transl. based automatic diagnosis of attention deficit hyperactivity disorder using
Psychiatry, vol. 9, no. 1, Nov. 2019. functional and structural MRI,’’ IEEE Access, vol. 5, pp. 23626–23636,
[95] V. Mato-Abad, I. Jiménez, R. García-Vázquez, J. Aldrey, 2017.
D. Rivero, P. Cacabelos, J. Andrade-Garda, J. Pías-Peleteiro, and [116] A. Riaz, M. Asad, E. Alonso, and G. Slabaugh, ‘‘DeepFMRI: End-to-
S. Rodríguez-Yáñez, ‘‘Using artificial neural networks for identifying end deep learning for functional connectivity and classification of ADHD
patients with mild cognitive impairment associated with depression using fMRI,’’ J. Neurosci. Methods, vol. 335, Apr. 2020, Art. no. 108506.
using neuropsychological test features,’’ Appl. Sci., vol. 8, no. 9, p. 1629, [117] Z. Wang, Y. Sun, Q. Shen, and L. Cao, ‘‘Dilated 3D convolutional
Sep. 2018. neural networks for brain MRI data classification,’’ IEEE Access, vol. 7,
[96] Y. Meng, W. Speier, M. K. Ong, and C. W. Arnold, ‘‘Bidirectional repre- pp. 134388–134398, 2019.
sentation learning from transformers using multimodal electronic health [118] S. Dey, A. R. Rao, and M. Shah, ‘‘Attributed graph distance measure for
record data to predict depression,’’ IEEE J. Biomed. Health Informat., automatic detection of attention deficit hyperactive disordered subjects,’’
vol. 25, no. 8, pp. 3121–3129, Aug. 2021. Frontiers Neural Circuits, vol. 8, p. 64, Jun. 2014.
VOLUME 11, 2023 86315

