EEG based Major Depressive disorder and Bipolar disorder detection using Neural Networks: A review
Sana Yasin, Syed Asad Hussain, Sinem Aslan, Imran Raza, Muhammad Muzammel, Alice Othmani
HHighlights
Neural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosisusing EEG signals: A review
Sana Yasin,Syed Asad Hussain,Sinem Aslan,Imran Raza,Muhammad Muzammel,Alice Othmani• A summary of clinical background of MDD and BD, Brain structure and EEG based brain activity,• EEG experimental protocols study for mental disorders diagnosis,• An extensive study of state-of-the-art shallow and deep neural networks models for EEG based clinical depressiondetection,• An extensive study of neural networks based approaches for Biploar Disorder diagnosis using EEG signals,• A discussion of state-of-the-art methods limitations and giving valuable recommendations for future research a r X i v : . [ q - b i o . N C ] S e p eural Networks based approaches for Major Depressive Disorder andBipolar Disorder Diagnosis using EEG signals: A review Sana Yasin a,b , Syed Asad Hussain a , Sinem Aslan c,d , Imran Raza a , Muhammad Muzammel e andAlice Othmani e , ∗ a Department of Computer Science, COMSATS University Islamabad, Lahore Campus Lahore,Pakistan b Department of Computer Science and Information Technology, Superior University Gold Campus, Lahore Pakistan c CaâĂŹ Foscari University of Venice, DAIS & ECLT, Venice, Italy d Ege University, International Computer Institute, Izmir, Turkey e UniversitÃľ Paris-Est CrÃľteil (UPEC), LISSI, Vitry sur Seine 94400, France
A R T I C L E I N F O
Keywords :Electroencephalogram(EEG)Major Depressive disorder(MDD)Bipolar disorder(BD)Artificial neural networksbiomedical informatics
A B S T R A C T
Mental disorders represent critical public health challenges as they are leading contributors to theglobal burden of disease and intensely influence social and financial welfare of individuals. The presentcomprehensive review concentrate on the two mental disorders: Major depressive Disorder (MDD)and Bipolar Disorder (BD) with noteworthy publications during the last ten years. There’s a bigneed nowadays for phenotypic characterization of psychiatric disorders with biomarkers. Electroen-cephalography (EEG) signals could offer a rich signature for MDD and BD and then they could improveunderstanding of pathophysiological mechanisms underling these mental disorders. In this work, wefocus on the literature works adopting neural networks fed by EEG signals. Among those studies usingEEG and neural networks, we have discussed a variety of EEG based protocols, biomarkers and publicdatasets for depression and bipolar disorder detection. We conclude with a discussion and valuablerecommendations that will help to improve the reliability of developed models and for more accurateand more deterministic computational intelligence based systems in psychiatry. This review will proveto be a structured and valuable initial point for the researchers working on depression and bipolardisorders recognition by using EEG signals.
1. Introduction
Mental disorders represent critical public health chal-lenges because they are a leading contributors to the globalburden of disease and profoundly impact social and economicwelfare of the people. The World Health Organization pre-dicted that by the year 2020, mental disorders shall be thefirst cause of disability worldwide . According to WHO, itis predicted that over 264 million women and men of all agesbears some form of mental disorder, indicating that mentalhealth problems effect up to 27% of the general population atsome point during their lives . The total cost of mental, neu-rological, and substance use (MNS) disorders in the unitedstaes is over 210 billion, mostly related to indirect costs [1].MNS disorders, therefore, account for 35% of the overallburden of illness and are more costly than the combined bur-den of diabetes and cancers [2]. In contrast to other fieldsof medicine, psychiatry is still plagued by two problems:(1) a classification of mental disorders based on clinicalsymptoms of overlapping nosographic entities rather than oncausal factors; (2) a pharmacologic arsenal that only targetsclinical symptoms, mostly in an incomplete manner in a ma- ∗ Corresponding author: Associate professor. Dr. Alice OTHMANI [email protected] (A. Othmani)
ORCID (s): International Classification of Diseases or the Diagnostic and Sta-tistical Manual of Mental Disorders (DSM): jority of patients. Moreover, mapping diagnostic labels froma clinically defined nosology to varying biological indiceshas proven to be problematic [3].Therefore, we need MentalHealth Innovation and new ways to diagnose mental diseasesby finding new biomarkers. Artificial Intelligence (AI) canplay a key role in the psychiatry revolution. Multimodal Arti-ficial Intelligence-based approaches and technologies need tobe developed in order to advance our understanding and careof mental health, improve early precise diagnosis and prog-nosis, develop innovative treatments, and develop assistivetechnologies for longitudinal follow-up of the patient.Recently, automatic recognition of mental states and men-tal disorders has attracted considerable attention from com-puter vision and artificial intelligence community. While, ear-lier works mostly adopted Functional Magnetic ResonanceImaging (fMRI) [4], visual cues [5], self-rating scale [6] andsocial network analysis [7]. In recent years non-invasive sen-sors based devices, such as Electroencephalogram (EEG),have been widely employed in the literature. One of the mostremarkable research efforts has been made on developingefficient neural network-based approaches for EEG signalsanalysis for automatic assessment of mental disorders suchas Major Depressive Disorder (MDD) or Bipolar Disorder(BD).Electroencephalogram (EEG) is a non-invasive, effective,and powerful tool for recording the electrical activity of thebrain and for the diagnosis of various mental disorders suchas MDD [8], BD [9], anxiety [10], schizophrenia [11], andsleep disorders [12]. Due to these mental disorders or anoma-lies specifically depression and bipolar disorder, the body
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 1 of 29 eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review releases cortisol to the brain which affects the neurons pro-duction and communication and consequently slowing downthe functionality of some parts of brain and changing theelectrical activities patterns. The resulting voltage variationsfrom ionic current flows within the neurons of the brain couldhelp in the diagnosis of mental disorders like depression andbipolar disorder [13]. The development of robust approachesfor brain signals analysis is a challenging task because ofthe complexity, the unstructured nature of signals and thebig variability related to the person, to its age and its mentalhealth. Moreover, brain signals are frequently affected bydifferent types of noise due to eye blinking and muscularactivities during EEG recording [14].Major Depressive Disorder and Bipolar Disorder are men-tal disorders that affect physical health, sleep, appetite, at-tention level and they can lead to suicidal thoughts or ac-tions. Both disorders have many similarities, but they alsohave some crucial differences. Patients with Bipolar disorderexperience high mood swings that include excessive highs(mania or hypomania) and lows (depressive episodes). Inmania episode, patients feel overactive, full of energy andusually irritable. While they feel opposite during depressiveepisodes with a strong feeling of hopeless and loss of pleasurein most of activities. Further, in depressive episodes they donot experience any elevated and extreme feelings that maniaor hypomania patient face [15].The diagnosis of bipolar disorder is not always easy be-cause many of the symptoms overlap with depression [16].Bipolar patients most probably consult their doctor for thefirst time when they have a depressive episode, instead ofduring a manic or hypo-manic episode. Due to this reason,clinicians frequently misdiagnose bipolar disorder as depres-sion [16].In this paper, we present an exhaustive review of exist-ing neural networks-based approaches, i.e., both shallow anddeep architectures, for
Major Depressive Disorder (MDD)and
Bipolar Disorder (BD) diagnosis using EEG signals.With this work, we aim to provide a head start to the re-searchers with an up-to-date survey of advances of EEGbased depression and bipolar disorders detection techniquesfor further contributions in this field. To the best of ourknowledge, only a few surveys related to depression and EEGsignals analysis, i.e., [17, 8, 18, 19, 6, 20], have been pub-lished on depression diagnosis based on machine learning[17], computer-aided diagnosis (CADx) [8], speech analy-sis techniques[18], nutrition examination [19], self-ratingscale [6] and by smartphone application [20]. Some of thesurveys [21, 22, 23, 24, 25] cover the literature works ondeep learning-based approaches using EEG signals. How-ever, these techniques do not focus on depression diagnosisrather they consider other cognitive tasks [23, 24], such asclassification of brain signals [22], motor imagery [25], andBrain Computer Interfaces (BCI) [21]. To the best of ourknowledge, this paper is the first comprehensive survey onneural network approaches that adopt either shallow or deepneural networks and using EEG signals for MDD and BDdetection.
In the past studies, deep learning and mental disordersfields have been studied separately. Just a few years ago,crossovers between these two areas have been merged andresearchers have used deep learning for EEG-based mentaldisorders detection. Table 1 shows the existing surveysrelated to deep learning, Electroencephalogram (EEG) andmental disorders. To the best of our knowledge, this review isthe first comprehensive study of the latest improvements andfront lines of deep learning and artificial neural networks forMDD and BD recognition. In this survey, we have consideredpapers, most of which has been published in the last five years(since 2015). The contributions of this survey are as follows:1. A systematic review of artificial neural networks in-cluding shallow and deep learning-based approaches todetect Major depressive disorder (MDD) and Bipolardisorder using EEG, is presented to provide researchersan extensive understanding of this area of research.2. Discussion on standard deep learning techniques andstate-of-the-art models for MDD and BD detection, andproviding some guidelines for choosing the suitabledeep learning models.3. A review of applications and challenges of deep learn-ing and ANN-based MDD and BD detection. It alsohighlights some fascinating topics for future research.
The current review is organized in nine sections. Sec-tion2 covers search strategy and eligibility criteria. Section3describes the Electroencephalogram and brain structure thatfurther subdivide into Electroencephalogram background,brain structure and effect of depression and bipolar disor-der on brain structure. The section 4 elaborates the clinicalbackground of depression and summarizes the assessmentof clinical depression by verbal and nonverbal signs. Sec-tion 5 reviews the neural networks-based approaches, EEGexperimental protocols and public data sets for depressionrecognition, while Section 6 presents a clinical backgroundof bipolar disorder by verbal and nonverbal signs. Section7 encloses the information about the neural networks-basedapproaches and EEG experimental protocols for bipolar dis-order recognition. Discussion of the current review findingsand suggestions for future studies are comprised in Section8. At the end conclusion section revealed the potential ofshallow and deep neural network for depression and bipolardisorder recognition in section 9.
2. Search Strategy and Eligibility Criteria
We have searched IEEEXplore, PubMed, Embase, Springer,ScienceDirect and Web of Science, for articles publishedbetween January 2010, and May 2020 by using the follow-ing keywords: (âĂIJDepressionâĂİ OR âĂIJshallow neuralnetworkâĂİ OR âĂIJDeep learningâĂİ OR âĂIJElectroen-cephalogramâĂİ OR âĂIJCross-validationâĂİ OR âĂIJBipo-lar depressionâĂİ OR âĂIJArtificial neural networkâĂİ OR
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 2 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 1: Summary of existing surveys and reviews related to DL, EEG,MDD and BD where DL stands for deep learning,EEG stands for Electroencephalogram, MDD stands for Major depressive disorder and BD stands for Bipolar disorder.Publication Years Topic ScopeDL EEG MDD BD[26] 2017 A survey on deep learning ✓ x x x[27] 2018 A survey on deep learning ✓ x x x[28] 2019 A survey on deep learning ✓ x x x[29] 2018 A survey on deep learning ✓ x x x[30] 2019 A survey on deep learning ✓ x x x[31] 2013 A survey on EEG x ✓ x x[32] 2020 A survey on EEG x ✓ x x[33] 2020 A survey on EEG x ✓ x x[34] 2019 A survey on EEG x ✓ x x[35] 2020 A survey on EEG x ✓ x x[36] 2014 A survey on depression x x ✓ x[37] 2017 A survey on depression x x ✓ x[7] 2020 A survey on depression x x ✓ x[38] 2020 A survey on depression x x ✓ x[3] 2019 A survey on depression x x ✓ x[21] 2019 A survey on deep learningand EEG ✓ ✓ x x[22] 2018 A survey on deep learningand EEG ✓ ✓ x x[23] 2018 A survey on deep learningand EEG ✓ ✓ x x[24] 2019 A survey on deep learningand EEG ✓ ✓ x x[25] 2020 A survey on deep learningand EEG ✓ ✓ x x[39] 2020 A survey on bipolar disorder x x x ✓ [40] 2020 A survey on depression andbipolar disorder x x ✓ ✓ [41] 2009 A survey on EEG based bipo-lar disorder x ✓ x ✓ Our Work 2019 Survey on EEG based MDDand BD detection using neu-ral network ✓ ✓ ✓ ✓ âĂIJuni polar depressionâĂİ OR âĂIJEEG base depressionâĂİOR âĂIJbipolar depressionâĂİ OR âĂIJMajor depressive dis-orderâĂİ OR âĂIJâĂİ OR âĂIJRecurrent neural networkâĂİOR âĂIJDeep neural network OR âĂIJBDI-IIâĂİ OR âĂIJDSM-IVâĂİ OR âĂIJPHQ-9âĂİ OR âĂIJPersistent depressive dis-orderâĂİ OR âĂIJDeep Learning ModelsâĂİ OR âĂIJEEGbio-markersâĂİ OR âĂIJDeep Feature ExtractionâĂİ ORâĂIJFFNN OR Feed forward neural networkâĂİ OR âĂIJF-BNN OR Feed backword neural networkâĂİ OR âĂIJDis-criminatives Deep learning modelsâĂİ OR âĂIJRepresen-tative Deep learning ModelsâĂİ OR âĂIJGenerative DeepLearning ModelsâĂİ OR âĂIJHybrid Deep learning Mod-elsâĂİ OR âĂIJDepression diagnosis’s techniquesâĂİ ORâĂIJConvolutional neural network OR CNNâĂİ OR âĂIJMildDepressionâĂİ OR âĂIJMLPâĂİ OR âĂIJMental StatesâĂİOR âĂIJ EEG ArtifactsâĂİ OR âĂIJBipolar DepressionâĂİ OR âĂIJManic DisorderâĂİ OR âĂIJDeep Belief Network-sâĂİ OR âĂIJLSTM OR Long Short Term Memory. Wealso explored the articles that cite the ones that we found bythe key-words mentioned above. There were no languagerestrictions.We performed this systematic review by conforming tothe PRISMA statement [42] that helps to improve the re-porting of systematic reviews and meta-analyses. Eligibilitycriteria of this review includes the suitable depictions of dif-ferent shallow and deep neural network techniques for theautomatic assessment of depression by using EEG and rep-resentation of scientifically acquired data and generation ofreal-time results. Different technical reports and procedures[43], [44] of systematic reviews are followed to completethis survey. Publication dates of the EEG based depressionstudies meeting the criteria of ten years is considered. Only
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 3 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 2: Keywords and web resources used in the current review.Web Resources Theme KeywordsElsevier Major Depressive DisorderSpringer MDDGoogle Scholar Depression Clinical DepressionMayo Clinic Unipolar disorderPubMed Mild DepressionScopus BDACM Digital Library ManiaAmerican Psychiatric Associa-tion Bipolar Disorder Manic depressionEmbase HypomaniaNASA Manic-depressive illnessWeb of Science Shallow and deep neural networkIEEE Xplore Digital Library FFNN or Feed forward neural networkWorld Health Organization Deep Learning FBNN or Feed backword neural net-workCNN or Convolutional neural networkRNN or Recurrent neural networkEEG based depressionElectroencephalogram EEG-based unipolar depressionEEG based bipolar disorderEEG biomarkersthose subjects are included in this survey that have more than13 depression severity scores and no prior history of drugand medication. The main aim of this review is to sum up allthe MDD and BD detection techniques that are performedby the shallow and neural network. The Inclusion criteria forthe current review encloses 1) adequate depiction of EEGand neural networks based automatic depression assessmentclassifiers and 2) demonstration of scientifically derived data,generating concrete and accurate results. Strictly clinical stud-ies, as well as depression detection approaches relying onother techniques than EEG,are not included. The keywordsused to search electronic databases and related resources arelisted in Table 2. These keywords were used interchangeably,in combinations of two or more, with either âĂIJORâĂİ orâĂIJANDâĂİ operands.
3. Electroencephalogram (EEG) and BrainStructure:
Electroencephalogram (EEG) is an electro-biological mea-surement method that records the electrical activity of thebrain signals that are highly random and encloses valuableinformation about the brain parts [45]. It is extensively usedby physicians and researchers to study brain functions and todetect neurological syndromes. Among different depressiondetection techniques (like audio [46], facial [47], text [48]and MRI [4]) EEG method is the most reliable due to its easeof use (i.e., it requires a simple placement of electrodes) andits high temporal resolution. Due to the reliability factor ofEEG, it is also utilized in EEG Bio metrics for Person Verifi- cation [49]. Moreover, it records brainâĂŹs electrical activityover sometime by providing non-invasive and cost-effectivesolutions [50].Depression and bipolar anomalies usually indicate dys-function in the human brain. Abnormal shape of EEG signalsappears as variations in the signalsâĂŹ patterns for particularstates of the patient and EEG reacts to the biotic activitiesof the brain, for the accurate detection of the brain abnor-malities [50]. Both normal and depressed EEG signals aredisordered and composite in nature with refined differencesreflecting different brain activities of the depressive and nor-mal groups that cannot be formulated easily through visualinterpretations.