[119] J. Du, L. Wang, B. Jie, and D. Zhang, ‘‘Network-based classification [139] P. P. Thoduparambil, A. Dominic, and S. M. Varghese, ‘‘EEG-based deep
of ADHD patients using discriminative subnetwork selection and graph learning model for the automatic detection of clinical depression,’’ Phys.
kernel PCA,’’ Computerized Med. Imag. Graph., vol. 52, pp. 82–88, Eng. Sci. Med., vol. 43, no. 4, pp. 1349–1360, Dec. 2020.
Sep. 2016. [140] H. W. Loh, C. P. Ooi, E. Aydemir, T. Tuncer, S. Dogan, and U. R. Acharya,
[120] N. A. Khan, S. A. Waheeb, A. Riaz, and X. Shang, ‘‘A novel knowl- ‘‘Decision support system for major depression detection using spectro-
edge distillation-based feature selection for the classification of ADHD,’’ gram and convolution neural network with EEG signals,’’ Expert Syst.,
Biomolecules, vol. 11, no. 8, p. 1093, Jul. 2021. vol. 39, no. 3, Mar. 2022, Art. no. e12773.
[121] B. Miao, L. L. Zhang, J. L. Guan, Q. F. Meng, and Y. L. Zhang, ‘‘Classi- [141] W. Mumtaz and A. Qayyum, ‘‘A deep learning framework for auto-
fication of ADHD individuals and neurotypicals using reliable RELIEF: matic diagnosis of unipolar depression,’’ Int. J. Med. Informat., vol. 132,
A resting-state study,’’ IEEE Access, vol. 7, pp. 62163–62171, 2019. Dec. 2019, Art. no. 103983.
[122] L. Shao, Y. Xu, and D. Fu, ‘‘Classification of ADHD with bi-objective [142] A. Saeedi, M. Saeedi, A. Maghsoudi, and A. Shalbaf, ‘‘Major depressive
optimization,’’ J. Biomed. Informat., vol. 84, pp. 164–170, Aug. 2018. disorder diagnosis based on effective connectivity in EEG signals: A con-
[123] X. Peng, P. Lin, T. Zhang, and J. Wang, ‘‘Extreme learning machine-based volutional neural network and long short-term memory approach,’’ Cogn.
classification of ADHD using brain structural MRI data,’’ PLoS ONE, Neurodyn., vol. 15, no. 2, pp. 239–252, Apr. 2021.
vol. 8, no. 11, Nov. 2013, Art. no. e79476. [143] A. Seal, R. Bajpai, J. Agnihotri, A. Yazidi, E. Herrera-Viedma, and
[124] P. Preetha and R. Mallika, ‘‘Normalization and deep learning based O. Krejcar, ‘‘DeprNet: A deep convolution neural network framework for
attention deficit hyperactivity disorder classification,’’ J. Intell. Fuzzy detecting depression using EEG,’’ IEEE Trans. Instrum. Meas., vol. 70,
Syst., vol. 40, no. 4, pp. 7613–7621, Apr. 2021. pp. 1–13, 2021.
[125] S. Liu, L. Zhao, J. Zhao, B. Li, and S.-H. Wang, ‘‘Attention [144] G. Sharma, A. Parashar, and A. M. Joshi, ‘‘DepHNN: A novel hybrid
deficit/hyperactivity disorder classification based on deep spatio- neural network for electroencephalogram (EEG)-based screening of
temporal features of functional magnetic resonance imaging,’’ Biomed. depression,’’ Biomed. Signal Process. Control, vol. 66, Apr. 2021,
Signal Process. Control, vol. 71, Jan. 2022, Art. no. 103239. Art. no. 102393.
[145] U. R. Acharya, V. K. Sudarshan, H. Adeli, J. Santhosh, J. E. W. Koh,
[126] T. Eslami and F. Saeed, ‘‘Similarity based classification of ADHD using
S. D. Puthankatti, and A. Adeli, ‘‘A novel depression diagnosis index
singular value decomposition,’’ in Proc. 15th ACM Int. Conf. Comput.
using nonlinear features in EEG signals,’’ Eur. Neurol., vol. 74, nos. 1–2,
Frontiers, 2018, pp. 19–25.
pp. 79–83, 2015.
[127] G. Deshpande, P. Wang, D. Rangaprakash, and B. Wilamowski, ‘‘Fully
[146] H. Akbari, M. T. Sadiq, and A. U. Rehman, ‘‘Classification of normal
connected cascade artificial neural network architecture for atten-
and depressed EEG signals based on centered correntropy of rhythms in
tion deficit hyperactivity disorder classification from functional mag-
empirical wavelet transform domain,’’ Health Inf. Sci. Syst., vol. 9, no. 1,
netic resonance imaging data,’’ IEEE Trans. Cybern., vol. 45, no. 12,
Dec. 2021.
pp. 2668–2679, Dec. 2015.
[147] E. Aydemir, T. Tuncer, S. Dogan, R. Gururajan, and U. R. Acharya,
[128] S. Ghiassian, R. Greiner, P. Jin, and M. R. G. Brown, ‘‘Using functional ‘‘Automated major depressive disorder detection using melamine pattern
or structural magnetic resonance images and personal characteristic data with EEG signals,’’ Int. J. Speech Technol., vol. 51, no. 9, pp. 6449–6466,
to identify ADHD and autism,’’ PLoS ONE, vol. 11, no. 12, Dec. 2016, Sep. 2021.
Art. no. e0166934. [148] G. M. Bairy, U. C. Niranjan, and S. D. Puthankattil, ‘‘Automated classifi-
[129] A. Anderson, P. K. Douglas, W. T. Kerr, V. S. Haynes, A. L. Yuille, J. Xie, cation of depression EEG signals using wavelet entropies and energies,’’
Y. N. Wu, J. A. Brown, and M. S. Cohen, ‘‘Non-negative matrix factor- J. Mech. Med. Biol., vol. 16, no. 3, May 2016, Art. no. 1650035.
ization of multimodal MRI, fMRI and phenotypic data reveals differential [149] L. J. Duan, H. L. Liu, H. F. Duan, Y. H. Qiao, and C. M. Wang, ‘‘Classifi-
changes in default mode subnetworks in ADHD,’’ NeuroImage, vol. 102, cation of depression based on local binary pattern and singular spectrum
pp. 207–219, Nov. 2014. analysis,’’ in Algorithms and Architectures for Parallel Processing. Cham,
[130] P. Wang, X. Zhao, J. Zhong, and Y. Zhou, ‘‘Localization and diagnosis of Switzerland: Springer, 2020.
attention-deficit/hyperactivity disorder,’’ Healthcare, vol. 9, no. 4, p. 372, [150] W. Liu, C. Zhang, X. Wang, J. Xu, Y. Chang, T. Ristaniemi, and F. Cong,
Mar. 2021. ‘‘Functional connectivity of major depression disorder using ongoing
[131] A. M. S. Aradhya, V. Subbaraju, S. Sundaram, and N. Sundararajan, EEG during music perception,’’ Clin. Neurophysiol., vol. 131, no. 10,
‘‘Discriminant spatial filtering method (DSFM) for the identification and pp. 2413–2422, Oct. 2020.
analysis of abnormal resting state brain activities,’’ Expert Syst. Appl., [151] S. Mahato, N. Goyal, D. Ram, and S. Paul, ‘‘Detection of depression and
vol. 181, Nov. 2021, Art. no. 115074. scaling of severity using six channel EEG data,’’ J. Med. Syst., vol. 44,
[132] S. Farzi, S. Kianian, and I. Rastkhadive, ‘‘Diagnosis of attention deficit no. 7, p. 118, Jul. 2020.
hyperactivity disorder using deep belief network based on greedy [152] U. Raghavendra, A. Gudigar, Y. Chakole, P. Kasula, D. P. Subha,
approach,’’ in Proc. 5th Int. Symp. Comput. Bus. Intell. (ISCBI), N. A. Kadri, E. J. Ciaccio, and U. R. Acharya, ‘‘Automated detection
Aug. 2017, pp. 96–99. and screening of depression using continuous wavelet transform with
[133] N. Qiang, Q. Dong, H. Liang, B. Ge, S. Zhang, C. Zhang, J. Gao, and electroencephalogram signals,’’ Expert Syst., vol. 40, no. 4, May 2023,
Y. Sun, ‘‘A novel ADHD classification method based on resting state tem- Art. no. e12803.
poral templates (RSTT) using spatiotemporal attention auto-encoder,’’ [153] W. Mumtaz, S. S. A. Ali, M. A. M. Yasin, and A. S. Malik, ‘‘A machine
Neural Comput. Appl., vol. 34, no. 10, pp. 7815–7833, May 2022. learning framework involving EEG-based functional connectivity to diag-
[134] U. R. Acharya, S. L. Oh, Y. Hagiwara, J. H. Tan, H. Adeli, and nose major depressive disorder (MDD),’’ Med. Biol. Eng. Comput.,
D. P. Subha, ‘‘Automated EEG-based screening of depression using deep vol. 56, no. 2, pp. 233–246, Feb. 2018.
convolutional neural network,’’ Comput. Methods Programs Biomed.,
vol. 161, pp. 103–113, Jul. 2018.
[135] D. M. Khan, N. Yahya, N. Kamel, and I. Faye, ‘‘Automated diagnosis
of major depressive disorder using brain effective connectivity and 3D
convolutional neural network,’’ IEEE Access, vol. 9, pp. 8835–8846,
2021.
[136] X. Li, R. La, Y. Wang, J. Niu, S. Zeng, S. Sun, and J. Zhu, ‘‘EEG-based
mild depression recognition using convolutional neural network,’’ Med. CHRISTIAN NASH (Student Member, IEEE)
Biol. Eng. Comput., vol. 57, no. 6, pp. 1341–1352, Jun. 2019. received the B.Eng. degree in electrical and
[137] C. Uyulan, T. T. Ergüzel, H. Unubol, M. Cebi, G. H. Sayar, M. Nezhad computer engineering from Newcastle University,
Asad, and N. Tarhan, ‘‘Major depressive disorder classification based on U.K., in 2020, where he is currently pursuing
different convolutional neural network models: Deep learning approach,’’ the Ph.D. degree with the Intelligent Sensing and
Clin. EEG Neurosci., vol. 52, no. 1, pp. 38–51, Jan. 2021. Communication Group.
[138] X. Li, X. Zhang, J. Zhu, W. Mao, S. Sun, Z. Wang, C. Xia, and B. Hu, His research interests include deep learning,
‘‘Depression recognition using machine learning methods with different medical video analysis, and mental health detec-
feature generation strategies,’’ Artif. Intell. Med., vol. 99, Aug. 2019, tion.
Art. no. 101696.
86316 VOLUME 11, 2023

RAJESH NAIR is currently a FRCPsych and a SYED MOHSEN NAQVI (Senior Member, IEEE)
Consultant Psychiatrist, leads a team with the received the Ph.D. degree from Loughborough
Adult ADHD Services, Cumbria, Northumber- University, U.K., in 2010.
land, Tyne and Wear NHS Foundation Trust, Eng- He is currently a Reader (Associate Professor)
land. He is also the U.K. Adult ADHD Network in multimodal signal and information processing,
(UKAAN) Lead of the North-East of England. the Director of the Intelligent Sensing Laboratory,
He has experience in both National Institute for and the Deputy Head of the Intelligent Sensing and
Health Research (NIHR) and industry funded Communications Research Group, Newcastle Uni-
research projects/trials e.g., digital medicines trial, versity, Newcastle, U.K. He has above 130 pub-
ketamine in electroconvulsive therapy (ECT) trial, lications in peer-reviewed articles in high impact
the lithium versus quetiapine in depression (LQD) trial with TEWV, and U.K. journals and proceedings of leading international conferences. He organized
landscape study. He has significant expertise with Adult ADHD. special sessions in FUSION (2013–2022), delivered seminars, and was a
Speaker with University Defence Research Collaboration (UDRC) Summer
Schools (2015–2017). He was involved in more than 15 research projects
funded by UKRI and Industry (e.g., EPSRC, BBSRC, MoD, UDRC, Thales,
Innovate U.K., and NHS). He also successfully supervised and graduated
above 20 Ph.D. His research interests include human action, activity, behav-
ior analyses, multiple human target detection, localization, tracking, human
speech enhancement and separation, and explainable AI, all for defence and
healthcare applications.
Dr. Naqvi is a fellow of the Higher Education Academy. He was an
Associate Editor of Journal of Signal Processing (Elsevier) (2018–2022).
He served two terms of an Associate Editor for IEEE TRANSACTIONS ON SIGNAL
PROCESSING (2019–2023). He has been an Associate Editor of IEEE/ACM
TRANSACTIONS ON AUDIO SPEECH AND LANGUAGE PROCESSING, since 2019.
VOLUME 11, 2023 86317