The first human EEG recording was performed in 1924by Hans Berger, who was a neuro psychiatrist at the Univer-sity of Jena Germany. He gives the German name âĂIJelek-trenkephalogrammâĂİ to EEG device that represents thegraphical representation of the electric currents generated inthe brain. Further, he presents that currents in the electricalpulses of the brain changes with respect to the functionalstatus of the brain, such as, sleep, epilepsy and anesthesia.This idea of Hans Berger revolutionize the medical field andthus helped to open a new branch of medical science calledneurophysiology. EEG signals can be divided into five dif-ferent categories based on their bandwidth i.e., alpha and beta , theta , delta and gamma , as illustrated in Fig. 1a. Alphaand beta waves can be used to represent conscious states;while, theta and delta waves are mostly used to represent Sana Yasin et al.:
Preprint submitted to Elsevier
Page 4 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 3: Electroencephalogram (EEG) bands and their characteristics.EEG Waves FrequencyRanges(Hz) Brain States Mostly foundedDelta waves 0.1-3Hz Unconscious/Sleeping Newborns and deepsleep phasesTheta Waves 4-8Hz Imagination drowsiness and sleepAlpha waves 8-13 Hz Relaxed/Conscious normal and relaxedsubjectsBeta waves 13-30Hz Conscious/Focused/Problem solving attentive or nervoussubjectsGamma waves 30-40Hz Conscious per-ception/ Peakperformance Attentive subjectsunconscious states [51]. Gamma rhythm has been attributedto sensory perception [52]. The frequency ranges and majorcharacteristics of brain waves are given in Table 3.EEG signals can be recorded using invasive and non-invasive approaches. In an invasive approach, EEG electrodesare implanted in the brain by surgery or can be implantedat forehead sites. This procedure is invasive. Next to it, themanufacturing costs of these invasive electrodes are veryhigh. While, the non-invasive electrodes are positioned onthe scalp [50]. In a non-invasive approach, EEG signalscan be recorded by using two types of electrodes, i.e., 1)wet electrodes and 2) dry electrodes. Wet electrodes areoften made of silver chloride (AgCl) and use a gel to createa conductive path between the electrodes and the skin byreducing the impedance value. The gel leakage can cause ashort circuit between different electrodes. Furthermore, theextensive use of gel can cause allergy or any other infection.Therefore, non-invasive dry electrodes are also proposed tomeasure the EEG signals. These electrodes do not require gelor any skin penetration. Also, these electrodes work perfectlyeven on the hairy sites [53, 54, 55, 56, 57].
The Human brain is an incredible part of the body thatcontrols all the body’s functions and interprets the informa-tion from the outside world. It is composed of cerebrum , cerebellum and brainstem that is enclosed within the skull.The cerebrum is a major part of the brain. It accomplishescomplex functions like inferring touch, visualization, hear-ing, speaking, cognitive, sensations, learning, and adequatecontrol of movement. The cerebrum is made of left and righthemispheres, that controls the opposite side of the body andhave distinctive fissures. Each hemisphere has four lobes:frontal, temporal, parietal, and occipital that are illustratedin Fig. 1b.Each lobe represents different information of the humanbrain, i.e., frontal lobe controls the consciousness, the tempo-ral lobe is responsible for the computation of complex stimuliand the senses of smell and sound, the parietal lobe repre-sents the sensual information and the management of objects while the occipital lobe provides information about the senseof sight. To extract the information about the depressiveand non depressive subjects, EEG electrodes are placed atdifferent lobes (frontal, temporal, parietal, and occipital) ofthe cortex, as can be seen in Fig. 1c. The placement of theelectrodes at the scalp is important, because different lobes ofthe cerebral cortex are responsible for giving the informationof electrical activities of the brain by mono polar and bipolarrecordings [57]. Mono polar recording extracts the voltagevariance among a reference electrode on the ear lobe and anelectrode on the scalp. In contrast, Bipolar electrodes gathersthe voltage variance among two scalp electrodes. Depression and Bipolar Disorder affects three portions ofthe brain: hippocampus (resides in the temporal lobe of thebrain), prefrontal cortex (located at the front of the frontallobe) and amygdala (the frontal portion of the temporal lobe)[61]. The hippocampus holds memories and controls theproduction of a hormone called cortisol. During depressionbody releases cortisol that becomes problematic when its ex-cessive amount is released and sent to the brain. People withMDD, face long-term exposure of increased cortisol levelswhich can slow down the production of new neuron. It alsocauses the neurons in the hippocampus to shrink, hence leadsto memory problems. The prefrontal cortex that resides inthe frontal lobe is responsible for controlling emotions, mak-ing decisions, and creating memories. When the amount ofcortisol exceeds in the brain, the prefrontal cortex gets shrink.The amygdala exists in the frontal portion of the temporallobe and it enables emotional responses. In depression andbipolar disorder patients, the amygdala becomes large andmore vigorous due to the continuous exposure of a high ratioof cortisol. An enlarged and manic amygdala, with irregularactivity in other portions of the brain, can consequence insleep disorders and activity patterns. Usually, cortisol levelsincrease in the morning and reduce at night. However, inMDD patients cortisol ratio is always higher even at night.A Literature survey at EEG based depression and bipolar
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 5 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review (a) EEG waves (Adapted from[58]).(b) Representation of Brain Lobes (Adapted from[59]).(c) Birds eye view of electrode placement (Adaptedfrom [60]).
Figure 1: a)Brain structure and EEG waves representation: b)Different brain lobes and their location; b) The placement ofEEG electrodes at four brain lobes using 10-20 internationalsystem. disorder detection represented in Table 9 of Section 3 showsthat the frontal lobe is more effected by depression comparedto the other brain lobes.
4. Clinical Background of Depression
Depression is a leading source of disability worldwideand significantly contributes to the global burden of disease.According to the World Health Organization (WHO) , de-pression is a common psychiatric disorder characterized bya persistent undesirable effect like sadness, lack of atten-tion, or pleasure in formerly satisfying or enjoyable activities.While the symptoms of the depression can be psychological(e.g., feeling hopeless, having continuous sadness, feelingsof guilt and low self- esteem, loss of interest, having suicidalthoughts, and so on), the link between the depression andphysical symptoms (e.g., headache, constipation, limb pain,stomach problems, pain at joints, back pain, tiredness, ap-petite and weight changes, sleep changes, and so on) has alsobeen reported in the literature [62, 63]. The causes of de-pression comprise complex relations among psychosomatic,social, and biotic factors. Life moments such as childhoodand teenage adversity, death, major events like losing a job,genetics, and substance abuse may increase the chance of de-pression. Pharmacological and psychosomatic treatments areavailable for moderate and severe depressive disorder. How-ever, in low- and middle-income states, treatment and carefacilities for depression are frequently absent or undersized[64]. WHO reports that approximately 76âĂŞ85% of peo-ple in such countries have a lack of access to the depressivetreatment they need.According to Diagnostic and Statistical Manual of Men-tal Disorders (DSM-5) of the American Psychiatric Asso-ciation (APA) [65], depression exists in various forms likeMajor Depressive Disorder (MDD), Disruptive Mood Dys-regulation Disorder (DMDD), Persistent Depressive Disor-der (Dysthymia), Premenstrual Dysphoric Disorder (PDD),Substance/Medication-Induced Depressive Disorder (S/M-IDD), Depressive Disorder Due to Another Medical Condi-tion (DDDAMC), and Other Specified Depressive Disorder(OSDD) or Unspecified Depressive Disorder (UDD).MDD, a.k.a. clinical depression , is treated as the mosttypical form of the disease, and it is diagnosed by the exis-tence of at least four of the following symptoms present forlonger than two weeks [65, 66]: changes in weight, changesin sleep, loss of energy almost every day, feelings of guiltand worthlessness, psychomotor agitation nearly every day,difficulty in concentrating, recurrent thoughts of death andsuicide. In [67], the etiology of MDD is associated with thegenetic, biological, hormonal, immunological, neurological,environmental factors, acute life events and neuroendocrino-logical mechanisms. While other types of depression havecommon symptoms with MDD, they mainly distinguish fromMDD in a number of attributes. For example, DMDD is nonepisodic and occurs in children and teenagers. It is diagnosedby an obstinately ill-tempered, annoyed mood, and recur-rent temper bursts. Dysthymia is an incessant and chronicform of depression. PDD causes severe irritability and ner-vousness in the week or two days before the period starts[68]. Substance/Medication-Induced Depressive Disorder Sana Yasin et al.:
Preprint submitted to Elsevier
Page 6 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review (S/M-IDD) is a persistent form of MDD and occurs during orafter substance intoxication. Depressive Disorder Due to An-other Medical Condition (DDDAMC) is caused by long-termillnesses and ongoing pain in the body.In the following subsections we will be focusing on diag-nosis and assessment of MDD, a.k.a clinical depression.
Clinical Depression or MDD assessment includes an ex-tensive checkup of the patient,including examination of themental state by discovering functional, relational, societalissues, and psychiatric history. Previous incidences of depres-sion or mood rise,reaction to past treatment, and comorbidmental health disorders are examined. Regarding the leveland effect of functional loss, or disability, clinical depressionis assessed in mild, moderate or severe levels. Based on thisassessment, the DSM-IV (Diagnostic and Statistical Manualof Mental Disorders) principles are used to make a diagnosis[69], and the treatment is initiated. The benchmarks for diag-nosing depression syndromes and their clinical consequencesare set by the Diagnostic and Statistical Manual of MentalDisorders (DSM-5) [70]. In addition to a questionnaire-basedapproach where verbal signs are considered in depression-assessment, another followed approach is using biomarkersin which the biological process of the human body that isongoing or has happened used as depression indication.
In this section, we highlight existing self-reported questioners-based depression detection techniques and their limitations.These questioners are designed for individuals and composedof questions to capture the signs of depression such as desper-ation, irritability, feeling guilty, and physical indications suchas tiredness, weight loss, and lack of attention in sex, as well.Many instruments have been used by clinicians for depressionassessment. Among these, we chose widely-accepted ones:PHQ-9, BDI, DSM-IV, CES-D, HAM-D.
The Patient Health Questionnaire (PHQ-9) is a multiple-choice self-report inventory and used to monitor the severityof depression [71]. It consists of nine items, each scoredfrom 0 to 3, which results with a total score varying from 0to 27. Based on the score range where the total score of thequestioners fall into, subjects are categorized into differentdepression levels like Minimal, Mild, Moderate, Moderatesevere, and Severe depression. Scoring system provided byPHQ-9 presented in Table 4. In [72, 60, 73], PHQ-9 is usedfor the participants’ selection for EEG based experiments.
Back Depression Inventory (BDI) [74], which is first pub-lished in 1961, is another questionnaire-based approach com-monly used as a valuation tool by healthcare specialists andscientists to diagnose depression and anxiety. It has three ver-sions BDI, BDI-1A and BDI-II. The first BDI was publishedin 1961 with 21 multiple choice questions against four pos-sible responses of the patients. BDI-1A, which is a revisedversion of BDI, was published in 1970. While BDI-1A pro-vided ease of use, it still had some drawbacks, such as it onlyaddress 6 criteria of DSM-III out of 9. At the latest version, BDI-II, which was introduced in 1996, four items that areBody Image Change, Weight Loss, Somatic Preoccupation,and Work Difficulty, are changed by Agitation, Worthless-ness, Concentration Difficulty, and Loss of Energy. Based onBDI-II, depression is classified into Minimal, Mild, Moder-ate, and Severe depression by the scoring system presented inTable 5. In [75, 56, 57], BDI is used as a psychometric test,i.e., every question has at least four possible responses, elon-gating in strength. A total high score shows the symptoms ofsevere depression.
Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) [76], which is published in 1994 by APA (Ameri-can Psychiatric Association), has been used for the diagnosisof It offers a consistent classification system for the identifi-cation of mental health disorders for both adults and children.[55, 77, 78], used DSM-IV to monitor perinatal depressionpatients for an EEG study. The updated version of DSM-IVwas introduced in May 2013 with the title DSM-5 insteadof DSM-V due to the limitation of the roman numerals [65].The new version includes some essential changes from theprevious one, such as replacing the roman numbers to Ara-bic numbers, excluding Asperger’s disorder and includingdisruptive mood dysregulation disorder.The
Center for Epidemiologic Studies Depression Scale (CES-D) was initially designed for the general population.Since from last few years, it has been used for the screening ofdepression patients in primary care centers [79]. The CES-Dcontains 20 self-report questions, scored on a 4-point scale,which evaluate the type and level of depression experiencedin the last few weeks. The CES-D can be applied to peopleof all ages. It has been verified across gender and culturalpopulations and achieve constant reliability and validity [80].The
Hamilton Depression Rating Scale (HAM-D) [81] isa depression assessment tool that consists of 17 items that areused for scoring. The HAM-D emphasis on the wakefulness,desperateness, self-destructive thoughts, suicidal thoughts,and actions. It is chiefly used to diagnose the depressionrecovery in individuals before, during and after the treatment.Questionnaire-based assessment approaches have severallimitations. First, they are prone to professional’s and pa-tient’s subjectivity, which hinders the objectivity of the pro-cess. Second, although depression presents different symp-toms [82] and has high co-morbidity, especially with anxi-ety [83], self questionnaire-based approaches are unable toevaluate differences across patient subgroups. Third, it isnot capable of excluding participants already diagnosed ashaving or being treated for depression. Forth, it frequentlyperforms a false diagnosis of bipolar disorder as a clinicaldepression [84].
A biomarker is defined as âĂIJa characteristic that isobjectively measured and evaluated as an indication of normalbiologic processes, pathogenic processes, or pharmacologicalresponses to a therapeutic interventionâĂİ [87]. The MajorDepressive Disorder (MDD) diagnosis and treatment can
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 7 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 4: PHQ-9 scoring standard for depression detection([85]).Depression Score Depressive Severity1-4 Minimal Depression5-9 Mild Depression10-14 Moderate Depression15-19 Moderately Severe Depression20-27 Severe DepressionTable 5: Back Depression Inventory BDI-II Scoring Standardfor depression detection ([86]).Depression Score Depression Levels0âĂŞ13 minimal depression14âĂŞ19 mild depression20-28 moderate depression29-63 Severe Depressionbe enriched by decreasing the useless treatment trials byhaving accurate predictive bio-markers. Biomarkers-basedassessment provides accurate predictions of depression andbipolar disorder and is capable of adequately measuring thechanges in disease conditions.The scientific community uses different invasive and non-invasive tools and techniques as biomarkers to understandthe mechanisms behind the MDD.The Anterior insula metabolism, the Hippocampal vol-ume, and the Subcallosal cingulate cortex metabolism are neuroimaging-based biomarkers [88], and they are used asTreatment Selection Biomarker (TSB) for major depressivedisorder. The limitation is that, these biomarkers may not bebeneficial for long-term depression treatment selection, andmay fail in severe depressive conditions. Literature worksreport that genetic factors contribute to developing depres-sion [89]. For example, it is shown in [90] that the gene3p25-26 was found in more than 800 families with recurrentdepression. [91] presents a review of EEG (Electroencephalo-grams) and ERP (Event-Related Potentials) based predictivebiomarkers for MDD. This study highlights 1) Alpha powerand asymmetry, 2) Theta band activations, 3) Antidepressanttreatment response (ATR) index, 4) Theta QEEG cordance,5) Referenced EEG (rEEG), 6) Rostral anterior cingulatecortex (rACC) activations and 7) machine learning as EEGbased predictive biomarkers 8) The P300 [92] and LDAEP[93] as ERP biomarkers. Wake and sleep EEG is also usedas a depression biomarker in [94] that provide an overviewof sleep variations in depression. It is reported in [95] that60âĂŞ90% of MDD patients with high severity suffers fromsleep disorders.In the past studies, while Alpha and Theta bands havebeen found to give information about the depression diag-nosis and recovery [96, 97], Gamma band was not well rec-ognized in depression diagnosis [98]. On the other hand, in[99] Gamma waves are declared as depression diagnostic bio-marker by presenting some significant findings on gammapulses. In comparison to Alpha, Beta and Theta waves, Gamma waves have some distinct attributes: 1) Gammapulses can accurately differentiate patients with major depres-sive disorder from healthy controls, under certain disorders,2) Gamma waves can discriminate uni polar disorder frombipolar, 3) several pharmacological and no pharmacologicaltreatments that counter depression also affect gamma. [100]adopts a variety of EEG features as depression diagnosis bio-markers extracted by linear and nonlinear methods and PhaseLagging Index (PLI) at the resting state of the patient.[97] used the frontal Theta asymmetry as a depressionbiomarker. More specifically, the EEG signals of 23 subjectswith MDD and 23 are recorded while they were listening tomusic. Results show that frontal Theta power and frontalTheta asymmetry increased significantly in healthy subjectsand decreased in depressed patients.[101] proposed to use multi-modal bio-markers (combi-nation of executive dysfunctions, motor activity, neurophysio-logical patterns) for MDD diagnosis since depressive disorderaffects not only mood but also psychomotor and cognitivefunctions. 20 MDD and 20 healthy subjects are selected. Itis shown that the multi modal bio-markers are more consis-tent in the identification of MDD patients than the unimodalbiomarkers are. [102] used brainwaves as a potential bio-marker for risk analysis of MDD. These brain waves are notonly helpful in the depression diagnosis but it also providesthe basic foundation for the accurate and reliable treatmentof depression.The EEG-based depressive bio-markers have several ad-vantages over neuroimaging techniques [50]. While, theneuroimaging techniques are less effective and unable to pro-vide information about the treatment, EEG bio-markers areeasy to use, non-invasive, have high temporal resolution, cost-effective and provide optimal treatment selection. Despitelots of advantages of EEG-based bio markers, they have sev-eral drawbacks, such as poor measurement in below areaof cortex and poor signal to noise ratio that require largenumber of participants for extracting useful information fromEEG. The major bio-markers that have been reported in theliterature for depression diagnosis are shown in Table 6.
5. Neural networks-based approaches forDepression recognition using EEG signals
The fields of Affective Computing (AC) and Neural Net-works (NNs) are useful in solving complex and multidimen-sional problems such as modeling social affective behaviorand mental disorders. These problems involving affectivedatasets need Deep Neural Networks (DNNs), neural net-works with two or more hidden layers, for effective temporalmodeling, and real-time performance analysis. The ability ofDNN to identify latent structures in raw, unlabeled, unstruc-tured, noisy, and incomplete EEG datasets makes it suitablefor EEG based depression diagnosis. The effectiveness ofDNN based solutions for depression diagnosis depends onthe EEG experimental protocols, placement, and types ofEEG electrodes, and the availability of EEG datasets. Thisalso requires investigating automatic depression assessment
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 8 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 6: Psychological Biomarkers for the assessment of depression.Ref EEG bio mark-ers ERP Biomarkers NeuroimagingBio markers Genes andother Biomarkers Common Findings[96] Alpha powerand Asymme-try(FAA) x x x FAA is able to distinguishMDD and healthy controlbased on gender, age and de-pression severity.[97] Theta band ac-tivation’s x x x Frontal theta asymmetry in-creased in normal subjectsbut reserved in depressed pa-tients during music listening.[103] Antidepressanttreatment re-sponse(ATR)index x x x ATR is a potential predictorand biomarker of NTR treat-ment of MDD pateints.[98] Theta QEEGcordance x x x The change in QEEG distin-gusish MDD and normal con-trol.[104] ReferencedEEG (rEEG) x x x It provides assistance inMDD medicine selection.[105] Rostral ante-rior cingulatecortex (rACC) x x x Low rAcc is a best responderto depression treatment.[94,106] Wake andsleep EEG x x x The variation in wake andsleep helps in depression de-tection.[99] Gamma waves x x x gamma can discriminate unipolar disorder from bipolar.[107] Resting stateEEG x x x Extracted EEG features helpsin depression diagnosis.[97] Frontal thetaAsymmetry x x x It provides reliable classi-fication in comparison togamma.[88] x x Anterior insulametabolism x It identify treatment out-comes of depressed patients.[108] x x Psycho motor re-tardation x Affected by depression.[88] x x Hippocampal vol-ume x Depresses patient have 19%small hippocampal.[109] x x Cognitive func-tions - Cognitive actives gets dam-aged by depression.[20] x x Subcallosal cin-gulate cortexmetabolism x Abnormal SCG distinguishthe MDD and control sub-jects.[110] x x x Vocal andFacial biomarkers It is less reliable in compari-son to EEG biomarker.[90,89] x x x 3p25-26 This genes is found in morethan 800 depressed familiesbut still immature.[92] x P300 x x It is used as indicator to mea-sure the severity of depres-sion.[93] x LDAEP x x The higher LDAEP valuesfound in depression patient.
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 9 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review methods.
The EEG experimental protocol for depression recogni-tion defines a standard set of rules such as number of par-ticipants, selection criteria of participants, EEG recordingduration,and so on. Different EEG-based experimental pro-tocols for depression recognition have been proposed in theliterature. This section discusses the EEG based experimentalprotocol for depression recognition using shallow and deepneural networks.
Participants:
The number of participants varies in dif-ferent studies, from one to two hundred, with a median of30 subjects. Most of these studies are based on a relativelysmall number of participants, therefore, it is hard to establishthe exactness and significance of the results. Table 7 presentsa summary of existing studies highlighting the number ofsubjects, gender, age group, prior history, and depression.According to Table 7,only ten out of fifty studies includethirty participants for the diagnosis of the EEG based depres-sion. Thirteen studies include 50 to 60 subjects with an equalnumber of depressed and normal participants.The remain-ing studies consider different numbers of subjects.Regardinggender selection, most of the studies include both male andfemale participants; While, few studies include just femaleparticipants for EEG based major depressive disorder detec-tion. Since men and women might perceive depression indifferent ways, so the participants from both genders must beincluded in studies.
Selection Criteria of Participants:
The selection of theparticipants is the most important phase of EEG based de-pression recognition. The quantitative (clinical questionnaire-based) and qualitative (EEG-based) methods are used as stan-dard tools for the selection of participants and data collec-tion. In EEG based mental disorder studies, several clinicalquestionnaire-based pre-testing techniques are used in Table8 for participantsâĂŹ recruitment, and then the patientâĂŹsselection for EEG recording. Beck Depression Inventory(BDI) [74] is a common valuation tool used by healthcarespecialists and researchers for pre-testing of depression andanxiety diagnosis. In this test, patients with MDD are selectedbased on different multiple-choice questions. The patientswith BDI-II score above 13 are considered as a depressivesubject. Overall high score of BDI-II shows the severity ofdepression. BDI is used in [53, 75, 114, 56, 57, 115, 116, 117,119, 126, 128, 138, 129, 136] for participants selection. Diag-nostic and Statistical Manual of Mental Disorders (DSM-IV)developed by the American Psychiatric Association (APA)has been used by several studies for the measurement of dif-ferent mental illnesses. In EEG based diagnosis, DSM-IVbased questionnaire is used for pre-psychometric test to as-sess depression. Most of the articles used it as pre-EEG test[53, 55, 77, 125, 130]. Hospital Anxiety and DepressionScale (HADS) is used in [53] to measure the different levelsof depression. Another participant selection method is knownas PHQ-9 that is a multiple-choice self-report inventory. It consists of 9 questions and is used as a selection/diagnostictool for mental health syndromes such as depression. Thesubjects are categorized into different levels of depressionsuch as minimal, mild, moderate, moderately severe, andsevere depression based on the standard score of the ques-tionnaire. The PHQ score standard is presented in Table 4.The most of the articles use it for pre-screening of the sub-jects [60, 73, 122, 135]. The participants are shortlisted forqualitative EEG based recording after the pre-screening test,and approval of the design of study from the ethics depart-ment. The participants perform specific tasks for a particularduration in a calm room under a resting state with a differentnumber of non-invasive wet or dry electrodes at the differentregions of the scalp. The recorded EEG signals contain a lotof noise that is removed for further processing of the signaland depression classification.
Placement and Types of EEGElectrodes:
The number of electrodes, their placement, andtype play an important role in the EEG based depression andbipolar disorder diagnosis due to multiple reasons. First, thetime required to set up the EEG device, second, ease of thepatient who wears the EEG device, and third the number offeatures to process [139].For these reasons, researchers haveused different numbers of electrodes and standards to acquireEEG signals from the scalp. According to Table 9, thirtyone studies out of fifty use wet electrodes to record the EEGsignals and the remaining use dry electrodes. The position ofthe electrodes at the scalp is also important because differentlobes of the cerebral cortex are responsible for processing theelectrical activities of the brain. As far as placement of elec-trodes at the scalp is concerned, two standards of 10-20 and10-10 exist in the literature. The 10-20 and 10-10 electrodesplacement systems are based on an international standard thatdescribes the location of electrodes at the scalp. In 10-20 stan-dard, "10" and "20" represent the 10% or 20% inter-electrodedistance and in 10-10 standard,"10" and "10" shows the 10%inter-electrode distance. These standard electrode placementsystems are based on the correlation between the location ofan electrode and the underlying area of the cerebral cortex.The even numbers at the scalp represent the right hemisphereand odd numbers refer to the left hemisphere. The lettersF, T, C, P, and O stand for Frontal, Temporal, Central, Pari-etal and Occipital respectively. They are used to identify thebrain lobes and place the electrodes at the scalp. The pointz refers to the midline of the brain. In the 10-20 electrodestandard, the smallest number is closer to the midline andvice versa. A bird’s eye view of electrode placement is shownin figure 1c. According to Table 9, twenty-five studies adopt10-20 standard, only two use 10-10 standard, and remainingdo not mention any information about electrode placement.According to Table 9, number of electrodes varies in differentstudies. Researchers decide about the number of electrodesaccording to their requirements and diagnostic criteria.
Due to the sensitive nature of depression data and forprivacy and confidentiality reasons, very few public datasetsare available for EEG based depression diagnosis, therefore,
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 10 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 7: Participants information in EEG experiments for depression detection where D for Depressed, H for healthy, F forfemale and M for male is used.Ref Subjects Gender Age Group(Mean, Âś SD) OtherInforma-tion Depression Types[53] 63(33D+30H) 27F,36M 38-40 (38.22, 15.64) - Unipolar depression[111] 30(15D+15H) - 20-50 - Bipolar depression[75] 51 15F,36M 18-24 (20.96, 1.95) - Mild depression[54] 30(15D+15H) - 20-50 - Bipolar depression[112] 60(30D+30H) - 20-50 - Bipolar depression[113] 10 - - - Depression[114] 60 30F,30M 10,32 (32.4, 10.5) - Minimal to severe depression[55] 28(14D+14H) - - Righthanded Depression[56] 15 - - - Depression[57] 89 - - - Uni and Bipolar depression[115] 60(30D+30H) 32F,28M 20-50 - Bipolar depression[116] 30 16F,14M 20-50 - Depression[72] 213(92D+121H) - - - Depression[117] 60(30D+30H) 20-50 - - Depression[118] 16 - - - Depression Level[119] 51 15F,36M 18-24 Righthanded Mild depression[60] 116 50F,66M 19-25 - Mild depression[77] 55 - - - Major depressive disorder[73] 34(17D+17H) 17F,17M 30-33 (33.35, 12.36) - Depression[120] 22 - - - Mild Depression[121] 25(12D+13H) 25F 30-42 (24.23, 6.33) Righthanded Pervasive depression[122] 178(86D+92H) - - - Depression[78] 24(12D+12H) 10F,14M 20-28 - Major Depressive Disorder[123] 30(15D+15H) 16F,14M 20-50 - Depression[124] 30D - 20-50 - Depression[125] 64(34D+30H) 26F,38M 15-38 - Major Depressive Disorder[126] 25(13D+12H) 25F,0M (24.23, 6.33) Righthanded Major depressive disorder[127] 23(12D+11H) 12F,11M 21-55 (26.4, 10.6) - Depression[128] 16 - - - Depression Level[129] 51(24D+27H) 15F,36M 18-24 (20.96, 1.95) - Mild depression[130] 64(34D+30H) 26F,38M 12-40 - Major depressive disorder[131] 12(12D) 6F,6M 20-35 - Depression[132] 30(10D+10H+10S)- 13-53 (31.5) - Depression[133] 22(12D+11H) 10F,12M 20-24 - Depression[134] 20(10D+10H) 10F,10M - - Depression[135] 116(63D+53H) - 19-25 - Depression[136] 63(33D+30H) 27F,36M 38-40 - Unipolar depression[137] 22(12D+10H) - 69-70 (69.81) - Depressionmost research groups use their datasets. This section presentsthe few publicly available datasets for EEG based depressiondiagnoses as shown in Table 10.
Healthy Brain Network : HBN is a public biobank of Healthy Brain Network Data set link http://fcon_1000.projects.nitrc.org/indGi/cmi_healthy_brain_network/sharing_neuro.html data that is launched by the child mind institute. The majorgoal of HBN is to produce a dataset that captures a widerange of heterogeneity and impairment that occurs in evolv-ing psychopathology. The HBN contains information aboutdepressive disorder, behavioral, intellectual, eye tracking, andphenotype data, by using multi-modal EEG and brain imag-
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 11 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 8: Participants selection and experimentation tasks in EEG studies for depression detection.Ref EEG Selection Criteria Subject Inclusion cri-teria Tasks Depression criteria[53] No psychotic disease DSM-IV BDI,HADS (BDI-II,HADS)score > > > > > > > > > > > > > > > > Sana Yasin et al.:
Preprint submitted to Elsevier
Page 12 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 9: EEG devices,number of channels and their placement used for EEG experiment for depression detection.Ref EEG Device Electrode Brain Lobes Placement/Standard Types ofElectrodes[53] EEG cap 19 - 10-20 Wet[111] Pair electrode 2 Channel pair left/right half - Wet[75] (HCGSN) 128 - 10-10 Dry[54] (HCGSN) 2 Channel pair Left/right half - Dry[112] Bipolar Montage 2 Channel pair Frontal lobe 10-20 Wet[113] Salivary cortisol 14 - - Wet[114] - 19 - - Wet[55] (HCGSN) 16 - 10-20 Dry[56] Procomp 1(F4) - - Dry[57] Procomp 1(F4) - - Dry[115] - 24 - 10-20 Wet[116] - 2 Channel pair Left/right half - Wet[72] - 3(Fp1,Fp2,Fpz) Frontal lobe 10-20 Wet[117] Bipolar Montage 2 Chanel pair Frontal lobe 10-20 Wet[118] Ag-Cl electrodes 3 Frontal lobe 10-20 Wet[119] Geodesic,HCGSN 16 - 10-20 Dry[60] 32 Channel hard-ware 16 - - Wet[77] (Compumedics/Neuroscan) 6 Frontal lobe 10-20 Wet[73] HydroCel GSN(128) 64 - 10-20 Dry[120] SynAmps 61 - 10-20 Wet[121] Mobile EEG belt 3(Fp1, Fp2 and Fpz) Frontal lobe 10-20 Dry[122] - 3(FP1,FP2,FPz) Frontal lobe 10-20 Wet[78] - 7(FP1-FP8) Frontal lobe 10-20 Wet[123] - 2 Channel pair Left/right half 10-20 Wet[124] Bipolar Montage 2 Channel pair Left/right half 10-20 Wet[125] ProComm 1(F4) Frontal lobe - Wet[126] EEG belt 3 Frontal lobe 10-20 Wet[127] Brain amp 6(Fp1, Fp2, F3, F4,P3 and P4) Prefrontal,parietalcortex 10-20 Wet[128] - 3 Frontal lobe 10-20 Wet[138] Geodesic sensornet (HCGSN) 128 - 10-20 Dry[129] HCGSN 128 - 10-10 Dry[130] - 19 Temporal,parietal10-20 Wet[131] - 3(Fp1, FpZ and Fp2) Prefrontallobe 10-20 Wet[116] - 2 Channel pair(FP1-T3)FP2-T4) left/right half 10-20 Wet[132] HCGSN 16(Fp1-FP8, C3-C4,P3-P4, T3-T6,O1-O2) Frontal,earlobe 10-20 Wet[133] - 5(F4,F3,Cz,M1,M2) Frontal,earLobe 10-20 Wet[134] - 64 - 10-20 Wet[135] 32 channels wetEEG hardware 16 place at wholescalp 10-20 Wet[136] EEG cap 19 - 10-20 Wet[137] Nihon kohden JE-207A easy cap 57 - - Wet
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 13 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review ing MRI. In the HBN, a clinical assessment of mental healthand learning disorders is performed. The HBN protocols in-clude the behavioral and physical measures, family structure,stress, trauma, cognition, and language tasks. About 10,000subjects of New York with ages between 5 and 21years par-ticipated in data collection. The clinical team consisted of amixture of psychologists and social workers. The safety andeligibility of the participants was confirmed by prescreeninginterview. The screening interview collected informationabout the subjectâĂŹs psychiatric and medicinal history.