Machine Learning in ADHD and Depression Mental Health Diagnosis A Survey

Uploaded by

Copyright:

Available Formats

Machine Learning in ADHD and Depression Mental Health Diagnosis A Survey

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Machine Learning in ADHD and Depression Mental Health Diagnosis A Survey

Uploaded by

Copyright:

Available Formats

Received 11 July 2023, accepted 31 July 2023, date of publication 10 August 2023, date of current version 17 August 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3304236

Machine Learning in ADHD and Depression

I. INTRODUCTION due to overlapping symptoms and the potential side effects of

86298 VOLUME 11, 2023

VOLUME 11, 2023 86299

such as epidemiology, neurology, paediatrics, primary care,

As both the DSM-V and ICD-11 are very similar in nature,

86300 VOLUME 11, 2023

rating scales like Conners-3 [14], offering insights into an

VOLUME 11, 2023 86301

TABLE 2. Interpretation of the total scores when evaluating depression

FIGURE 4. An advanced wearable EEG head cap [109]. The headcaps

86302 VOLUME 11, 2023

VOLUME 11, 2023 86303

for depression evaluate their machine learning models on

86304 VOLUME 11, 2023

VOLUME 11, 2023 86305

86306 VOLUME 11, 2023

VOLUME 11, 2023 86307

86308 VOLUME 11, 2023

VOLUME 11, 2023 86309

VOLUME 11, 2023

VOLUME 11, 2023

86312 VOLUME 11, 2023

VOLUME 11, 2023 86313

86314 VOLUME 11, 2023

VOLUME 11, 2023 86315

86316 VOLUME 11, 2023

VOLUME 11, 2023 86317

You might also like