EMBARC (Establishing Moderators/Mediators for aBiosignature of Antidepressant Response in Clinical Care) is a public data set made available by the National Institute ofMental Health (NIMH). It recognizes the neurological sign ofreaction to antidepressant treatment by using the resting-stateEEG and machine learning algorithm. A total of 16 posteriorelectrodes were comprised of P1, P2, P3, P4, P5, P6, P7, P8,PO3, PO4, PO7, PO8, POz, O1, O2, and Oz. [140].
Depresjon public dataset is used in [141], it contains mo-tor activity recordings of 23 unipolar and bipolar depressedpatients and 32 healthy controls . Trans diagnostic cohorts [142] is a publicly availabledataset that evaluates the brief transdiagnostic cognitive-behavioral group therapy (TCBGT) for the treatment of anxi-ety and depression patients. It contains 287 participants inprimary care with depression and anxiety disorders. Thesepatients spent approximately 5 weeks with TCBGT. ANOVAtests that have mixed design capabilities have been used forstatistical analysis and to check the effects of treatment.
Multimodal Resource for Studying Information Pro-cessing in the Developing Brain (MIPDB) focuses on neuro-phenotyping of psychiatric and healthy populations of di-mensional and multi-domains. They intellectualize mentaldisorder in terms of domain-general discrepancies, insteadof considering a single factor[143].
Patient Repository for EEG data and computationaltools (PREDICT) is a high volume publicly available datasetthat contains EEG data. There exist several data reposito-ries which contain imaging and patient-specific data. PatientRepository For EEG Data + Computational Tool (PRED+ CT) is one of the few sites that offer EEG based patient-specific data. It provides a centralized platform by categoriz-ing psychiatric and neural patients based on EEG data storage,tasks, and computational tools. In PRED+CT, EEG data im-plementation is performed in MATLAB toolbox based onpatient/control, symptom scores, age, and sex [144]. In this section, automatic depression assessment methodsare investigated by different prepossessing techniques, Neuralnetworks (NNs), and deep learning-based approaches fordepression detection. Depresjon Data set link https://datasets.simula.no/depresjon/ Patient Repository for EEG data and computational tools (PREDICT)Data set Link
PatientRepositoryforEEGdataandcomputationaltools(PREDICT
The EEG signal recording is a time-consuming proce-dure in which depressed patients perform some tasks. Duringrecording, EEG signals are contaminated by undesired orpolluted signals called artifacts. The artifacts that occur dueto the patient body movement, heartbeat, eye blinks, musclemovement are called physiological artifacts. The artifactsthat occur due to the electrodes placement, environmentalnoise, and device error are known as non-physiological ar-tifacts [147]. These artifacts affect the quality of the ac-tual EEG signals; hence it is important to sanitize usefuldata from contaminated EEG signals through preprocess-ing of the EEG data. In this phase, different undesired arti-facts are filtered by using different noise removal and ar-tifact elimination algorithms to prepare data for the nextstage. Table 11 shows the shows the different artifact re-moval techniques. Mumtaz et al. [53], claim that raw EEGsignals have poor resolution due to the low Signal-to-Noiseratio (SNR). Therefore, to enhance the performance of EEGsignals, multiple source eye correction (MSEC) techniqueis used to remove the undesired signals. The authors ofreferences [111, 54, 117, 116, 123, 124, 131, 135] used aNotch filter with 50Hz to remove power line interventionand to sanitize EEG signals from artifacts. While, articles[75, 113, 114, 121, 123, 129, 130, 134, 135] used a low passand high-pass-filter with 40Hz and 1Hz cutoff frequency tofilter the noise. Adaptive noise canceller and fast ICA is usedin [55, 137] to remove inaccurate information from the falseEEG recordings. As per our findings, the low pass, high pass,and Notch filter are frequently used to remove the artifacts inEEG based depression detection.
Neural Networks (NNs) [148] are non-parametric, flexi-ble, and parallel computing systems that consist of neuronslayers and weighted links in which information is transferredfrom the input neurons to the output neurons in a forwardor backward way. In recent years, ANN-based approachesare used in EEG studies for classification and diagnosis ofmajor depressive disorder. ANN can classify nonlinear rela-tions and incorporates high-order dealings between predictivevariables to produce accurate results. Most of the articles[57, 77] use it for the classification of the unipolar and bipolardepression and have achieved good accuracy of 89.09% byimplementing the pre-treatment cordance of frontal QEEG.Multi-layer FFNN, Back Propagation Neural Network(BPNN), and Enhanced probabilistic neural network (EPNN)have been used in [60, 121, 78] for discriminating MDD andnon-MDD patients. In the test, EEG recording of depressedand age control subjects were collected under the 10 cross-validation technique. EEG electrodes were placed at theprefrontal (Fp1 and Fp2), frontal (F3, F4, F7, and F8), cere-bral (C3 and C4), and temporal (T3, T4, T5 and T6) regionsof the brain. The results indicate that the signals collectedfrom the cerebral (C3 and C4) region gives slightly higheraccuracy as compared to the other brain regions. Multi-layer
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 14 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 10: Public EEG datasets for depression diagnosis.Ref Dataset Name PatientGroup Task dur-ing EEGrecording Number of Pa-tients EEG system Electrodes[145] HealthyBrain Net-work(HBN) Depression Resting state 10,000 128-channel EEGgeodesic hydrocelsystem 128[140] EMBARC Depression Resting state 675 - 16[141] Depresjon Depression Motor activity 55(23D+32H) - -[142] Trans diagnos-tic cohorts - - 287 - -[146] PREDICT(DepressionRest) Depression/High BDI Resting state 46 Neuroscan 64[143] MIPDB Depressive Resting state 126 - 109[144] PREDICT(DepressionRL) DepressionRL/ HighBDI Reinforcementlearning 46 Neuroscan 64FFNN performs better than BPNN and EPNN by achieving95% classification accuracy.In [56, 115, 116], a comparative study of FFNN, neuro-fuzzy networks, relative wavelet energy (RWE), and proba-bilistic neural networks (PNN) has been performed for differ-entiating the depressed and normal patients through EEG sig-nals. Classification abilities of neural, neuro-fuzzy network,and relative wavelet energy (RWE) networks are authenti-cated by the EEG recordings. The FFNN leads the PNN,neuro-fuzzy networks, and relative wavelet energy (RWE)achieving 100% classification accuracy with time and waveletenergy as the input features.EEG entropies based depression detection has been per-formed in [123, 149] by comparing the Probabilistic NeuralNetwork (PNN) performance with Radial Basis Function(RBF) and different machine learning classifiers. Resultsshow that PNN performs better with 99.5% classificationaccuracy.
Deep learning is the most popular area of research andcurrently, it is extensively used for the classification of EEGsignals in comparison to other approaches[150]. A literaturesurvey shows that approximately there are six different ar-eas of EEG:sleep,seizure prediction,mental workload,motorimagery,emotion recognition and depression that use deeplearning for the detection and classification purposes [151].According to the latest survey on deep learning [151], it isobserved that there are 16% articles for EEG based emotionrecognition system, 22% for motor imagery, 16% for mentalworkload, 14% for seizure detection, 9% for sleep stage scor-ing, 10% of event-related potential detection, 2% for EEGbased depression and 8% for AlzheimerâĂŹs, gender andabnormal signal classification(Please see Fig. 2). Herein, wesee that the motor imagery is the most frequently exploredarea but the use of deep learning in EEG based depression recognition is not much reported so far. In [53] authors claimthat there are several articles on the automatic diagnosis ofdepression with traditional classifiers, however, there are onlytwo or three articles on deep learning methods for EEG baseddepression detection. In the existing studies, questionnaires[37], visual cues [5], regression and machine learning-basedtechniques [17] are used for the depression recognition andtreatment. Besides all these efforts, the depression detec-tion field still needs improvements. To enhance the featureextraction and classification accuracy for depression recogni-tion and assessment tasks, deep learning is used for the lasttwo years. The Convolutional Neural Network (CNN) is themost famous model of deep learning and is widely used inmost of the EEG studies to diagnose major depressive dis-order. The one Dimensional Convolutional Neural Network(1DCNN) is used with Long Short Term Memory (LSTM)for the classification of unipolar depression [53]. This hy-brid architecture achieves 95.97% accuracy by automaticallylearning EEG patterns. The cerebral cortex is divided intothe left and right hemispheres of the brain. In [111], the au-thors used Deep Convolutional Neural Networks to classifydepressive and non-depressive subjects against the left andright hemispheres. Results show that the right hemisphereis more accurate in depression classification as compared tothe left hemisphere with 93.5% accuracy. Mild depressionrecognition of college students has been performed [75] bythe convnets framework. The convnets framework achieves85.62% accuracy with 24-fold cross-validation. Automaticdepression detection using deep representation and sequencelearning is performed in [54]. These two hybrid architecturesof deep learning provide sequence learning and extract thetemporal properties of the EEG signals. It achieves 97.66%classification accuracy by placing the electrodes at the leftand the right hemisphere of the brain.The Neucube architecture of deep learning with 10 cross-validations is applied in [120] with the resting state of EEG.
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 15 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 11: Artifacts and noise filtering approaches in EEG based depression detection.Ref Artifact filtering technique[53] MSEC[111] Notch filter:50Hz[75] High pass, net station waveform[54] Notch filter:50HZ[112] -[113] Low-pass,high-pass, notch filter[114] Highpass, butterworth,low-pass[55] Adaptive noise canceller, Hzband-pass,fastICA[56] -[57] Band pass filter 40HZ[115] Total variation Filtering(TVP)[116] Visual inspection, total variation filtering (TVF)[72] Kalman filter,adaptive predictor filter[117] Notch filter:50Hz[118] -[119] Net station waveform tool[60] High pass,Low pass, and Notch filter[77] Band pass filter (0.15âĂŞ30Hz)[73] FastICA,hanning filter[120] Off line ICA[121] Low pass filter at 50HZ[116] Notch filter at 50HZ[122] Band pass filter[78] Wavelet filter[123] Notch and low pass filter at 50HZ[124] Notch filter at 50HZ[127] Convolutional filter[128] ANFIS[138] Net station waveform tools[129] Low pass,high pass filter[130] ICA,notch,low pass,high pass filter[131] Notch filter[116] Visual inspection,50-Hz notch filter[132] EEGLAB toolbox[133] Stationary wavelet transform (SWT)[134] Lowpass and highpass filter[135] Lowpass,highpass and notch filter[136] MSEC[137] ICA and butterworth filterThe neucube architecture is compared with Multi-layer Per-ceptron (MLP) and other traditional machine learning meth-ods. The results show that neucube performs better than MLPand all other traditional machine learning algorithms such asSVM, decision tree and logistic regression. Deep learning-based depression detection with frontal imagery EEG chan-nels is performed in [136]. The Oxford Net (VGG16) modelwith max pooling and softmax activation function isused. The results show that OxfordNet (VGG16) has 87.5%classification accuracy. Functional connectivity and EEG based mild depression detection has been performed in [152]by using a deep learning approach. Initially, abnormal func-tional connectivity is measured by using the graph theory, andthen CNN is applied for the classification of the depressionwith 80.74% accuracy.Deep neural network approaches have also been used forfeature extraction for transformation of the huge volume ofinput data into imperative features sets. In Major Depres-sive Disorder (MDD), the feature extraction phase selects themost imperative features or information from the EEG sig-
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 16 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Figure 2:
Ratio of Deep Learning studies adopting EEG signals for a variety of applications referred to [151]. nals for depression classification. The feature extraction parthas a great influence on the accuracy of the results. In [53]temporal feature is extracted efficiently by using the One Di-mensional Convolutional Network (1DCNN). The first layerof the CNN layer maps the 64 features, the next layer reducedit from 64 to 48, and the third layer includes 24 features onlyby using the pooling layer and gets better experimental results.In [111] authors do not use any manual set of features to befed into a depression classifier. The given CNN model has thecapability to self-learn and select the unique features duringtraining process without a separate feature selection step. In[75] pre-trained model based on transfer learning is used toextract the power spectrum of EEG signals observing thedifferences between depressed and non-depressed patientsin alpha, beta, and theta frequency bands. The pre-trainedmodel removes the last output layer of CNN and then usesthe entire network as a fixed feature extractor. In [54] bothlocal and long term feature selection/extraction and depres-sion classification operations are automatically performedby using an end-to-end single framework and EEG signalsas input. Convolutional and pooling layers of deep learningmodels are used in [113] for feature extraction. The benefitof these layers is that they enhance the classification accu-racy by extracting the features from the nearest neighboringpixels. Brain has a nonlinear and complex system so non-linear features including fuzzy entropy FuzzyEn and fractaldimensions (KFD and FFD) are used in [114]. Ensemblelearning and deep learning approaches are applied in [55]for the processing of features from brain signals, power spec-tral density and activity are extracted as original features.Neucube model as a feature extractor is applied in [120]. Itachieves not only high classification accuracy, but also re-veals patterns of brain activities relevant to the two classesdepressive and non-depressive subjects. The performancesof a variety of EEG based deep learning and neural network approaches for depression detection are presented in Table12.
6. Clinical background of Bipolar Disorder
Bipolar Disorder(BD) or manic-depressive illness (MDI)is a dangerous neural disorder that is predicted by mood un-certainty and may start from early ages (i.e., in infants andteenagers). Individuals with BD face recurrently swings be-tween depressive and manic episodes [15]. According to theAmerican Psychiatric Association [16], there are four majorcategories of bipolar disorder: bipolar I disorder , that hasone atleast one full episode of mania or diverse episodes ofmania and depression.
Bipolar II disorder has no manicepisode, minimum one hypomanic episode and many depres-sive episodes. Patients with cyclothymic disorder has manyhypomanic and depressive episodes.
Bipolar disorder in-cludes both depressive and hypomanic episodes alternatively.Bipolar disorder and depression have various resemblances,yet they have some crucial dissimilarities [157]. Bipolar dis-order has experience of high mood swings, different episodesof depression and periods of excessive highs (also known asmania). In contrast, depression is just one state of bipolardisorder that is above from the low feeling. It’s a deep griefor hollowness that an individual can’t manage. Sometimes,bipolar patients feel disheartened, valueless, and tired andmight fail to concentrate in possessions that they had enjoyedbefore the illness. Clinical depression (also known as majordepressive disorder or MDD) frequently drives with sleepdifficulties, alterations in hunger, and trouble focused. Itcan lead to suicidal activities or actions. People with majordepressive disorder do not practice any extreme, elevated feel-ings that bipolar patients face (such as mania or hypomania).In bipolar disorder, the symptoms of intense periods withlow and high moods do not follow a proper pattern. Somebipolar patients experience the same mood state for a long pe-
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 17 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 12: Performance of different EEG based deep learning and neural network approaches for depression detectionpresentation in terms of accuracy.Ref Model used Results(Accuracy)[53] RNN(LSTM),1DCNN 98.32%[111] CNN 95.49%[75] CNN 85.62%[54] CNN+RNN(LSTM) 99.12%[112] CNN 99.31%[113] CNN 86.62%[114] Fuzzy function based on neural net-work(FFNN) 87.5%[55] DL(CNN) 84.75%[56] ANN and Fuzzy 13.97% and76.88%[57] ANN 89.89%[115] FFNN and PNN 58.75% and 98.75%[116] ANN 98.11%[72] ANN 72.56%[117] Deep learning (CNN) 99.3%[118] ANN 91.7%[119] Multi model deep learning, 83.42%[60] ANN, 95%[77] ANN, 89.09%[73] DL(CNN), 77.20%[120] Deep learning, 90%[121] BPNN 94.2%[116] FNN 98.11%[122] ANN,DBN 78.24%[78] EPNN 91.3 %[123] PNN 99.5%[149] ANN -[153] MLP 80%[124] RNN(LSTM) 80%[125] MLPNN 93.33%[126] BPNN 94.2%[127] CNN 79.08%[128] ANFIS,NPR(neural network tool) NPR 88.32,ANFIS 91.7%[138] ANN(BNMLP) 83.42%[129] CNN 85.62%[130] MLPNN 93.33%[131] FBNN 70%[116] ANN 98.11%[132] ANN -[133] BP neural network -[134] ANN -[135] ANN 95%[154] PNN 98%[155] Deep learning(SPN) -[136] Deep learning 87.5%[137] Multilayer perceptron(MLP) 95.45%[156] Deep neural network 95.45%[152] Deep learning(CNN) 80.74%
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 18 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review riod before shifting to the opposite mood and vise versa [158].These depressive periods occur for a week, month, or some-times even for a year. Bipolar severity varies with respect toperson and occurrence time. Bipolar disorder symptoms arebroadly classified into two types: 1) External and physical,and ii) Internal and physiological symptoms.Bipolar disorder have a strong physical effect on patient’sbody and creates a lot of physical illness like flu, palpitations,diarrhea, abdominal pain, nausea and vomiting, high pulseand heart rates, higher blood pressure, weight and appetitechanges, fast speaking and poor attention,strangely high sexdrive, enhanced energy and less need for sleep.[159, 160].The primary physiological symptoms of bipolar disorderare penetrating and unpredictable. It includes severe moodswings, extreme pleasure, hopefulness and excitement, rapidchanges from being glad to being ill-tempered, irritated, ag-gressive and becoming more thoughtless. In the followingsubsections we will focus on diagnosis and assessments ofbipolar disorder like clinical bipolar disorder.Irregular physical features of the brain or an inequalityin certain brain chemicals may be the main causes of bipolardisorder. The assessment of bipolar disorder is not alwayseasy like other mental disorders [161]. Bipolar patient mostprobably goes to their consultant for the first time when theyhave a depressive episode instead of during a manic or hy-pomanic episode. Due to this reason, in the beginning clini-cians frequently misdiagnose bipolar disorder as depression.By considering all limitations of bipolar disorder, scientistsand researchers have introduced some automatic and reliablesources for the diagnosis and treatment of bipolar disorder.Nothing is more significant than diagnosing a patient withbipolar disorder or manic-depressive illness (MDI), as onlyaccurate diagnosis can lead to a proper effective treatment.Electroencephalogram (EEG) is a top rated neuroimagingtechnique that is becoming a central focus of researchersfor the past few years and widely used to diagnose mentaldisorders.
Like major depressive disorder, questionnaire based as-sessment tools are also used for bipolar disorder participantsto investigate their physiological responses. The Young Ma-nia Rating Scale (YMRS), Hamilton Depression Rating Scale(HDRS) and Structured Clinical Interview (SCID), are thefew bipolar assessment tools that are widely used for partic-ipants selection [57]. The YMRS [162] is one of the mostwidely used rating scale to evaluate the mania symptoms. Itconsists of 11 items to evaluate the clinical condition of pa-tients that appeared in the last 48 hours. The purpose of eachitem is to measure the abnormality of it. Next to it, HDRSScale is used to measure the severity of the disease. Based onthe clinical feature and symptoms, patients are categorizedinto different diseases like (YMRS=12) show bipolar disor-der, (YMRS=3) shows depression and (YMRS=2) showsethymia.The HDRS [163] originally published in 1960 and used to evaluate the recovery process (and also to measure theseverity) of major depressive disorder and bipolar disorder.It contains ten multiple choice based questions that providesan indication to the bipolar depression. It usually works forthe adults and measure the severity level of depression andmania according to the HDRS total score. The SCID [164]is a structured diagnostic tool that developed in 1990 andworked with different versions of DSM to determine theiraxis. It is organized into different modules and each moduleis used to detect unique type of disorder i.e., SCID-I diag-nose the mental disorders, SCID-II determine the personalitydisorders and SCID-5 diagnose the anxiety, eating, gamblingand sleeping disorders.
The pathophysiology of bipolar disorder is complex, multifactorial, and not fully understandable. To overcome this com-plexity, the biomarkers based assessment not only facilitatesthe diagnosis and monitoring of complex bipolar disorder,yet also provide biological effects of treatment. These assess-ments devise a new hypotheses about the causes and patho-physiology of bipolar disorder. The peripheral biomarkerslike neurotrophins, oxidative stress and neuroinflammationare used in [165] to measure the illness activity of bipolardisorder. The neurotrophins [166] is a family of proteins thatinduce the survival, development, and function of neurons.It shows distinct patterns in the different stages of bipolardisorder; therefore, [167] use it as brain-derived neurotrophicfactor (BDNF) bipolar disorder biomarker. The oxidativestress is an imbalance of antioxidants and free radicals inthe human body. Its ratio increased during the bipolar disor-der episodes; therefore, in [168] oxidative stress is used todiagnose the bipolar disorder. The neuroinflammation is a lo-calized physical condition of the body that becomes reddenedor swollen during bipolar disorder. The authors of [169] useit for bipolar disorder study. The spectral entropy modula-tion that quantifies the EEG signal degree of uncertainty getsdeficit during manic episodes of bipolar disorder, so in [9] itis used as a bipolar disorder and schizophrenia biomarker foraltered function. The brain oscillations and lithium responsevariate during bipolar disorder condition and used as bipolardisorder biomarker in [170] .
7. Neural Networks based approaches forBipolar disorder recognition using EEGsignals
Bipolar disorder (BD) is a composite disorder that os-cillates between two mood states depression and mania andis often misdiagnosed by physicians [15]. It is an enduringcondition in which BD patients spend most of their life ina misery of symptoms of depression, which complicates ac-curate identification and analysis of bipolar disorder. Thefundamental clinical challenge of this disease is the differen-tiation between BD patients and the patients showing symp-toms of general depression. Neural networks (NNs) based
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 19 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review approaches offer new ways to predict bipolar disorder recog-nition and clinical outcomes for individuals and reduce thecomplexity of bipolar disorder identification process. NNsprocess the psychiatric data that is relayed on the brain struc-ture and by resolving the complex real-world problems whichis otherwise difficult for conventional tools and techniques.In contrast to the human data processing ability, NNs pro-vide better precision and time effective solutions for patternrecognition and prediction problems.Multi-layer perceptron (MLP) is the most popular super-vised neural network that is widely used for classification andassessments. It is a class of FFNNs (Feed Forward NeuralNetworks) that contains at least three layers: an input layer,a hidden layer and an output layer. In [171], MLP is usedas BD (Bipolar Disorder) classifier with five neurons andone hidden layer. To acquire the best conceivable precision,several discriminative features can be extracted from the EEGrecordings by using the four diverse feature selection algo-rithms. These algorithms are CMIM(Conditional MutualInformation Maximization),MIM(Mutual Information Maxi-mization),DISR(Double Input Symmetrical Relevance) andFCBF(Fast Correlation Based Filter). The extracted featurescan be fed to the MLP for classification purpose. Resultsshows that MLP achieve 91.83% classification accuracy forbipolar disorder sub types and normal subjects [171]. Theartificial neural networks and quantitative EEG have beenused in [172] for differentiating fronto temporal dementiafrom late-onset bipolar disorder. All patients are assessedby the clinical MRI scan(Magnetic Resonance Imaging) andEEG(Electroencephalogram). The results represents that acombination of EEG and MRI with ANN classifier givesbetter classification results as compared to EEG and MRIseparately.As one of the major unbearable neural syndrome, BD(bipolardisorder) is commonly misdiagnosed as UD(unipolar disor-der) , that further leads to suboptimal cure and poor results.Therefore,the classification of UD and BD at initial phasescan therefore support to assist effective and precise treatment.In [57] artificial neural network classifier with quantitativeEEG is used as a biomarker for the classification of the unipo-lar and bipolar disorder and it achieved an accuracy of 89.89%.A bipolar depression patient experiences both manic and de-pressive periods. To classify these two depressive periods,the feedforward neural network (FFNN) and probabilisticneural network (PNN) is used in [115]. The FFNN showed98.75% classification accuracy while PNN achieved an ac-curacy of only 46.5%. The results show that FFNNs givebetter classification results for bipolar disorder as comparedto PNN.Classification of bipolar EEG signals in normal and de-pressive condition has been performed in [116] by usingrelative wavelet energy (RWE) and an artificial feed forwardneural network. The performance of the artificial neural net-work was assessed by the classification accuracy and its valueof 98.11% shows its unlimited potential for classifying nor-mal and depressive subjects. In [173] convolutional neuralnetwork with electroencephalography features are used for the precise diagnosis of the depression sub types. It achieves99.5% accuracy in the classification of unipolar vs healthysubjects and 85% in the discrimination of bipolar vs healthysubjects. In [174], bipolar and schizophrenia disorder di-agnosis is achieved by using an artificial neural network.It achieves 90% classification accuracy among the bipolar,schizophrenia and healthy subjects.
EEG experimental protocols are a set of rules that aredefined before EEG recording. It includes the number ofsubjects that have participated in the EEG based study, selec-tion criteria of participants, placement standard and types ofelectrodes that are used for recording the bipolar activity.
Participants:
The ratio of the number of participants forthe diagnosis of bipolar disorder varies in different studiesaccording to their resources and requirements. Participants’strength, gender, age group, prior history of medication hasa great effect on the diagnosis of bipolar disorder; therefore,most of the studies consider it as their initial protocols require-ments as shown in Table 13. As far as participant strength isconcerned, [171, 172] include 38 subjects with the age groupin the range of 15 to 16 years and 18:20 ratio, which means 18subjects belongs to the bipolar disorder type I and 20 belongsto the bipolar disorder type II. No distinction about gender ismentioned in this study. To classify the unipolar, bipolar andhealthy subjects [57, 115, 116, 175] have performed EEGexperiments for 30 ,60, 89 and 134 subjects in which halfof the participants are depressed and the others are healthy.Participants age varies from 20 to 50 years with no discrimi-nation of gender. The classification of bipolar, schizophreniaand healthy subjects is performed in [174] in which 35 par-ticipants have schizophrenia, 35 are with bipolar illness andthe remaining 35 are healthy subjects. From the availableliterature of bipolar disorder diagnosis, it is observed thatthere is no standard ratio of number of participants and itvaries in different studies according to their resources and re-quirements of the experiment.
Selection Criteria of BipolarDisorder Participants:
The participants are selected basedon two major parameters of self-reported psychometric testand inclusion criteria of participants. For all subjects, theinclusion criteria are history of epilepsy, head injury, psychi-atric disorders and effect of illegal drugs as shown in Table14. The Diagnostic and Statistical Manual of Mental Disor-ders (DSM-IV) and Back Depression Inventory (BDI-II) arethe two major self-reported psychometric tests that are usedin bipolar studies [115, 116, 175, 174] for primary selectionof the participants. Based on inclusion criteria and psycho-metric test, only those participants are considered for EEGstudy that have no psychotic disorder and their self-reportedpsychometric test score is above 14. Interview by a psychi-atrist is also conducted in some studies [111] to ensure thatthe selected participants are taking no medication to treatbipolarity and depression. After the selection of the partici-pants, a ten minutes EEG experiment is performed with openand close eye conditions for each subject.
Placement and
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 20 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 13: Participants information in EEG experiments for bipolar disorder detection.Ref Subjects Gender Age Group(Mean,ÂśSD) Bipolarity[171] 38(18BDI,20BDII) 18F,20M 15-16 (15.7,1.5) Bipolar disorder I, II[172] 38(18bvFTD+20BD) - 52-77(64 years) Bipolar disorder[57] 89 - - Unipolar, bipolar[115] 60(30D+30H) 32F,28M 20-50 Bipolar depression[116] 30 16F,14M 20-50 Bipolar depression[175] 134 (75D+59H) - 18-54(35,4.2) Bipolar disorder[174] 105(35SZ+35BD+35H) - - Bipolar disorderTable 14: Participant selection and experimentation tasks in EEG studies for bipolar disorder detection.Ref EEG Selection Crite-ria Subject Inclu-sion criteria Tasks Bipolar Disordercriteria[171] No psychotic DSM-IV,BDI - (BDI-II score>14[172] Brain hospital patient DSM-5 - DSM-5 score[57] No mental disorder Watch Pics BDI(II)score 14-28[115] No mental disorder BDI-II - BDI(II)score 14-28[116] No neurological his-tory - -[175] Heroin addiction aa-tients DSM-IV Stay relaxedand awake DSM-IV score[174] Age no more than 65 DSM-IV - -
Types of EEG Electrodes:
The electrode placement andtypes of electrodes play a major role in EEG based bipolardisorder data acquisition. Minor mistakes in electrode place-ment pollute the overall EEG results that further affect theclassification of the EEG signals. The two major interna-tional electrode placement standards at the scalp are 10-10and 10-20 [176]. According to the literature presented inTable 15 most of the studies used international placementstandards [171, 172, 115] but few create their own electrodeplacement strategy [57, 116]. Wet and dry types of electrodesare used in EEG based studies to diagnose bipolar disorder.Wet electrodes are usually made of silver and silver chloridematerial and applied on a scalp by using the electrolytic gelmaterial that works as a conductor between the skin and thewet electrodes. Dry electrodes consist of a single metal andcan be directly placed on the scalp without the need to applythe conductive gel. Dry electrodes are the most efficient andeasy to use as compared to wet electrodes. As patients feelcomfortable with dry electrodes therefore latest technologiesprefer to use them [177].
Several public datasets, such as BioGPS [178], Bipo-lar Disorder Neuroimaging Database [179], Bipolar Disor-der Phenome Database [180] etc., exist for bipolar disorderrecognition. However, after an extensive literature survey,we found that no EEG based public data sets are available forbipolar disorder recognition research.
Automation of bipolar disorder assessment is a majorchallenge for the research community. The automatic meth-ods make the investigation procedure quick and easy. Inthese methods, features are automatically extracted from theEEG signals and then based on these features a mental disor-der is diagnosed. In this article, artificial neural network anddeep learning-based approaches are considered for automatedbipolar disorders diagnosis.
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 21 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 15: EEG devices,number of channels and their placement used for EEG experiment for bipolar disorder detection.Ref EEG Device Electrode Brain Lobes PlacementStandard Types ofElectrode[171] EEG cap 19 - 10-20 Wet[172] 19 Frontal andtemporal 10-20 Wet[57] Procomp 1(F4) - - Dry[115] - 24 - 10-20 Wet[116] - Two channelpair Left/right half - Wet[175] âĂIJNeuroscan/scan LTâĂİneuro-headset - - - -[174] - - - - -Table 16: Artifacts and noise filtering approaches in EEGbased bipolar disorder detection.Ref Artifact Filtering Technique[171] Bandpass, notch filter andvisual inspection[172] Bandpass filter at 0.15 to 70 Hz[57] Bandpass filter 40HZ[115] Total variation Filtering(TVP)[116] Notch filter at 50HZ[175] Independent component analysis[174] -
The pre-processing or artifact removal is a major part ofEEG data acquisition, except this activity no analysis canbe performed directly on the EEG data. It is necessary tofilter the EEG signals from the different physiological andnon-physiological artifacts and interference before using it.The researchers used different techniques and tools to man-ually or automatically remove the artifacts from the EEGsignals as shown in Table: 16. The band pass, notch andICA (Independent Component Analysis) are the importantfilters that are used in most of the bipolar disorder studies[171, 172, 57, 116, 175] to remove the noise, interferenceand physiological and non-physiological artifacts with 40 Hzto 70Hz frequency range. These filters remove the artifactsfrom EEG signals based on the frequency and amplitude levelof the EEG signals (upper and lower level of the signals). TheTotal Variation Filtering (TVP) and visual inspection are sig-nal filtering techniques that are used in some studies [115]and remove noise manually from the signals.
For the past few decades, neural networks are widely usedin EEG based studies due to their ability of self-learning andproducing the output that is not limited to the input providedto them. Different models of ANN are used in EEG basedbipolar disorder diagnosis and classification as shown in Ta- Table 17: Neural Network based approaches for EEG basedbipolar disorder detection and their accuracy level.Ref Model used Results(Accuracy)[171] MLP neural net-work 91.83%[172] ANN 76%[57] ANN 89.89%[115] FFNN and PNN 58.75%,98.75%[116] FNN 98.11%[175] ANN -[174] ANN 90%ble:17. The Multi-layer perceptron (MLP), Feed forwardneural network (FFNN), Probabilistic neural network (PNN)and Artificial neural network (ANN) are the models that areused in bipolar disorder studies [171, 172] with online fea-ture extraction mechanism. They achieve a 91.83%, 58.75%,98.75% and 90% classification accuracy respectively.
Bipolar disorder is often confused with major depressivedisorder and other mental disorders. It is characterized bypersistent depression and mania. The Complications andinterruptions in the diagnosis of bipolar disorder delay effec-tive treatment of patients. Bipolar disorder is misdiagnosedas recurrent MDD in 60% of patients that seek treatmentfor depression[181].To overcome all these limitations and toprovide effective diagnosis and treatment to bipolar patients,there is a need to use deep learning based EEG for bipolardisorder diagnosis.
8. Discussion
Our umbrella review provided an up-to-date overviewand synthesis of the literature of neural networks-based ap-proaches using EEG signals for the diagnosis of Major De-
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 22 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
Table 18: Bipolar disorder type and their criteria defined byDSM-IV.Bipolar DisorderType (DSM-IV) criteriaBipolar disorder typeI In bipolar disorder type I onefull episode of mania or di-verse episodes of mania anddepression can occurs.Bipolar disorder typeII In bipolar disorder type II ,nomanic episode, minimum onehypomanic episode and manydepressive episodes occurs.Cyclothymic disorder Many hypomanic and depres-sive episodes occurs.Bipolar disorder nototherwise specified Depressive and hypomanicepisodes changes rapidly.pressive Disorder and Bipolar Disorder. We included datafrom different sources like Healthy Brain Network (HBN),PREDICT, pubMed, IEEE explore, embase, google scholar,research gate and web of science etc. We noticed in this re-search area several limitations, which we discuss in the formof lists of points and to which we make suggestions for betterimprovement:
Few Existing Works:
Deep learning approaches for EEGbased depression diagnosis have been found in only 4-5 recentarticles from 2018-2020 [53, 54, 111]. In case of EEG basedbipolar disorder no deep learning based article exists for diag-nosis. Despite the few existing works in deep learning-basedapproaches for EEG depression and bipolar disorders analy-sis, deep learning seems to be promising in many researchareas. Deep learning can play a key role in developing a moreaccurate biomarkers.
Data Availability:
The significant issue regarding thedata is availability. The most of the techniques reviewed hereare relied on private datasets, very few public data sets existfor EEG based depression diagnosis due to the sensitive na-ture of depression data, and for privacy and confidentialityreasons. In bipolar case no public dataset is available for EEGbased bipolar disorder diagnosis. However, for collective in-novation and financial growth, there should be some standardpublic datasets. Open datasets provide new opportunities foradministrations to collaborate with people and evaluate pub-lic services by giving access to the data about those services.Further, these datasets can be useful to evaluate and validatemethodologies/approaches presented by different researchers.
AI for a psychiatry revolution:
EEG is a non-invasivestrong biomarker from the AI revolution, that mental healthand psychiatry does not benefit yet. Suggestion: This domainneeds to attract more attention from the scientific communityespecially the computer vision community to develop moreinnovative methods and applications for a better diagnosis of mental disorders.
Deep learning for more complex patterns:
The fea-ture extraction phase has a great effect on the precision andaccuracy level of the research. The processing is applied onlyon the data that is extracted from the feature extraction phase.However, the majority of the approaches reviewed here usedthe conventional machine learning approaches for featureextraction that does not predict how many features shouldbe extracted for a high classification accuracy. Suggestion:Feature extraction phase should be automatic and predictablefor higher classification ratio.
Need for Multimodal approaches:
The multimodal ap-proaches of EEG and its fusion with other modalities areless explored in literature, despite the fact that multimodalfusion performs better than unimodal approaches in manyother applications [182]. Suggestion: EEG based multimodeldeep learning approaches should be introduced in depressionand bipolar disorder detection to enhance the classificationaccuracy.
EEG signals and noise:
Most of the studies reviewedhere use conventional artifacts removal techniques likewise:visual inspection,notch filter, low,band and high pass filterto sanitize the polluted EEG signals.These conventional ap-proaches cannot filter the signals accurately and signals stillcarry a lot of noise, which yield a negative impact on de-pression detection/ classification accuracy [183]. Suggestion:The signal pre-processing or filtering can be automated byusing some deep learning approaches.
Deep learning and interpretability:
The problem ofinterpretability of the results of neural networks based ap-proaches still exist. These neural network approaches can beregarded as black boxes. Without the interpretability of thesealgorithm results, it can be difficult for a physician to tell apatient about their depression/disorder and not providing anygenuine reason. Decision-making in such uncertain situationsis a problem of liability, ethics and beyond pure performance.Therefore, the final decision of algorithm should be inter-pretable [184]. Suggestion:The neural networks should beinterpretable to enhance the understanding level and to workin uncertain scenarios.
Brain Lobes:
More than 90% of reviewed studies usethe frontal lobe and left/right hemisphere for depression andbipolar disorder detection as given in Table 9, and achieve aremarkable classification accuracy. It can be a strong moti-vation for the new researchers for future analysis on frontallobe.
9. Conclusion
Mental disorders are highly prevalent and disabling healthcondition. Numerous studies explored the use of EEG sig-nals to diagnose the functioning of brain activity. In thissurvey paper, the focus is giving to EEG signals as a strongbiomarker for Major Depressive Disorder and Bipolar Disor-der and an extensive study of the state-of-the-art shallow anddeep neural networks based methods is giving for MDD andBD diagnosis and assessment. While EEG based methods
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 23 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review for MDD diagnosis attracts the attention of the computer vi-sion community, the EEG based methods for BD diagnosis isless explored in the literature and needs more consideration.The EEG based experimental protocols and methods couldhelp the scientific community to better understand mentaldisorders and to design strong biomarkers for their diagnosisand assessment.Deep neural networks offers high classification accuracyamong the depressed and healthy control subjects in com-parison to shallow neural networks based methods. Severalclinical research issues remain to be addressed scientificallyin this field. Thus, a set of recommendations are addressedin the discussion section to offer ways to guide the emergingcollaborations and interactions towards the most fruitful out-comes.In conclusion, the exhaustive review of available evidencehighlights the considerable potential in the application ofEEG-based methods for the assessment and monitoring ofMDD and BD. However, despite the good prediction perfor-mance of neural networks based methods, they lack sufficientmodel explainability that has shown its impact in areas suchas quantitative research and then prevent the community tofurther develop reproducible and deterministic protocols andto achieve clinically useful results. Thus, explainable neuralnetworks based methods are needed for mental health diag-nosis and assessment. Finally, we hope this paper is anotherstep towards harnessing the full potential of AI for mentalhealth diagnosis.
Declaration of Competing Interest
The author(s) declare(s) that there is no conflict of inter-est.
References [1] W. Chow, M. Doane, J. Sheehan, L. Alphs, H. Le, Le h. economicburden among patients with major depressive disorder: an analysisof healthcare resource use, work productivity, and direct and indirectcosts by depression severity, Am J Manag Care 16 (2019) e188–e196.[2] N. Sartorius, Depression and diabetes, Dialogues in clinical neuro-science 20 (1) (2018) 47.[3] T. T. T. Tran, N. B. Nguyen, M. A. Luong, T. H. A. Bui, T. D.Phan, T. H. Ngo, H. Minas, T. Q. Nguyen, et al., Stress, anxiety anddepression in clinical nurses in vietnam: a cross-sectional survey andcluster analysis, International journal of mental health systems 13 (1)(2019) 3.[4] K.-M. Han, D. De Berardis, M. Fornaro, Y.-K. Kim, Differentiatingbetween bipolar and unipolar depression in functional and structuralmri studies, Progress in Neuro-Psychopharmacology and BiologicalPsychiatry 91 (2019) 20–27.[5] A. Pampouchidou, P. Simos, K. Marias, F. Meriaudeau, F. Yang,M. Pediaditis, M. Tsiknakis, Automatic assessment of depressionbased on visual cues: A systematic review, IEEE Transactions onAffective Computing (2017).[6] F. Koyama, T. Yoda, T. Hirao, Insomnia and depression: Japanesehospital workers questionnaire survey, Open Medicine 12 (1) (2017)391–398.[7] A. Zafar, S. Chitnis, Survey of depression detection using socialnetworking sites via data mining, in: 2020 10th International Confer-ence on Cloud Computing, Data Science & Engineering (Confluence),IEEE, 2020, pp. 88–93. [8] U. R. Acharya, V. K. Sudarshan, H. Adeli, J. Santhosh, J. E. Koh,A. Adeli, Computer-aided diagnosis of depression using eeg signals,European neurology 73 (5-6) (2015) 329–336.[9] F. Vellante, F. Ferri, G. Baroni, P. Croce, D. Migliorati, M. Pettoruso,D. De Berardis, G. Martinotti, F. Zappasodi, M. Di Giannantonio,Euthymic bipolar disorder patients and eeg microstates: a neuralsignature of their abnormal self experience?, Journal of AffectiveDisorders (2020).[10] Y. Wang, B. McCane, N. McNaughton, Z. Huang, P. Neo, et al., Anxi-etydecoder: An eeg-based anxiety predictor using a 3-d convolutionalneural network, in: 2019 International Joint Conference on NeuralNetworks (IJCNN), IEEE, 2019, pp. 1–8.[11] M. Hébert, C. Mérette, A.-M. Gagné, T. Paccalet, I. Moreau, J. Lavoie,M. Maziade, The electroretinogram may differentiate schizophreniafrom bipolar disorder, Biological psychiatry 87 (3) (2020) 263–270.[12] S. I. Dimitriadis, C. I. Salis, D. Liparas, A sleep disorder detectionmodel based on eeg cross-frequency coupling and random forest,medRxiv (2020).[13] S. Mahato, S. Paul, Electroencephalogram (eeg) signal analysis fordiagnosis of major depressive disorder (mdd): a review, in: Nano-electronics, Circuits and Communication Systems, Springer, 2019,pp. 323–335.[14] S. Phadikar, N. Sinha, R. Ghosh, Automatic eye blink artifact re-moval from eeg signal using wavelet transform with heuristicallyoptimized threshold, IEEE Journal of Biomedical and Health Infor-matics (2020).[15] H. Grunze, Bipolar disorder, in: Neurobiology of brain disorders,Elsevier, 2015, pp. 655–673.[16] M. L. Phillips, D. J. Kupfer, Bipolar disorder diagnosis: challengesand future directions, The Lancet 381 (9878) (2013) 1663–1671.[17] M. Č. Radenković, V. L. Lopez, Machine learning approaches fordetecting the depression from resting-state electroencephalogram(eeg): A review study, arXiv preprint arXiv:1909.03115 (2019).[18] A. Malviya, R. Meharkure, R. Narsinghani, V. Sheth, P. Meshram,Depression detection through speech analysis: A survey, InternationalJournal of Scientific Research in Computer Science, Engineering andInformation Technology (2019) 712–716.[19] J. Oh, K. Yun, U. Maoz, T.-S. Kim, J.-H. Chae, Identifying depressionin the national health and nutrition examination survey data using adeep learning algorithm, Journal of affective disorders 257 (2019)623–631.[20] A. Kerst, J. Zielasek, W. Gaebel, Smartphone applications for de-pression: a systematic literature review and a survey of health careprofessionalsâĂŹ attitudes towards their use in clinical practice, Euro-pean archives of psychiatry and clinical neuroscience 270 (2) (2020)139–152.[21] X. Zhang, L. Yao, X. Wang, J. Monaghan, D. Mcalpine, Y. Zhang,A survey on deep learning based brain computer interface: Recentadvances and new frontiers, arXiv preprint arXiv:1905.04149 (2019).[22] M. Iftikhar, S. A. Khan, A. Hassan, A survey of deep learning andtraditional approaches for eeg signal processing and classification,in: 2018 IEEE 9th Annual Information Technology, Electronics andMobile Communication Conference (IEMCON), IEEE, 2018, pp.395–400.[23] L. Bozhkov, P. Georgieva, Overview of deep learning architecturesfor eeg-based brain imaging, in: 2018 International Joint Conferenceon Neural Networks (IJCNN), IEEE, 2018, pp. 1–7.[24] A. S. Fathima, S. Mythili, Deep learning technique for feature classi-fication of eeg to access studentâĂŹs mental status: A survey, Inter-national Research Journal of Engineering and Technology (IRJET)(2019).[25] A. Saikia, S. Paul, Application of deep learning for eeg, in: Handbookof Research on Advancements of Artificial Intelligence in HealthcareEngineering, IGI Global, 2020, pp. 106–123.[26] W. Liu, Z. Wang, X. Liu, N. Zeng, Y. Liu, F. E. Alsaadi, A survey ofdeep neural network architectures and their applications, Neurocom-puting 234 (2017) 11–26.[27] S. Pouyanfar, S. Sadiq, Y. Yan, H. Tian, Y. Tao, M. P. Reyes, M.-
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 24 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review
L. Shyu, S.-C. Chen, S. Iyengar, A survey on deep learning: Al-gorithms, techniques, and applications, ACM Computing Surveys(CSUR) 51 (5) (2018) 1–36.[28] J. Pamina, B. Raja, Survey on deep learning algorithms, InternationalJournal of Emerging Technology and Innovative Engineering 5 (1)(2019).[29] C. Tan, F. Sun, T. Kong, W. Zhang, C. Yang, C. Liu, A survey ondeep transfer learning, in: International conference on artificial neuralnetworks, Springer, 2018, pp. 270–279.[30] M. Z. Alom, T. M. Taha, C. Yakopcic, S. Westberg, P. Sidike, M. S.Nasrin, M. Hasan, B. C. Van Essen, A. A. Awwal, V. K. Asari, Astate-of-the-art survey on deep learning theory and architectures,Electronics 8 (3) (2019) 292.[31] H.-J. Hwang, S. Kim, S. Choi, C.-H. Im, Eeg-based brain-computer in-terfaces: a thorough literature survey, International Journal of Human-Computer Interaction 29 (12) (2013) 814–826.[32] M. Rashid, N. Sulaiman, M. Mustafa, S. Khatun, B. S. Bari, M. J.Hasan, Recent trends and open challenges in eeg based brain-computer interface systems, in: InECCE2019, Springer, 2020, pp.367–378.[33] A. Khosla, P. Khandnor, T. Chand, A comparative analysis of signalprocessing and classification methods for different applications basedon eeg signals, Biocybernetics and Biomedical Engineering (2020).[34] S. Bhattacharjee, S. Ghatak, S. Dutta, B. Chatterjee, M. Gupta, Asurvey on comparison analysis between eeg signal and mri for brainstroke detection, in: Emerging Technologies in Data Mining andInformation Security, Springer, 2019, pp. 377–382.[35] X. Gu, Z. Cao, A. Jolfaei, P. Xu, D. Wu, T.-P. Jung, C.-T. Lin, Eeg-based brain-computer interfaces (bcis): A survey of recent studieson signal sensing technologies and computational intelligence ap-proaches and their applications, arXiv preprint arXiv:2001.11337(2020).[36] A. Alhassan, A. R. Ziblim, S. Muntaka, A survey on depressionamong infertile women in ghana, BMC women’s health 14 (1) (2014)1–6.[37] F. Koyama, T. Yoda, T. Hirao, Insomnia and depression: Japanesehospital workers questionnaire survey, Open Medicine 12 (1) (2017)391–398.[38] K. Heiden-Rootes, A. Wiegand, D. Thomas, R. M. Moore, K. A.Ross, A national survey on depression, internalized homophobia,college religiosity, and climate of acceptance on college campusesfor sexual minority adults, Journal of Homosexuality 67 (4) (2020)435–451.[39] J. N. Miller, D. W. Black, Bipolar disorder and suicide: A review,Current psychiatry reports 22 (2) (2020) 6.[40] Y.-M. Bai, M.-H. Chen, J.-W. Hsu, K.-L. Huang, P.-C. Tu, W.-C.Chang, T.-P. Su, C. T. Li, W.-C. Lin, S.-J. Tsai, A comparison study ofmetabolic profiles, immunity, and brain gray matter volumes betweenpatients with bipolar disorder and depressive disorder, Journal ofNeuroinflammation 17 (1) (2020) 42.[41] R. Degabriele, J. Lagopoulos, A review of eeg and erp studies inbipolar disorder, Acta Neuropsychiatrica 21 (2) (2009) 58–66.[42] G. Zhang, W. Wang, Z. Yang, S. Zhan, F. Sun, Introduction to prisma-ci extension statement and checklist systematic reviews on complexinterventions, Zhonghua liu xing bing xue za zhi= Zhonghua liux-ingbingxue zazhi 40 (7) (2019) 832–838.[43] M. Staples, M. Niazi, Experiences using systematic review guidelines,Journal of Systems and Software 80 (9) (2007) 1425–1437.[44] D. Strech, N. Sofaer, How to write a systematic review of reasons,Journal of Medical Ethics 38 (2) (2012) 121–126.[45] D. P. Subha, P. K. Joseph, R. Acharya, C. M. Lim, Eeg signal analysis:a survey, Journal of medical systems 34 (2) (2010) 195–212.[46] X. Ma, H. Yang, Q. Chen, D. Huang, Y. Wang, Depaudionet: Anefficient deep model for audio based depression classification, in: Pro-ceedings of the 6th international workshop on audio/visual emotionchallenge, 2016, pp. 35–42.[47] M. Nasir, A. Jati, P. G. Shivakumar, S. Nallan Chakravarthula, P. Geor-giou, Multimodal and multiresolution depression detection from speech and facial landmark features, in: Proceedings of the 6th in-ternational workshop on audio/visual emotion challenge, 2016, pp.43–50.[48] B. Sun, Y. Zhang, J. He, L. Yu, Q. Xu, D. Li, Z. Wang, A randomforest regression method with selected-text feature for depressionassessment, in: Proceedings of the 7th Annual Workshop on Au-dio/Visual Emotion Challenge, 2017, pp. 61–68.[49] B. Goudiaby, A. Othmani, A. Nait-Ali, Eeg biometrics for personverification, in: Hidden Biometrics, Springer, 2020, pp. 45–69.[50] S. Siuly, Y. Li, Y. Zhang, Electroencephalogram (eeg) and its back-ground, in: EEG Signal Analysis and Classification, Springer, 2016,pp. 3–21.[51] N. Sharma, T. Gedeon, Objective measures, sensors and computa-tional techniques for stress recognition and classification: A survey,Computer methods and programs in biomedicine 108 (3) (2012)1287–1301.[52] C. S. Nayak, A. C. Anilkumar, Eeg normal waveforms, in: StatPearls[Internet], StatPearls Publishing, 2019.[53] W. Mumtaz, A. Qayyum, A deep learning framework for automaticdiagnosis of unipolar depression, International journal of medicalinformatics 132 (2019) 103983.[54] B. Ay, O. Yildirim, M. Talo, U. B. Baloglu, G. Aydin, S. D. Puthankat-til, U. R. Acharya, Automated depression detection using deep repre-sentation and sequence learning with eeg signals, Journal of medicalsystems 43 (7) (2019) 205.[55] X. Li, X. Zhang, J. Zhu, W. Mao, S. Sun, Z. Wang, C. Xia, B. Hu,Depression recognition using machine learning methods with differ-ent feature generation strategies, Artificial intelligence in medicine99 (2019) 101696.[56] B. Mohammadzadeh, M. Khodabandelu, M. Lotfizadeh, Comparingdiagnosis of depression in depressed patients by eeg, based on twoalgorithms: Artificial nerve networks and neuro-fuzy networks, Inter-national Journal of Epidemiologic Research 3 (3) (2016) 246–258.[57] T. T. Erguzel, G. H. Sayar, N. Tarhan, Artificial intelligence approachto classify unipolar and bipolar depressive disorders, Neural Com-puting and Applications 27 (6) (2016) 1607–1616.[58] S. M. Alarcao, M. J. Fonseca, Emotions recognition using eeg signals:A survey, IEEE Transactions on Affective Computing 10 (3) (2017)374–393.[59] T. Alotaiby, F. E. Abd El-Samie, S. A. Alshebeili, I. Ahmad, A reviewof channel selection algorithms for eeg signal processing, EURASIPJournal on Advances in Signal Processing 2015 (1) (2015) 66.[60] Y. Mohan, S. S. Chee, D. K. P. Xin, L. P. Foong, Artificial neuralnetwork for classification of depressive and normal in eeg, in: 2016IEEE EMBS conference on biomedical engineering and sciences(IECBES), IEEE, 2016, pp. 286–290.[61] T. J. L. Erica Cirino, The effects of depression on the brain (February8, 2017).URL [62] G. E. Simon, M. VonKorff, M. Piccinelli, C. Fullerton, J. Ormel,An international study of the relation between somatic symptomsand depression, New England journal of medicine 341 (18) (1999)1329–1335.[63] T. M. H., The link between depression and physical symptoms, Pri-mary care companion to the Journal of clinical psychiatry (2004)12âĂŞ16.[64] G. G. M, Depression and associated physical diseases and symptoms,Dialogues in clinical neuroscience 8(2) (2006) 259âĂŞ265.[65] M. Guha, Diagnostic and statistical manual of mental disorders: Dsm-5, Reference Reviews (2014).[66] A. Pampouchidou, P. Simos, K. Marias, F. Meriaudeau, F. Yang,M. Pediaditis, M. Tsiknakis, Automatic assessment of depressionbased on visual cues: A systematic review, IEEE Transactions onAffective Computing (2017).[67] N. R. Council, et al., Depression in parents, parenting, and children:Opportunities to improve identification, treatment, and prevention,National Academies Press, 2009.
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 25 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review [68] A. Rapkin, A review of treatment of premenstrual syndrome & pre-menstrual dysphoric disorder, Psychoneuroendocrinology 28 (2003)39–53.[69] G. Serra, A. Koukopoulos, L. De Chiara, A. E. Koukopoulos, G. Sani,L. Tondo, P. Girardi, D. Reginaldi, R. J. Baldessarini, Early clinicalpredictors of long-term morbidity in major depressive disorder, Earlyintervention in psychiatry 13 (4) (2019) 999–1002.[70] A. P. Association, A. P. Association, et al., Diagnostic and statisticalmanual of mental disorders: Dsm-5 (2013).[71] K. Kroenke, R. L. Spitzer, J. B. Williams, The phq-9: validityof a brief depression severity measure, Journal of general internalmedicine 16 (9) (2001) 606–613.[72] H. Cai, J. Han, Y. Chen, X. Sha, Z. Wang, B. Hu, J. Yang, L. Feng,Z. Ding, Y. Chen, et al., A pervasive approach to eeg-based depressiondetection, Complexity 2018 (2018).[73] W. Mao, J. Zhu, X. Li, X. Zhang, S. Sun, Resting state eeg baseddepression recognition research using deep learning method, in: In-ternational Conference on Brain Informatics, Springer, 2018, pp.329–338.[74] A. T. Beck, R. A. Steer, G. K. Brown, et al., Beck depressioninventory-ii, San Antonio 78 (2) (1996) 490–498.[75] X. Li, R. La, Y. Wang, J. Niu, S. Zeng, S. Sun, J. Zhu, Eeg-based milddepression recognition using convolutional neural network, Medical& biological engineering & computing 57 (6) (2019) 1341–1352.[76] A. P. Association, et al., American psychiatric association: Diag-nostic and statistical manual of mental disorders, (p. 81), Arlington:American Psychiatric Association (2013).[77] T. T. Erguzel, S. Ozekes, S. Gultekin, N. Tarhan, G. H. Sayar,A. Bayram, Neural network based response prediction of rtms inmajor depressive disorder using qeeg cordance, Psychiatry investiga-tion 12 (1) (2015) 61.[78] M. Ahmadlou, H. Adeli, A. Adeli, Fractality analysis of frontal brainin major depressive disorder, International Journal of Psychophysiol-ogy 85 (2) (2012) 206–211.[79] L. S. Radloff, A self-report depression scale for research in the generalpopulation, Applied psychol Measurements 1 (1977) 385–401.[80] R. M. Saracino, H. Cham, B. Rosenfeld, C. J. Nelson, Confirmatoryfactor analysis of the center for epidemiologic studies depressionscale in oncology with examination of invariance between youngerand older patients., European Journal of Psychological Assessment(2018).[81] N. Timmerby, J. H. Andersen, S. Søndergaard, S. D. Østergaard,P. Bech, A systematic review of the clinimetric properties of the6-item version of the hamilton depression rating scale (ham-d6),Psychotherapy and psychosomatics 86 (3) (2017) 141–149.[82] B. D. Nelson, E. M. Kessel, D. N. Klein, S. A. Shankman, Depressionsymptom dimensions and asymmetrical frontal cortical activity whileanticipating reward, Psychophysiology 55 (1) (2018) e12892.[83] R. Nusslock, K. Walden, E. Harmon-Jones, Asymmetrical frontalcortical activity associated with differential risk for mood and anxietydisorder symptoms: An rdoc perspective, International Journal ofPsychophysiology 98 (2) (2015) 249–261.[84] S. M. Eack, C. G. Greeno, B.-J. Lee, Limitations of the patient healthquestionnaire in identifying anxiety and depression in communitymental health: many cases are undetected, Research on social workpractice 16 (6) (2006) 625–631.[85] K. Kroenke, R. L. Spitzer, The phq-9: a new depression diagnosticand severity measure, Psychiatric annals 32 (9) (2002) 509–515.[86] N. Hagiwara, Y. Omiya, S. Shinohara, M. Nakamura, M. Higuchi,S. Mitsuyoshi, H. Yasunaga, S. Tokuno, Validity of mind monitoringsystem as a mental health indicator using voice, Adv. Sci. Technol.Eng. Syst. J 2 (3) (2017) 338–344.[87] B. D. W. Group, A. J. Atkinson Jr, W. A. Colburn, V. G. DeGruttola,D. L. DeMets, G. J. Downing, D. F. Hoth, J. A. Oates, C. C. Peck,R. T. Schooley, et al., Biomarkers and surrogate endpoints: preferreddefinitions and conceptual framework, Clinical pharmacology &therapeutics 69 (3) (2001) 89–95.[88] B. W. Dunlop, H. S. Mayberg, Neuroimaging-based biomarkers for treatment selection in major depressive disorder, Dialogues in clinicalneuroscience 16 (4) (2014) 479.[89] M. Shadrina, E. A. Bondarenko, P. A. Slominsky, Genetics factors inmajor depression disease, Frontiers in psychiatry 9 (2018) 334.[90] T. J. L. Stephanie Faris, Is depression genetic? (July 25, 2017).URL [91] W. Mumtaz, A. S. Malik, M. A. M. Yasin, L. Xia, Review on eeg anderp predictive biomarkers for major depressive disorder, BiomedicalSignal Processing and Control 22 (2015) 85–98.[92] S. M. Tripathi, N. Mishra, R. K. Tripathi, K. Gurnani, P300 latencyas an indicator of severity in major depressive disorder, Industrialpsychiatry journal 24 (2) (2015) 163.[93] C. Wyss, The ldaep as a potential biomarker for central serotonergicactivity: challenges to overcome, Ph.D. thesis, University of Zurich(2015).[94] A. Steiger, M. Kimura, Wake and sleep eeg provide biomarkers indepression, Journal of psychiatric research 44 (4) (2010) 242–252.[95] P. L. Franzen, D. J. Buysse, Sleep disturbances and depression: riskrelationships for subsequent depression and therapeutic implications,Dialogues in clinical neuroscience 10 (4) (2008) 473.[96] N. Van Der Vinne, M. A. Vollebregt, M. J. Van Putten, M. Arns,Frontal alpha asymmetry as a diagnostic marker in depression: Factor fiction? a meta-analysis, Neuroimage: clinical 16 (2017) 79–87.[97] A. Dharmadhikari, A. Tandle, S. Jaiswal, V. Sawant, V. Vahia, N. Jog,et al., Frontal theta asymmetry as a biomarker of depression, EastAsian Archives of Psychiatry 28 (1) (2018) 17.[98] A. M. Hunter, T. X. Nghiem, I. A. Cook, D. E. Krantz, M. J. Minzen-berg, A. F. Leuchter, Change in quantitative eeg theta cordance asa potential predictor of repetitive transcranial magnetic stimulationclinical outcome in major depressive disorder, Clinical EEG andNeuroscience 49 (5) (2018) 306–315.[99] P. J. Fitzgerald, B. O. Watson, Gamma oscillations as a biomarker formajor depression: an emerging topic, Translational psychiatry 8 (1)(2018) 1–7.[100] S. Sun, J. Li, H. Chen, T. Gong, X. Li, B. Hu, A study of resting-state eeg biomarkers for depression recognition, arXiv preprintarXiv:2002.11039 (2020).[101] P. C. Koo, C. Berger, G. Kronenberg, J. Bartz, P. Wybitul, O. Reis,J. Hoeppner, Combined cognitive, psychomotor and electrophysio-logical biomarkers in major depressive disorder, European archivesof psychiatry and clinical neuroscience 269 (7) (2019) 823–832.[102] P. Fernández-Palleiro, T. Rivera-Baltanás, D. Rodrigues-Amorim,S. Fernández-Gil, M. del Carmen Vallejo-Curto, M. Álvarez-Ariza,M. López, C. Rodriguez-Jamardo, J. Luis Benavente, E. de Las Heras,et al., Brainwaves oscillations as a potential biomarker for major de-pression disorder risk, Clinical EEG and Neuroscience 51 (1) (2020)3–9.[103] M. M. Caudill, A. M. Hunter, I. A. Cook, A. F. Leuchter, The an-tidepressant treatment response index as a predictor of reboxetinetreatment outcome in major depressive disorder, Clinical EEG andneuroscience 46 (4) (2015) 277–284.[104] C. DeBattista, G. Kinrys, D. Hoffman, C. Goldstein, J. Zajecka,J. Kocsis, M. Teicher, S. Potkin, A. Preda, G. Multani, et al., Theuse of referenced-eeg (reeg) in assisting medication selection for thetreatment of depression, Journal of psychiatric research 45 (1) (2011)64–75.[105] M. Arns, A. Etkin, U. Hegerl, L. M. Williams, C. DeBattista, D. M.Palmer, P. B. Fitzgerald, A. Harris, R. deBeuss, E. Gordon, Frontaland rostral anterior cingulate (racc) theta eeg in depression: Implica-tions for treatment outcome?, European Neuropsychopharmacology25 (8) (2015) 1190–1200.[106] A. Wichniak, A. Wierzbicka, W. Jernajczyk, Sleep as a biomarker fordepression, International review of psychiatry 25 (5) (2013) 632–645.[107] S. Sun, J. Li, H. Chen, T. Gong, X. Li, B. Hu, A study of resting-state eeg biomarkers for depression recognition, arXiv preprintarXiv:2002.11039 (2020).[108] Z. S. Syed, K. Sidorov, D. Marshall, Depression severity predictionbased on biomarkers of psychomotor retardation, in: Proceedings
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 26 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review of the 7th Annual Workshop on Audio/Visual Emotion Challenge,2017, pp. 37–43.[109] O. J. Robinson, B. J. Sahakian, Cognitive biomarkers in depression.(2013).[110] J. R. Williamson, T. F. Quatieri, B. S. Helfer, G. Ciccarelli, D. D.Mehta, Vocal and facial biomarkers of depression based on motorincoordination and timing, in: Proceedings of the 4th InternationalWorkshop on Audio/Visual Emotion Challenge, 2014, pp. 65–72.[111] U. R. Acharya, S. L. Oh, Y. Hagiwara, J. H. Tan, H. Adeli, D. P.Subha, Automated eeg-based screening of depression using deepconvolutional neural network, Computer methods and programs inbiomedicine 161 (2018) 103–113.[112] P. Sandheep, S. Vineeth, M. Poulose, D. Subha, Performance analysisof deep learning cnn in classification of depression eeg signals, in:TENCON 2019-2019 IEEE Region 10 Conference (TENCON), IEEE,2019, pp. 1339–1344.[113] H. Jebelli, M. M. Khalili, S. Lee, Mobile eeg-based workersâĂŹstress recognition by applying deep neural network, in: Advances inInformatics and Computing in Civil and Construction Engineering,Springer, 2019, pp. 173–180.[114] Y. Mohammadi, M. Hajian, M. H. Moradi, Discrimination of depres-sion levels using machine learning methods on eeg signals, in: 201927th Iranian Conference on Electrical Engineering (ICEE), IEEE,2019, pp. 1765–1769.[115] S. D. Puthankattil, P. K. Joseph, Half-wave segment feature extractionof eeg signals of patients with depression and performance evaluationof neural network classifiers, Journal of Mechanics in Medicine andBiology 17 (01) (2017) 1750006.[116] S. D. Puthankattil, P. K. Joseph, Classification of eeg signals innormal and depression conditions by ann using rwe and signal en-tropy, Journal of Mechanics in Medicine and biology 12 (04) (2012)1240019.[117] P. Sandheep, S. Vineeth, M. Poulose, D. Subha, Performance analysisof deep learning cnn in classification of depression eeg signals, in:TENCON 2019-2019 IEEE Region 10 Conference (TENCON), IEEE,2019, pp. 1339–1344.[118] H. Mallikarjun, H. Suresh, Depression level prediction using eeg sig-nal processing, in: 2014 International Conference on ContemporaryComputing and Informatics (IC3I), IEEE, 2014, pp. 928–933.[119] J. Zhu, Y. Wang, R. La, J. Zhan, J. Niu, S. Zeng, X. Hu, Multi-modal mild depression recognition based on eeg-em synchronizationacquisition network, IEEE Access 7 (2019) 28196–28210.[120] D. Shah, G. Y. Wang, M. Doborjeh, Z. Doborjeh, N. Kasabov, Deeplearning of eeg data in the neucube brain-inspired spiking neuralnetwork architecture for a better understanding of depression, in:International Conference on Neural Information Processing, Springer,2019, pp. 195–206.[121] X. Zhang, B. Hu, L. Zhou, P. Moore, J. Chen, An eeg based pervasivedepression detection for females, in: Joint International Conferenceon Pervasive Computing and the Networked World, Springer, 2012,pp. 848–861.[122] H. Cai, X. Sha, X. Han, S. Wei, B. Hu, Pervasive eeg diagnosisof depression using deep belief network with three-electrodes eegcollector, in: 2016 IEEE International Conference on Bioinformaticsand Biomedicine (BIBM), IEEE, 2016, pp. 1239–1246.[123] O. Faust, P. C. A. Ang, S. D. Puthankattil, P. K. Joseph, Depressiondiagnosis support system based on eeg signal entropies, Journal ofmechanics in medicine and biology 14 (03) (2014) 1450035.[124] S. D. Kumar, D. Subha, Prediction of depression from eeg signalusing long short term memory (lstm), in: 2019 3rd InternationalConference on Trends in Electronics and Informatics (ICOEI), IEEE,2019, pp. 1248–1253.[125] S. Mahato, S. Paul, Detection of major depressive disorder usinglinear and non-linear features from eeg signals, Microsystem Tech-nologies 25 (3) (2019) 1065–1076.[126] X. Zhang, B. Hu, L. Zhou, P. Moore, J. Chen, An eeg based pervasivedepression detection for females, in: Joint International Conferenceon Pervasive Computing and the Networked World, Springer, 2012, pp. 848–861.[127] Z. Wan, J. Huang, H. Zhang, H. Zhou, J. Yang, N. Zhong, Hybrideeg-net: A convolutional neural network for eeg feature learning anddepression discrimination, IEEE Access 8 (2020) 30332–30342.[128] H. Mallikarjun, H. Suresh, Depression level prediction using eeg sig-nal processing, in: 2014 International Conference on ContemporaryComputing and Informatics (IC3I), IEEE, 2014, pp. 928–933.[129] X. Li, R. La, Y. Wang, J. Niu, S. Zeng, S. Sun, J. Zhu, Eeg-based milddepression recognition using convolutional neural network, Medical& biological engineering & computing 57 (6) (2019) 1341–1352.[130] S. Mahato, S. Paul, Detection of major depressive disorder usinglinear and non-linear features from eeg signals, Microsystem Tech-nologies 25 (3) (2019) 1065–1076.[131] H. Cai, Z. Wang, Y. Zhang, Y. Chen, B. Hu, A virtual-reality basedneurofeedback game framework for depression rehabilitation usingpervasive three-electrode eeg collector, in: Proceedings of the 12thChinese Conference on Computer Supported Cooperative Work andSocial Computing, 2017, pp. 173–176.[132] Y.-j. Li, F.-y. Fan, Classification of schizophrenia and depression byeeg with anns, in: 2005 IEEE Engineering in Medicine and Biology27th Annual Conference, IEEE, 2006, pp. 2679–2682.[133] H. Peng, B. Hu, Q. Liu, Q. Dong, Q. Zhao, P. Moore, User-centereddepression prevention: An eeg approach to pervasive healthcare, in:2011 5th International Conference on Pervasive Computing Tech-nologies for Healthcare (PervasiveHealth) and Workshops, IEEE,2011, pp. 325–330.[134] Y. Katyal, S. V. Alur, S. Dwivedi, R. Menaka, Eeg signal and videoanalysis based depression indication, in: 2014 IEEE InternationalConference on Advanced Communications, Control and ComputingTechnologies, IEEE, 2014, pp. 1353–1360.[135] Y. Mohan, S. S. Chee, D. K. P. Xin, L. P. Foong, Artificial neuralnetwork for classification of depressive and normal in eeg, in: 2016IEEE EMBS conference on biomedical engineering and sciences(IECBES), IEEE, 2016, pp. 286–290.[136] H. Kwon, S. Kang, W. Park, J. Park, Y. Lee, Deep learning basedpre-screening method for depression with imagery frontal eeg chan-nels, International Conference on Information and CommunicationTechnology Convergence (ICTC) (2019).[137] I.-M. Spyrou, C. Frantzidis, C. Bratsas, I. Antoniou, P. D. Bamidis,Geriatric depression symptoms coexisting with cognitive decline:a comparison of classification methodologies, Biomedical SignalProcessing and Control 25 (2016) 118–129.[138] J. Zhu, Y. Wang, R. La, J. Zhan, J. Niu, S. Zeng, X. Hu, Multi-modal mild depression recognition based on eeg-em synchronizationacquisition network, IEEE Access 7 (2019) 28196–28210.[139] L. LIU, R. ZHOU, An important neural indicator of measuring de-pression: The asymmetry of resting frontal activity, Advances inPsychological Science 23 (6) (2015) 1000–1008.[140] W. Wu, Y. Zhang, J. Jiang, M. V. Lucas, G. A. Fonzo, C. E. Rolle,C. Cooper, C. Chin-Fatt, N. Krepel, C. A. Cornelssen, et al., Anelectroencephalographic signature predicts antidepressant responsein major depression, Nature biotechnology 38 (4) (2020) 439–447.[141] E. Garcia-Ceja, M. Riegler, P. Jakobsen, J. Tørresen, T. Nordgreen,K. J. Oedegaard, O. B. Fasmer, Depresjon: a motor activity databaseof depression episodes in unipolar and bipolar patients, in: Proceed-ings of the 9th ACM Multimedia Systems Conference, 2018, pp.472–477.[142] H. Kristjánsdóttir, P. M. Salkovskis, B. H. Sigurdsson, E. Sigurds-son, A. Agnarsdóttir, J. F. Sigurdsson, Transdiagnostic cognitivebehavioural treatment and the impact of co-morbidity: An open trialin a cohort of primary care patients, Nordic journal of psychiatry70 (3) (2016) 215–223.[143] N. Langer, E. J. Ho, L. M. Alexander, H. Y. Xu, R. K. Jozanovic,S. Henin, A. Petroni, S. Cohen, E. T. Marcelle, L. C. Parra, et al., Aresource for assessing information processing in the developing brainusing eeg and eye tracking, Scientific data 4 (1) (2017) 1–20.[144] J. F. Cavanagh, A. Napolitano, C. Wu, A. Mueen, The patient reposi-tory for eeg data+ computational tools (pred+ ct), Frontiers in neu-
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 27 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review roinformatics 11 (2017) 67.[145] L. M. Alexander, J. Escalera, L. Ai, C. Andreotti, K. Febre, A. Man-gone, N. Vega-Potler, N. Langer, A. Alexander, M. Kovacs, S. Litke,B. O’Hagan, J. Andersen, B. Bronstein, A. Bui, M. Bushey, H. Butler,V. Castagna, N. Camacho, E. Chan, D. Citera, J. Clucas, S. Co-hen, S. Dufek, M. Eaves, B. Fradera, J. Gardner, N. Grant-Villegas,G. Green, C. Gregory, E. Hart, S. Harris, M. Horton, D. Kahn,K. Kabotyanski, B. Karmel, S. P. Kelly, K. Kleinman, B. Koo,E. Kramer, E. Lennon, C. Lord, G. Mantello, A. Margolis, K. R.Merikangas, J. Milham, G. Minniti, R. Neuhaus, A. Levine, Y. Os-man, L. C. Parra, K. R. Pugh, A. Racanello, A. Restrepo, T. Saltzman,B. Septimus, R. Tobe, R. Waltz, A. Williams, A. Yeo, F. X. Castel-lanos, A. Klein, T. Paus, B. L. Leventhal, R. C. Craddock, H. S.Koplewicz, M. P. Milham, An open resource for transdiagnostic re-search in pediatric mental health and learning disorders, Scientificdata 4 (2017) 170181.[146] J. F. Cavanagh, A. W. Bismark, M. J. Frank, J. J. Allen, Multipledissociations between comorbid depression and anxiety on reward andpunishment processing: Evidence from computationally informedeeg, Computational Psychiatry 3 (2019) 1–17.[147] X. Jiang, G.-B. Bian, Z. Tian, Removal of artifacts from eeg signals:a review, Sensors 19 (5) (2019) 987.[148] A. K. Jain, J. Mao, K. M. Mohiuddin, Artificial neural networks: Atutorial, Computer 29 (3) (1996) 31–44.[149] A. ollah Ansari, M. Khalili, Diagnosis of major depressive disorderwith neural network models, International Journal of ElectronicsCommunication and Computer Engineering (IJECCE) 5 (5) (2014)1183–1186.[150] M. Bachmann, L. Päeske, K. Kalev, K. Aarma, A. Lehtmets, P. Ööpik,J. Lass, H. Hinrikus, Methods for classifying depression in singlechannel eeg using linear and nonlinear signal analysis, Computermethods and programs in biomedicine 155 (2018) 11–17.[151] A. Craik, Y. He, J. L. Contreras-Vidal, Deep learning for electroen-cephalogram (eeg) classification tasks: a review, Journal of neuralengineering 16 (3) (2019) 031001.[152] X. Li, R. La, Y. Wang, B. Hu, X. Zhang, A deep learning approach formild depression recognition based on functional connectivity usingelectroencephalography, Frontiers in Neuroscience 14 (2020).[153] S. Mitra, S. N. Sarbadhikari, S. K. Pal, An mlp-based model foridentifying qeeg in depression, International journal of bio-medicalcomputing 43 (3) (1996) 179–187.[154] U. R. Acharya, V. K. Sudarshan, H. Adeli, J. Santhosh, J. E. Koh,A. Adeli, Computer-aided diagnosis of depression using eeg signals,European neurology 73 (5-6) (2015) 329–336.[155] D. Shah, G. Y. Wang, M. Doborjeh, Z. Doborjeh, N. Kasabov, Deeplearning of eeg data in the neucube brain-inspired spiking neuralnetwork architecture for a better understanding of depression, in:International Conference on Neural Information Processing, Springer,2019, pp. 195–206.[156] Y. Xiaolong, W. Xiaoyun, J. Xinyi, L. Hongbo, Classification of de-pression with brain network characteristics based on multiphase mapdeep neural network equilibrium compensation, Journal of MedicalImaging and Health Informatics 10 (1) (2020) 134–138.[157] P. B. Mitchell, A. Frankland, D. Hadzi-Pavlovic, G. Roberts, J. Corry,A. Wright, C. K. Loo, M. Breakspear, Comparison of depressiveepisodes in bipolar disorder and in major depressive disorder withinbipolar disorder pedigrees, The British Journal of Psychiatry 199 (4)(2011) 303–309.[158] L. Samalin, F. Bellivier, B. Giordana, L. Yon, V. Milhiet, W. El-Hage, P. Courtet, E. Hacques, N. Bedira, A. Dillenschneider, et al.,PatientsâĂŹ perspectives on residual symptoms in bipolar disorder: afocus group study, The Journal of nervous and mental disease 202 (7)(2014) 550–555.[159] T. J. L. Kristeen Cherney, Effects of bipolar disorder on the body(October 25, 2018).URL [160] M. De Hert, C. U. Correll, J. Bobes, M. Cetkovich-Bakmas, D. Cohen, I. Asai, J. Detraux, S. Gautam, H.-J. Möller, D. M. Ndetei, et al.,Physical illness in patients with severe mental disorders. i. prevalence,impact of medications and disparities in health care, World psychiatry10 (1) (2011) 52.[161] J. Angst, A. Gamma, F. Benazzi, V. Ajdacic, D. Eich, W. Rössler,Diagnostic issues in bipolar disorder, European Neuropsychopharma-cology 13 (2003) 43–50.[162] R. Young, J. Biggs, V. Ziegler, D. Meyer, Young mania rating scale,Handbook of psychiatric measures (2000) 540–542.[163] J. B. Williams, A structured interview guide for the hamilton de-pression rating scale, Archives of general psychiatry 45 (8) (1988)742–747.[164] M. B. First, Structured clinical interview for the dsm (scid), Theencyclopedia of clinical psychology (2014) 1–6.[165] F. Kapczinski, F. Dal-Pizzol, A. L. Teixeira, P. V. Magalhaes,M. Kauer-SantâĂŹAnna, F. Klamt, J. C. F. Moreira, M. A. de Bitten-court Pasquali, G. R. Fries, J. Quevedo, et al., Peripheral biomarkersand illness activity in bipolar disorder, Journal of psychiatric research45 (2) (2011) 156–161.[166] G. Scola, A. C. Andreazza, The role of neurotrophins in bipolardisorder, Progress in Neuro-Psychopharmacology and BiologicalPsychiatry 56 (2015) 122–128.[167] B. S. Fernandes, M. L. Molendijk, C. A. Köhler, J. C. Soares, C. M. G.Leite, R. Machado-Vieira, T. L. Ribeiro, J. C. Silva, P. M. Sales,J. Quevedo, et al., Peripheral brain-derived neurotrophic factor (bdnf)as a biomarker in bipolar disorder: a meta-analysis of 52 studies,BMC medicine 13 (1) (2015) 289.[168] A. C. Andreazza, M. Kauer-Sant’Anna, B. N. Frey, D. J. Bond,F. Kapczinski, L. T. Young, L. N. Yatham, Oxidative stress markersin bipolar disorder: a meta-analysis, Journal of affective disorders111 (2-3) (2008) 135–144.[169] F. Benedetti, V. Aggio, M. L. Pratesi, G. Greco, R. Furlan, Neuroin-flammation in bipolar depression, Frontiers in Psychiatry 11 (2020)71.[170] M. İ. Atagün, Brain oscillations in bipolar disorder and lithium-induced changes, Neuropsychiatric disease and treatment 12 (2016)589.[171] A. Khaleghi, A. Sheikhani, M. R. Mohammadi, A. M. Nasrabadi, S. R.Vand, H. Zarafshan, M. Moeini, Eeg classification of adolescentswith type i and type ii of bipolar disorder, Australasian physical &engineering sciences in medicine 38 (4) (2015) 551–559.[172] S. Z. Metin, T. T. Erguzel, G. Ertan, C. Salcini, B. Kocarslan, M. Cebi,B. Metin, O. Tanridag, N. Tarhan, The use of quantitative eeg for dif-ferentiating frontotemporal dementia from late-onset bipolar disorder,Clinical EEG and Neuroscience 49 (3) (2018) 171–176.[173] M. Zelenina, D. Prata, Machine learning with electroencephalographyfeatures for precise diagnosis of depression subtypes, arXiv preprintarXiv:1908.11217 (2019).[174] M. B. Fonseca, R. S. de Andrades, S. de Lima Bach, C. D. Wiener,J. P. Oses, et al., Bipolar and schizophrenia disorders diagnosis usingartificial neural network, Neuroscience and Medicine 9 (04) (2018)209.[175] T. T. Erguzel, C. Uyulan, B. Unsalver, A. Evrensel, M. Cebi, C. O.Noyan, B. Metin, G. Eryilmaz, G. H. Sayar, N. Tarhan, Entropy: Apromising eeg biomarker dichotomizing subjects with opioid usedisorder and healthy controls, Clinical EEG and Neuroscience (2020)1550059420905724.[176] M. Seeck, L. Koessler, T. Bast, F. Leijten, C. Michel, C. Baumgartner,B. He, S. Beniczky, The standardized eeg electrode array of the ifcn,Clinical Neurophysiology 128 (10) (2017) 2070–2077.[177] J. W. Kam, S. Griffin, A. Shen, S. Patel, H. Hinrichs, H.-J. Heinze,L. Y. Deouell, R. T. Knight, Systematic comparison between a wire-less eeg system with dry electrodes and a wired eeg system with wetelectrodes, NeuroImage 184 (2019) 119–129.[178] C. Wu, I. MacLeod, A. I. Su, Biogps and mygene. info: organizingonline, gene-centric information, Nucleic acids research 41 (D1)(2013) D561–D565.[179] M. L. Phillips, H. A. Swartz, A critical appraisal of neuroimaging
Sana Yasin et al.:
Preprint submitted to Elsevier
Page 28 of 29eural Networks based approaches for Major Depressive Disorder and Bipolar Disorder Diagnosis using EEG signals: A review studies of bipolar disorder: toward a new conceptualization of under-lying neural circuitry and a road map for future research, AmericanJournal of Psychiatry 171 (8) (2014) 829–843.[180] J. Potash, J. Toolan, J. Steele, E. Miller, J. Pearl, P. Zandi, T. Schulze,L. Kassem, S. Simpson, V. Lopez, D. MacKinnon, F. Mcmahon, Thebipolar disorder phenome database: A resource for genetic studies,The American journal of psychiatry 164 (2007) 1229–37. doi:10.1176/appi.ajp.2007.06122045 .[181] F. K. Goodwin, K. R. Jamison, Manic-depressive illness: bipolardisorders and recurrent depression, Vol. 1, Oxford University Press,2007.[182] E. Rejaibi, A. Komaty, F. Meriaudeau, S. Agrebi, A. Othmani,Mfcc-based recurrent neural network for automatic clinical de-pression recognition and assessment from speech, arXiv preprintarXiv:1909.07208 (2019).[183] N. M. N. Leite, E. T. Pereira, E. C. Gurjao, L. R. Veloso, Deepconvolutional autoencoder for eeg noise filtering, in: 2018 IEEE In-ternational Conference on Bioinformatics and Biomedicine (BIBM),IEEE, 2018, pp. 2605–2612.[184] W. Samek, T. Wiegand, K.-R. Müller, Explainable artificial intel-ligence: Understanding, visualizing and interpreting deep learningmodels, arXiv preprint arXiv:1708.08296 (2017).
Sana Yasin et al.: