Machine-Learning-Based Approaches for Post-Traumatic Stress Disorder Diagnosis Using Video and EEG Sensors: A Review

According to the World Health Organization, approximately six in 100 people suffer from post-traumatic stress disorder (PTSD) at some point in their lives. PTSD is a mental disorder with a set of symptoms observed in a person following exposure to a life-threatening event or witness a death. The primary characteristic of PTSD is the persistence over time of certain symptoms: flashbacks, traumatic nightmares, acute stress, and symptoms of depression. In a stressful situation, the whole body goes into tension. This has an energy cost and affects the voice, breathing, daily gestures, and face, which are referred to in this survey article as external symptoms. The diagnosis of PTSD is based on the Diagnostic and Statistical Manual of Mental Disorders (DSM) classification using standard questionnaires where the patient self-reports their condition based on their symptoms. Indeed, these questionnaires have several limitations and give an imprecise diagnosis. Sensors and wearable-based technologies can play a key role in improving the diagnosis, the prognosis, and the assistance of PTSD. Recently, computer-aided diagnosis (CAD) systems have been proposed for PTSD detection based on external symptoms and brain activity using video and electroencephalography (EEG) sensors. To the best of our knowledge, this is the first survey article that gives a literature overview of machine-learning-based approaches for PTSD diagnosis using video and EEG sensors. In addition, a comparison between existing approaches and a discussion about the potential avenues for future PTSD research is provided.

The diagnosis of PTSD is based on the Diagnostic and Statistical Manual of Mental Disorders (also known by the acronym DSM) classification using standard questionnaires where the patient self-reports their condition based on their symptoms.Indeed, these questionnaires have several limitations and give an imprecise diagnosis.
Sensors and wearable based technologies can play a key role in improving the diagnosis, the prognosis and the assistance of PTSD.Recently, Computer-Aided Diagnosis (CAD) systems have been proposed for PTSD detection based on external symptoms and brain activity using video and EEG sensors.To the best of our knowledge, this is the first survey paper that gives a literature overview of machine learning based approaches for PTSD diagnosis using video and EEG sensors.In addition, a comparison between existing approaches and a discussion about the potential avenues for future PTSD research is provided.

I. INTRODUCTION
T HE potential use of smartphones, social media, wearable sensors and technologies, and neuroimaging has revolutionized many mental health practices.As a result of this revolution, a huge amount of data can now be collected as well as analyzed scalable and using sophisticated machine learning techniques, resulting in the integration of clinical decision support systems to be beneficial for a wide variety of conditions [40,54].As an example, Post-Traumatic Stress Disorder (PTSD) is a severe mental illness.This chronic and debilitating mental disorder is triggered by catastrophic life events, such as military combat, sexual assault, and natural disasters.Due to the heterogeneity of PTSD symptoms and the lack of measurements, diagnosing it presents a great challenge for clinicians and researchers.
There is a high prevalence of trauma events around the world, including death threats, sexual violence, and serious injuries.According to a study combining data from 24 countries, 70% of individuals will experience at least one potentially traumatic event during their lifetime (from 28.6% in Bulgaria to 84.6% in Ukraine (pre-war)) [30].

A. Clinical Background of PTSD
Post-traumatic Stress Disorder (PTSD) is a mental health disorder that can happen after experiencing (or witnessing) a traumatic event.The symptoms of PTSD include intense intrusions (flashbacks, nightmares, etc. ), anxiety, negative cognition or mood, hyperarousal, and persistent avoidance of stimuli pertaining to the trauma.[21].In addition to the symptoms of anxiety disorders, PTSD sufferers also experience anhedonia, dysphoria, and dissociations.The Diagnostic and Statistical Manual of Mental Disorders (DSM-V) moved PTSD to the Trauma-and Stressor-Related Disorders category instead of the anxiety disorders category [21].In addition, 68.5% of PTSD patients had comorbid Major Depressive Disorder (MDD), which indicated that PTSD is difficult to diagnose in part because of its similarities to anxiety disorders and depression [14].Females are more likely to suffer from PTSD than males, with PTSD prevalence in the US at 6.8% and 70 percent for females, compared to 30 percent for males [9].Additionally, 9.2% of Canadians have lived too with lifetime PTSD [11].
A number of assessment instruments have been developed and empirically evaluated in the past several decades to measure trauma exposure and PTSD, as well as related syndromes.Numerous measures have been validated and widely adopted internationally, including self-report questionnaires, structured interviews [46], and psychophysiological procedures [22].
In the past, PTSD was diagnosed by medical professionals in a clinical setting, like most other debilitating psychological disorders.In order to diagnose the condition, multiple inperson interviews would be conducted as part of a comprehensive psychological assessment.Occasionally, blood tests and self-reports are an additional requirement for filtering out specific psychological affection to collect information.In primary care settings, patients with post-traumatic stress disorder often go undiagnosed because it is confused with other mental disorders such as depression due to similar symptoms and also because of the patient subjectivity and denial of the disease, this impacts management and the prescription of appropriate treatments.Clinicians are also often biased about all these selfreports that are subject to several limitations and biases such as; (1) introspective ability, which means the subjects may not be able to assess themselves accurately, and (2) the rating scales bias, which means yes or no can be too preventive, numerical rules also can be vague and subject to an individual inclination to give an exciting response to all queries.

B. Investigating Automatic PTSD Assessment
A clinician's observations and self-reported measures are used to diagnose PTSD, such as the PCL-5 (a 20-item selfreport measure designed to assess the 20 DSM-V symptoms of PTSD) [33], Despite the popularity of CAPS-5 (Clinician-Administered PTSD Scale) and IES-R (Impact of Event Scale-Revised), the need for unbiased, objective, and automated methods continues to grow.In the literature, it has been demonstrated that EEG signals are capable of analyzing cerebral activity and detecting and discriminating mental disorders such as MDD and bipolar disorder.Several modalities have been explored recently for the automatic diagnosis of PTSD, as well as many other psychological disorders [63].Several studies in the literature used NLP and machine learning to diagnose clinical depression, a common mental disorder [71,55,50] using sentiment analysis based on text data which was cheap, non-invasive, and easily available online.In addition, machine learning has been used not only to diagnose mental illness [60], but also to predict treatment outcomes and to predict treatment outcome in mental health [42].
Self-reported questionnaires may prove difficult to gather from some individuals, such as infants and young children, so In this survey paper, a systematic and extensive review of clinical and non-clinical methods such as self-questionnaires, video, and EEG, using machine learning-based approaches to detect PTSD is presented to provide researchers with extensive knowledge of this area of research.The different steps of the audio, visual and EEG based single and multi-modal based approaches for PTSD diagnosis and assessment will be investigated in this paper (Fig. 8).
To the best of our knowledge, no comprehensive review on PTSD computer-aided diagnosis systems using selfquestionnaires, audiovisual, and EEG-based approaches has been reported so far.The findings of this review will be very useful and a structured starting point for researchers to study clinical and non-clinical PTSD recognition through machine learning based approaches.

C. Structure of the Paper
The current review is organized in six sections.Section II covers our search strategy and the considered eligibility criteria.The section III deals with state-of-the-art literature on computer-aided diagnosis approaches for PTSD using video data.Section IV gives an overview about the existing machine learning-based approaches for the recognition of PTSD using EEG data.The next Section ?? discusses and highlights the findings of this survey.The last Section V concludes this study and gives suggestions for future work.
The two main parts of this review, presented in Section III and Section IV, outlines the existing approaches for the preprocessing, the feature extraction, the features selection and fusion, the machine learning model selection, training and testing, and the public datasets for PTSD detection using the two modalities, the video and the EEG. Figure 1 illustrates how we divided our survey content into video and EEG sections using the standard ML pipeline.
This survey concludes with recommendations that will help to improve the reliability of the future machine learning-based approaches for PTSD diagnosis and assessment.

II. INCLUSION CRITERIA
Eligibility criteria of this survey includes the suitable depictions of different machine learning techniques for the automatic recognition and assessment of PTSD using video and EEG data.A number of databases were searched for related papers, such as Google Scholar, IEEE Xplore Digital Library, Elsevier, ScienceDirect, Frontiers, and the US National Library of Medicine by using a set of keywords that we define in the following sections.These keywords have been used in combinations of two or more, with either AND or OR operands.
Afterward, we removed duplicates and excluded papers based on their title and abstract (for example, we excluded papers based on MEG and fMRI), and after this screening step, we may take a more detailed look at the paper's content.Figure 2 illustrates the pipeline for both sections.
The articles that cite the ones that we found are checked and those in the scope of this survey are included.This survey is performed by respecting the PRISMA-CI [45], an extension of PRISMA that reports guidelines about systematic reviews.
1) Video and Audio Papers: A number of keywords were used for the video/audio section, such as PTSD diagnosis, machine learning, deep learning, multimodal, validation, models, image, audio, and video processing.There have been several keywords combined and used interchangeably, for example, (1) PTSD diagnosis is the purpose of the paper, (2) machine learning algorithms are used, (3) raw data are audio or video, and (4) the origin of the data should be mentioned.Study selection (8 studies) was conducted between 2014 and 2022.One particularity to note is that all recent studies are based on supervised machine learning algorithms (notably neural networks).The website of the connected paper was particularly useful for searching for studies 1 .The engine takes a paper title as input, then displays all other related papers as a graph, where popular papers (that are frequently cited) are represented by bigger circles (nodes), and more recent papers are represented by darker colors.
2) EEG Papers: The search has been performed using keywords such as machine learning, PTSD, diagnosis, EEG, post-traumatic stress disorder, and deep learning to select papers around 150 studied.We excluded papers that used MEG, and fMRI or that investigated traumatic brain injury, anxiety, and depression and selected related studies, that use machine learning approaches to classify PTSD with HCs as well as with another disorder [74], some studies containing more than one disorder in their datasets [58], that include only EEG and ECG data.

III. PTSD DIAGNOSIS AND ASSESSMENT USING CLINICAL VIDEOS
Computer-based video analysis is known and effective tool for diagnosing physical illnesses like tumors [7] and cruciate ligament injuries [8], as well as mental illnesses such as clinical depression [5,71,55,56,50,57], or also to predict mental disorders like depression relapse [69,68].
In this section, we summarize the different existing approaches for PTSD diagnosis and assessment using clinical videos.Thus, an overview of the existing PTSD video datasets and the data collection procedure are explained, as well as the different steps of the pipeline of the proposed approaches, the used machine learning models and a comparison of the performances of the different approaches is presented to give a summary and an interesting starting point to the researchers in this research field.
A. PTSD Videos Datasets 1) Data Collection Procedure: In the scientific literature of video-based approaches for PTSD diagnosis, different data collection procedures have been proposed.Indeed, databases were composed of either video or audio recordings.Clinical interviews were recorded either directly by humans or by a virtual interviewer controlled by humans in another room.The interviewees present symptoms of PTSD (DIAC-WOZ2 , FEMH 3 ) or several mental illnesses (Aurora Dataset4 ).In addition, certain datasets are based on recordings without a particular clinical protocol (TIMIT 5 , PTSD-in-the-wild 6 ).The PTSD-in-the-Wild dataset is collected mainly from Youtube and aims to recognize PTSD in unconstrained environments, it exhibits "natural" and big variability in acquisition conditions (different lighting, facial expression, resolution, age, pose, gender, race, background, etc.).
2) Reported Datasets: All reported datasets are compiled in table I.In the literature, several video datasets have been considered for the diagnosis and the assessment of PTSD and the study of the mental health of the people suffering from PTSD.In fact, even if the study aims to study PTSD, several used datasets are designated for other mental illnesses (Anxiety, Depression, Stress) or for more general usecases like emotion analysis (TIMIT [4], RAVDESS [38]).
It has been shown no big impact of the chosen dataset on the accuracy and the performances of the proposed model.Different modalities are made available on these datasets: some  contains only audio [37], while others gives video resources.
There are also other data modalities such as text and clinical questionnaires (the PHQ-8) [47]), and biomedical measures [62].Additionally, several datasets can be used in pre-training deep neural networks, especially when the approach is based on transfer learning, such as the study [39] where three datasets (TIMIT, Youtube, and Ohio hospital interviews) are used in the training (pre-training and fine-tuning) the proposed models .

B. Proposed Approaches for the Automatic Assessment of PTSD using Video
Video analysis is one of the most challenging topics in computer vision field as it involves retrieving information, detecting temporal and spatial events, and recognizing and analyzing or segmenting objects and scenes in a video.Medical services can be improved by using video analysis, an emerging discipline.In fact, image-based sensors are more and more widely used for surveillance and monitoring [70].The image and the sequence of images are used to extract relevant information for continuous patient monitoring of wide range of areas such mental health applications [70].
Indeed, video has proven to be effective in diagnosing certain mental or physical disorders [5,71,55,56,50,57].Indeed, Post-Traumatic Stress Disorder (PTSD) is one of the mental disorders that has attracted the attention of AI community and several machine-learning based approaches have been developed for its diagnosis and assessment based on video of the clinical interviews and they follow the steps of pipeline in Fig. 3.
The scientific literature on video analysis is extensive; however, there are several common patterns in these studies.The proposed approaches share the same pipeline and steps: (1) Data pre-processing, (2) Features extraction, (3) Features selection and fusion and (4) Classification and Model evaluation.Fig. 1 summarizes the different stages or steps of the processing pipeline.The machine learning based approaches for PTSD diagnosis and assessment using video follows also this general pipeline and propose the same steps.In the following, we give a detailed description of each step of the general pipeline of the existing machine learning-based approaches for PTSD diagnosis and assessment using video data.
a) Data Pre-processing: Video is an image sequence with an audio signal.Data pre-processing is required as one of the first steps in the machine learning pipeline to improve the quality of the data.In this way, the model can learn better and features can be extracted more accurately.In the scientific literature for PTSD diagnosis from the video, it is also necessary to divide the data to have a training set and a test set.Data pre-processing approaches include: • Audio normalization: A normalized recording has its peak amplitude or peak root mean square value or perceived volume increased or decreased over time to reach a predefined level by adjusting the amplitude or root mean square value.
Since the amplitudes of a long audio sample can vary widely, it is necessary to normalize the signal s and thus the normalized signal as described in [52][66] is defined as : with s the short term energy.In some cases, the normalization is done directly on the numerical values of the feature vectors, for example on the interval of [0,1] [73].• Audio signal windowing: refers to splitting the input signal into temporal segments and a rectangular window, which is a simple truncation function to avoid power leakage, considered essential to use smoothing windows.
In the audio signal processing domain, the process of windowing a signal involves multiplying the time record by a smoothing window of finite length whose amplitude varies smoothly and gradually towards zero at the edges.
Windows often used in speech pre-processing in PTSD computer-aided diagnosis systems are the Blackman window [29] and the Hamming/Hann window [29][66].
• Audio signal restoration : The process involves several operations and depends primarily on the original signal's state.In the scientific literature, however, the following operations have been found to be the most prevalent in diagnosing PTSD through videos:  -Pre-emphasis: Speech preprocessing can remove the lip's spectral contribution effectively by amplifying the input frequency range most sensitive to noise (often using a high pass filter) [66].-Noize removal: The microphone itself produces noise as well as noise from the environment.The Linear Predictive Coding (LPC) filter is an effective way to reduce it in both cases of speech processing.[29] -Silence elimination: When a patient speaks, there are several moments of silence that should be eliminated due to the fact that they do not bring any information.Two approaches are used to eliminate these moments: short-term energy (STE) and zero crossing rate (ZCR).[66] • Video restoration : In most studies, the capture quality and face-framing were controlled to ensure that little preprocessing was required for the visual part.As a result of the pre-processing, saturation and noise are removed [64].• Video time framing : In order to extract features efficiently, it is important to subdivide videos into temporal frames so that the process can be controlled.
In most cases, a frame size of 25ms and an overlap of 10ms are used [32][25] b) Low-level Features Extraction from Face: The face contains several features that are relevant to the diagnosis of PTSD, including: • Action units : Psychologists Paul Ekman and Wallace Friesen developed the facial action coding system (FACS) to describe facial movements.As a result, the FACS has become the primary tool for studying facial expressions, and contractions and relaxations of the face are broken down into action units (AU).The majority of existing video-based approaches for PTSD assessment use the AUs as one of their visual features [35][73] [32][35] [64].
The Computer Expression Recognition Toolbox (CERT) [17] is a global framework for automatic AUs extraction in real-time [17].In this approach, face detection is accomplished by extending the Viola-Jones method, facial feature detection by using detectors and linear regression, feature extraction by using a Fast Fourier transform, and Action Unit Recognition is accomplished by using a linear support vector machine (SVM) for each action unit.A similar approach  [64][73] [35] can be achieved by the openface python package [36].This way the proceeding may seem outdated, nevertheless it is still very competitive especially for PTSD diagnosis even compared to the most recent deep learning approaches.This action unit feature is thus used to determine the temporal variation of facial expressions, either through velocity (first-order derivative) or acceleration (secondorder derivative) [32] or by the Motion History Histogram (MHH) approach [35].
• Geometrical features : The geometrical features are based on particular points (Fig. 5b) of the face extracted using Facial Landmark Detection, from these points, distances and areas are deduced, these provide information both on the global topology of the face and on isolated and precise micro-movements.Distances are normalized by dividing the values of the face's total width and mean surface over the whole interview.
c) High-level Features Extraction from Face: PTSD causes fluctuations in mood, arousal, and alertness, which makes the face of person with PTSD different from the faces of healthy people.Due to their high performance, deep neural networks are increasingly used to distinguish patients with mental disorders, and it is the most common method in recent studies to extract video features [73] [32][35][64].Thus, highlevel PTSD features or deep encoding are learned from the facial images via deep learning as shown in Fig. 8.
Two types of visual features have been investigated for PTSD classification: the facial and the movement features [73].Deep facial features have been implemented in the Openface python package [36], which is based on triplet Loss architecture 6. Openface is a framework that implements modern facial behavior analysis algorithms including: facial Action Units (AUs) recognition and landmark detection, head pose and eye gaze tracking.
The extracted AUs were used to compute a facial expressivity score for seven discrete emotions, a peak of expressivity, facial expressivity index and expressivity peak count for PTSD classification [73].While to analyze the movement, head movement, attentiveness and pupil dilation rate are extracted from raw images in the video using OpenFace package [73].
These low-level facial and movement features and others voice and speech features are fed to a two layers deep neural networks to extract high-level features for PTSD and MDD classification in [73].These features are not the most effective in PTSD classification but they have good ranking according to their predictive importance [73].
Another recent approach for PTSD classification based on high-level deep features learned using an hybrid CNN-RNN architecture is proposed in [72] as visual baseline model for PTSD classification and evaluated on the PTSD-in-the-wild dataset.ResNet50v2 originally trained on the ImageNet dataset is used for extracting meaningful features that are fed to a sequence model consisting of recurrent layers of LSTM.The proposed approach has shown high performances in binary classification of PTSD from faces images in the wild.
Several deep learning based approaches have been proposed in the literature for depression recognition using video, contrary to PTSD recognition which is a more recent research topic for the computer vision community.In fact, approximately half of people with post-traumatic stress disorder (PTSD) also suffer from Major Depressive Disorder (MDD) [65].Thus, all existing deep visual learning based approaches for depression recognition [5,71,55,56,50,57]   the anchor is compared to a matching input (positive) and a non-matching input (negative), the purpose is to reduce the distance/deviation between the anchor and the positive and increase the same between the anchor and the negative.
where T i are the extracted F 0 (the fundamental frequency) period lengths and N is the number of extracted F 0 periods.
While Shimmer feature is computed as: where A i are the extracted peak-to-peak amplitude data and N is the number of extracted fundamental frequency periods.
• I-Vector: I-Vector is the state-of-the-art approach for speaker verification task [15], an i-vector is a lowdimensional fixed size representation of an utterance that is very efficient in capturing spectral variability cues, let's define: where m is the mean super-vector obtained by concatenating mean parameters in a Gaussian Mixture Model (GMM) trained using data from different classes.M is the mean centered super-vector of the speech utterance derived utilizing the 0th and 1st order Baum-Welch statistics.Similarly, T is a matrix that defines a total variance space representing the speaker and channel variability's trained using Expectation-Maximization (EM) and v is a low dimensional factor that observes a standard normal distribution N(0; I), referred to as the i-vector.This feature was used for PTSD diagnosis and has outperformed all other audio features [28] [32] Another existing approach fed low-level features to deep neural networks to extract high-level features [67].Indeed, it aims to diagnose PTSD from an emotion recognition viewpoint, thus, the Mel-spectrum which is a low-level feature, is used as input to a CNN-LSTM (CNN: convolutional neural network, LSTM: Long short-term memory model), the CNN extracts the effective high features and catches the spatial structures, which contains several convolution and pooling layers as well as LSTM finds the interdependence within the learned features and finally a single long continuous linear vector is obtained.
Similarly, another variant of this approach present in [39] is based on deep belief network (DBN); Three categories of lowlevel speech features are used to pre-train multiple restricted Boltzmann machine (RBMs) layer by layer, and then the network is fine-tuned using class labels.The transfer-learning strategy has been implemented to overcome PTSD patients data scarcity.
VGGish-based deep neural network has been used recently in [72] as baseline model for binary audio classification and evaluated on the PTSD-in-the-wild dataset.The VGGish-based approach achieved good performances in the audio PTSD classification task using two strategies (Training/testing and 3-fold cross-validation).
f) Features Fusion: The features extracted from different modalities (audio and video) need to be merged to detect further correlations between them.The fact that multiple complementary features can be combined makes it possible to create a more powerful feature, even if the study is done on one modality [16].The multi modalities fusion strategies used in the scientific literature concerning the diagnosis of PTSD from video (sound and images) can be subdivided into three major categories: • Early fusion: Based on a feature-level, this fusion is a simplest and the most widely used technique for feature fusion that consists on concatenating the different features vectors into one high-dimensional single vector earlier [25] [39] [32] [28] [73].Otherwise, this technique is often associated with a dimensionality reduction technique like PCA (principal component analysis) [25] [28].• Late fusion: In a late fusion, models are processed independently with their own inputs, then the outputs are combined using arithmetic operations, most commonly averages [66] weights, or logical operators AND.It is often used with shallow learning based approaches decision tree based models [32][66] Or XGBoost algorithm [66], hence this fusion strategy also called decision-level fusion.
• Model-level fusion: In this family of techniques, the fusion is carried out itself and often used in neural network models, which perform the fusion between different modalities that is done by additional layers.Thus, in [67], the authors used CNNs to captures the dependencies between features and represent them for the LSTM.Similarly, graphical models used as well in this approach by adding a simple predictor such as a simple multivariate regression model [35].Finally, the fusion can be done by a DNN (Deep Neural Network) such in [73].Finally, there is another approach in the video-based PTSD CAD literature, which consists of the subspace learning methods that using infer latent spaces with multi-view data at training phase.[18] [12], In general, this technique does not imply the fusion of features, but rather it shows the fundamental relationship between the different modalities, which has the advantage of using high-cost features only in training phases [28].
g) Classification Step: Various classification methods and algorithms have been used as a last step in the pipeline of CAD systems of PTSD from video like deep neural networks [35][73] or boost algorithms [66] to more traditional classifiers like SVM [39][28] as well as Gaussian Naive Bayes [28].
Two classifiers have been investigated in the classification of PTSD features extracted from video using artificial neural networks: the first uses a traditional classifier and the second uses a fully connected layer within the deep neural network.In this study [39], both solutions have been studied and finally it has been demonstrated that SVM is better performing in this application of PTSD diagnosis.Despite the results, this study is not systematically best, indeed the papers which have the best precision in PTSD detection [67] [73] have favored the second approach by insisting on the fact that in a multi-modal study, the classification by the neural networks allows to detect unique probabilistic information from different modalities, which show better discriminatory accuracy for the PTSD diagnosis classification, the study [73] implements this approach by using a sigmoid binary classification with an Adam-optimizer and utilized dense layers for global features extraction [67].
h) Evaluation Metrics: This section illustrated the evaluations metrics used to evaluate the classification performance of the proposed video-based approaches for PTSD diagnosis: Note that some studies have been used different classifiers and different set of features for comparison purposes in that case the model and the features set with the best performance will be selected and mentioned in our comparative table.
To conclude, the overall results of the studies (see table II) are very promising.The recent proposed video-based approaches for the diagnosis and the assessment of PTSD using the deep learning techniques outperformed the state-ofthe-art approaches thanks to its capacity to learn very discriminant deep PTSD features.Hence, deep neural networks have demonstrated another time great potential in recognizing the mental state based on audiovisual data.Indeed, one could imagine a telemedicine service that uses digital biomarkers for PTSD diagnosis.
Therefore, there are still some limitations, the most common one is the size of the dataset.Large datasets are required to train the deep neural networks which can pose a problem in the medical field because of the difficulty of data acquisition [67][39] [35][73].Also, the cost of data collection is quite high, one of the solutions is the the subspace learning method presented in [28].Several approaches have been proposed in the literature for data augmentation and to deal with data scarcity and the high cost of data collection.We cite among these approaches: (1) The Generative Adversarial Network (GAN), a family of deep neural networks that learns to generate new data with the same distribution as the training set, (2) transfer learning that uses pre-trained deep neural network as a starting point for a new task and then leverages the knowledge gained from a previous task to improve generalization capacity about another new task and (3) simulation and transformation functions that apply some functions to modify the original signal or data and then generate new data.In addition, video data even for healthcare applications are universally considered sensitive and confidential.Patient records are confidential and doctors and medical practitioners do not have the right to share them or to use them without patient's permission.Health data, specially video health data, is very sensitive, as it comes within a person's intimate sphere and information.For that, it must be stored safely and out of reach of authorised users.Machinelearning based approaches need data but in the opposite side it must guarantee the privacy and the confidentiality of the data especially in mental health applications.We expect that this point will attract the attention of AI community in future and solutions will be proposed.Fig. 9: International 10-20 system for electrode placement [44] Also, all the studies have concluded that the audio modality (voice and speech) is more informative than the frame images modality, however the use of multi-modalities is more effective than using a single modality in all proposed approaches except for [66] where their approach succeed to achieve a very high diagnostic accuracy using only one modality (the audio).Finally the studies that are based only on features concatenation as the multi-modal fusion strategy, which ignore the potential dependencies between the modalities (especially the non-linear dependencies).

SIGNALS
Electroencephalography (EEG) are signals which are obtained through electrodes to measure voltage fluctuations across the scalp.The placement of the electrodes follow the international and widely used 10-20 system (figure 9).
Electroencephalogram is a non-invasive, effective and powerful tool for recording the electrical activity of the brain and for the diagnosis of various mental disorders [63].EEG signals are easy to record comparing to other methods such as fMRI and PET.EEG has a high temporal resolution typically up to 1000 Hz but it also suffers from a very low spatial resolution due to the diffusion of electrical signals passing through the scalp.The 10-20 system contains standardized 75 electrode placements on the scalp 9. Placements used for recording EEG signals differ in the discussed studies as some include only a subset of these electrodes.Some experiments used 19 electrodes [58] while others used 64 [48] for higher spatial resolution and only analyzed the 10 most commonly used channels such as F3, F4, C3, C4, T3, T4, P3, P4, O1 and O2.

A. PTSD EEG Datasets
In this study, we noted 8 EEG papers that are distinguishable from each other and discussed three major experiment categories: • Resting state studies.These studies typically take between 2-5 minutes of sitting in a dark silent room, eyes closed and record the EEG signals .[58,47,61,75,76].• Sleep studies.These studies record the EEG signals of subjects for a whole night of sleep (8 hours of sleep).
[48] • ERP studies.Event Related Potential (ERP) studies involve exposing the subjects to auditory and visual stimulus and then analyzing the effect of these stimuli using EEG signal.[43,74].
2) Reported Datasets: Table III describes the reported datasets containing EEG recording for PTSD diagnosis.The table lists the papers that cited the dataset, as well as the year it was collected.If the collection date is not mentioned in the paper, the release year is used instead.The table also includes the number of samples and PTSD samples in parentheses, channels and sampling rate for the EEG signal, type of the experiment, classes of subjects included in the study and data availability as well.
The reported datasets included 78 to 945 individuals belonging to both classes (or more [58]).Some of the most common exclusion criteria are: individuals with substance or alcohol abuse in recent months, severe depression, mood/anxiety disorder, brain injury, and medication that affect the brain's normal functioning.Furthermore, some studies excluded males or females so that the results are not affected by genders.While, sleep studies excluded individuals with sleep disorders [48].
3) Automatic Assessment of PTSD using EEG Signals: In order to preprocess and extract features from EEG data, several techniques have been employed due to the fact that EEG signals are extremely noisy and have low spatial resolution inherently.Analyzing EEG signals using machine learning follows the standard machine learning pipeline, as shown in Figure 1.These steps are important for the model to fit the data and produce accurate predictions.
The different steps of the analysis of EEG signals for PTSD diagnosis are further explained with detailed in the following subsections.
a) Pre-processing: Wet and dry electrodes of the EEG sensors are highly susceptible to various forms and sources of noise.In fact, dry electrodes are more susceptible to noise than wet electrodes due to their weak contact leading to high impedance [51].Among the numerous sources of noise, motion artifacts, thermal artifacts, flicker, line noise, and the half-cell effect at the skin-to-electrode interface are the most frequent ones reported in the literature [51].
Other sources of muscle, motion artifacts and other nonbrain physiological activities are often observed in the EEG signals such as eye movement, heart rhythms, and muscle activity [41].Therefore, EEG signals denoising is necessary   for better analysis and to improve the performance.Artifacts that disturb the brain signal can be physiological: blinking (figure 10b), eye-movement (figure 10a), electromyography artifacts due to muscle activities (figure 10c), heartbeats, head movement as well as non-physiological: power-line interference (50Hz or 60Hz), loose wiring, poor electrode connection or placement.The pre-processing of EEG signals is different from one study to another (important to note that the order of the pipeline's steps differ).In the following, we detail the important pre-processing steps proposed in the literature of EEG signals for PTSD diagnosis: • Re-referencing.Electroencephalogram signals are voltage fluctuations produced by electrodes across the scalp, and the recorded voltage is the potential difference between an electrode and a reference electrode.The selection of the reference electrode can be complicated since it will set the voltage on that electrode to zero, then it eliminate all the information that will come from that electrode and its associated brain region.For example some studies [75, 43] used M1, and M2 Mastoids as reference which are located near the ears.In EEG preprocessing, re-referencing is a procedure of combining the different electrode values to produce an artificial reference.In the 2020, some papers [47] used the common average reference.This method uses the average of all the electrodes as the zero point.This technique helps to overcome the problems of information loss as no one specific electrode is reduced to zero.• Filtering.The IIR Butter-worth bandpass filter is used to remove high frequencies in PTSD EEG signal analysis [47]).The definition of the high cutoff frequency differs as it ranges from 50Hz to 100Hz.The low cutoff frequency ranges from 0.  signals and cleaning the data for our machine-learning models requires removing gross artifacts.This step is usually performed by a well-trained expert who inspects the signal and marks the beginning and end of each artifact.[61] • Processing smaller artifacts.Smaller artifacts such as eye movement or blinks are removed or minimized.These artifacts are identified either visually by comparing peakto-peak amplitudes to a voltage threshold, or by comparing signal values/power to a moving average.For example 5s epochs are rejected when the EEG signal power in 26 Hz-50 Hz exceeded 4 times the moving average of 3 minutes in the 2020 paper [48].Smaller artifacts are more common and removing them will reduce the size of our data significantly.For this reason, some studies used mathematical approaches to minimize the influence of these artifacts without deleting the segments that contain them.In [47], it was proposed to use the established mathematical procedure implemented in CURRY 7 [1] to reduce ocular related artifacts.
• Temporal segmentation extraction: divide the signals into 2s to 5s segments [61,47,48].Some resting state studies rejected epochs where the participant was thought to be drowsy or fell asleep.They performed this by removing epochs where theta and alpha ratio exceeded [47].
b) Feature Extraction: Several techniques have been used for feature extraction to achieve better results with the machine learning models after pre-processing the EEG signals.As a result, some techniques in PTSD diagnosis using machine learning literature have a heavy mathematical background and extensive prior research to support their value.As a result, we have summarized some of the most commonly used techniques in this section.
Spectral analysis: The EEG signal is composed of rhythmic activities in a wide spectrum of frequencies, which can be converted into a frequency domain using spectral analysis.This can reveal how the EEG power is distributed over frequency [44].In mathematical analysis The Fourier transform (FT) has been developed to perform frequency decomposition of a time function.Fourier transform deals with continuous functions.Moreover, the discrete Fourier transform (DFT) is the FT version for discrete functions, which defined in the following equation: where k=0,1,...,N-1.For computational efficiency the Fast Fourier Transform (FFT) algorithm is most commonly used for spectral analysis instead of the Discrete Fourier Transform.Using the FT methods produces the Power Spectrum Density (PSD) which describes the contribution of each frequency to the time domain signal.Typically, PSD is analyzed in the well-known frequency bands such as delta [1-4 Hz], theta [4][5][6][7][8], alpha [8][9][10][11][12], beta [12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27][28][29][30], and gamma .In a recent study realized for the identification of veterans with PTSD based on EEG features collected during their sleep ( [48]), the mean and coefficient of variation of the log EEG power spectrum are used for each of its 12 frequency bands in each epoch as features fed to a predictive model.
Source localization: Source localization is the process of calculating the location and magnitude of the dipoles responsible for the electromagnetic activity inside the brain from the potential recorded through the electrodes.This process involves two complementary sub problems such as coined forward and inverse problems.The forward problem is the determination of the scalp potentials given the current sources' configuration in the brain and inverse problem involves estimating the location of sources from the recorded EEG signals [24].All source localization studies reported in this survey use the distributed source model to represent the sources as a large number of dipoles distributed evenly in the whole brain volume.
Another important parameter is the head model, which determines how the sources affect the electrode recordings.It defines in fact the geometrical shape, permeability and conductivity of the subjects head.For simple head models the forward problem can be easily calculated (even analytically) and for more complex and accurate models it can always be numerically calculated [44].By comparison, The Inverse problem is much more complex and challenging called illposed problem which has multiple valid solutions 11.
Fig. 11: The forward problem indicated on top takes the source activities as input and constructs the scalp potential map.The inverse problem indicated by the bottom arrow, which takes the EEG recordings and estimates the location of the source signals [44] .
Multiple source configurations can result in the same scalp potentials, which is a core problem of source localization.To overcome this ill-posed problem several methods have been developed to apply certain constraints to obtain a unique solution [59].In PTSD diagnosis using machine learning literature only two methods have been used weighted minimum norm (WMN) [47,61,75] and standard Low resolution electromagnetic tomography (sLORETA) [43] according to the best of our knowledge.The three WMN papers used openMEEG [13] software to construct the head model.In addition they used Brainstorm toolbox [19] software for WMN estimation of brain source activities.
Functional Connectivity (FC): Efficient communication between function-specific brain regions is essential for healthy cognitive functioning [20].To quantify the effects of connectivity for EEG signals, researchers have developed methods to measure the statistical dependencies between spatially remote regions of the brain and several techniques have been used to measure functional connectivity: • Coherence based measures.In frequency domain, coherence (magnitude squared coherence) assesses the linear relationship between two signals [58].
Where S XX is the individual spectral density of the signal X(t).S XY is the cross spectral power density Coh XY is the coherency index between X(t) and Y (t) which takes values from 0-1. larger coherence indicates higher statistical dependence between the two signals [44] • Phase synchronization based measures.In this technique, the phase difference between two signals is calculated by combining Fourier and Hilbert transforms.After extracting the phase difference, different indices are used, such as the PLV (Phase Locking Value): X H (t) is the Hilbert transform of band pass filtered signal X(t).P.V. is the Cauchy principal value the analytic signal X an is then obtained the amplitude A x (t) and the instantaneous phase ϕ x (t) of the signals X(t) are obtained ∆ϕ rel is the relative phase difference, ranges from 0 to 2π.Finally we can calculate the PLV index between the signal X(t) and Y (t) e j∆ϕ rel (tn) (5g) where N is the number of points.E{.} is the expectancy operator.Equation 5h is included to familiarize the reader with the expectancy operator as it is used extensively in the literature.PLV has two main advantages over coherence: it does not require stationary signals and it quantifies the phase relationships clearly.[6] The PLV range is [0, 1]. 1 indicates that the signals have constant phase difference across time and 0 indicates that the phase difference is uniformly distributed.[23] Another index present in PTSD diagnosis using machine learning literature [48] is the weighted Phase lag index (WPLI): 6b) In order to compare PLV and PLI methods, it is necessary to explain the problem of the common source, which consists of two sub-problems.The first is the active reference electrodes.Research has shown that the choice of reference electrodes (Mastoid, Cz electrode, common average reference) influences the extracted features such as ERP latency and amplitude, spectral powers, and FC measures [44].The second subproblem is the propagation of electromagnetic fields in the brain termed volume conduction, whereas the PLV is not robust against common sources [23].This is why researchers have proposed the PLI as less sensitive to these problems [10].A further improvement to the PLI is weighted PLI (WPLI) which eliminates the effects of relative phase discontinuity on PLI [23].
Furthermore, fully connected network can be applied to source localization features since it is also a time series and can be applied to sensor-level features as well.[61].
Spatial complex brain network analysis: The brain interconnections can be viewed as a complex network on different scales (neurons, regions, electrodes) through the well established graph theory foundations and the relatively modern complex network theory .We can analyze the brain network structurally and functionally by presenting the brain as a graph to extract new features that would improve the model's accuracy significantly.Studies using EEG signals usually use the electrodes as nodes and phase synchronisation measures such as PLV are commonly used to quantify the relation between every two nodes and to obtain a correlation matrix [47,61].Hence, the threshold-based correlation matrix obtained a connectivity matrix, which is similar to an adjacency matrix which represents the network graph.Now, the brain network graph is constructed and represented as an adjacency matrix.Many indices can be used as features for global-level (the whole network) and nodal-level (for every node): • clustering coefficient: indicates how node is wellconnected to neighborhood where cc i the clustering coefficient, l i show the number of edges between vertex i and his neighborhood, k i which is the number of vertices in i.
• average path length: The path length between two nodes is the number of edges that constitute the shortest path between two vertices.• efficiency: In complex network theory the efficiency measures how well the network exchanges information between nodes.Micro-states analysis: Dietrich Lehman et al. [2] discovered that few topographical maps of the electric potential over all the EEG electrodes can take a discrete number of states that remain stable for a short period of time (80 ms to 120 ms).Theses configurations are then termed micro-states.However, researchers found that 4-8 micro-states would explain 80% of the EEG data, whereas, recent studies use clustering algorithms such as k-means to find the best micro-states.This unsupervised learning techniques help to avoid bias and to increase accuracy.[44,76] c) Feature Selection: We use various techniques to transform our data into high-level features after feature extraction.The problem now is that our data may have a lot of dimensions.For example, nodal network measures have three measures x 19 channels x 15 frequency bands, which equals 855 features.In order to avoid the curse of dimensionality, feature selection plays an important role in reducing the number of features.In addition to reducing training time and improving model quality, this method helps avoid overfitting.These are some of the methods that were used to select features (we cited the reference accordingly at the end of each methods): 1) Fisher score: The quality of a feature is determined by the value assigned to instances that belong to the same class and the value assigned to instances that belong to different classes [26], With this intuition, the score for the i-th feature S i will be calculated as such: [47, 61] 2) Concordance Correlation Coefficient (ccc): In order to assess the reproducibility of data, the correlation coefficient has been developed in order to measure the agreement between two variables [3].[48].The ccc (ρ c ) is defined as such: where µ x and µ y are the means and σ 2 x and σ 2 y are the variances for the variables x and y respectively.ρ is the correlation coefficient between the two variables.[47,61] 3) Sequential Backward Selection (SBS): The Sequential Backward Selection method is a wrapper feature selection method.It estimates the feature subset based on the classifier's results.Following the removal of features from the set, SBS sequentially reaches or increases a criterion by removing features sequentially [75].4) Non-zero regression coefficients: Features that have non-zero regression coefficient for more than six or seven folds in K-fold cross validation in a logistic classification model [58].5) Relief-based algorithm: Relief and its multi-class extension Relief select features to separate instances from different classes [26].
where l is the number of sampled instances  (Fisher geodesic minimum distance to the mean) .There is no doubt that SVM is the most popular model used in PTSD diagnosis based on EEG and machine learning literature, and it provided the best results for 5 out of 8 studies.However, we cannot conclude that SVM will always be the best model.In the 2021 paper [58], the best classification results have been obtained using Random Forests model as well.In order to confirm which is the best model, we have to consider a lot of parameters like data size, feature dimensions, and data values.In 2022, authors • Specificity: Is the ratio of correctly identified negatives over all the negatives • ROC AUC: The receiver operating characteristic(ROC) curve plots the sensitivity versus 1 -specificity and the AUC is the area under the curve of the ROC f) Conclusions from the Experimental Results: Table IV Table .IV presents a comparative analysis of the existing studies and the different techniques applied for features extraction, features selection, deployed model, the used dataset and the performance of each proposed study.In [58], the best AUC score of 0.97 is reached with an accuracy for PTSD and HCs classification of acc = 0.92 which is the highest accuracy compared to the other reported studies.This study was performed in a dataset that contains seven classes of disorders, as shown in table III.With a large dataset of patients with multiple disorders, the model was trained to find distinct PTSD biomarkers and to classify PTSD patients with comorbidities.
The second best PTSD classification performance are represented in [75] that used PTSD and HCs in its dataset and achieved the best classification results with delta frequency band ([1-4 Hz]).Furthermore, many studies such as [47,75,61] have found the source-level features to improve the quality of the model and utilized the source localization techniques to overcome the low spatial resolution of EEG signals [47].
Additionally, the ERP study [43] confirmed reduced P300 amplitudes and delayed latency at Cz, Pz and T8.Moreover, it highlighted the importance of the cingulate gyrus source-level features as it helped discriminate PTSD patients from HCs and MDD patients.The only sleep study included reached an AUC of 0.83 and showed that C4-C3 synchronization in the alpha band was larger for PTSD patients.Deep learning-based approaches are more more sophisticated and used in EEG related applications [40,54].However, there are still lack of approaches based on deep learning models for EEG analysis for PTSD recognition as mentioned in survey [53] despite the fact that deep learning can improve the ERP classification results up to 30% as shown in [53].
Similarly, Recurrent Neural Networks (RNNs) or Long Short-Term Memory (LSTM) or Transformers have not yet been implemented for PTSD diagnosis based on EEG signals.
It is anticipated that further future advanced deep learningbased studies will improve the current results, and perhaps then CAD will be widely available and used, thanks to the enormous amounts of EEG data that are produced every day, the academics' increasing interest in CAD, and the breakthroughs in deep learning and signal processing.
With the recent progress, flexible and cost-effective EEG can be practically used in ubiquitous bio-potential monitoring.In addition, high-quality signals and comfortable user experience in bio-potential recording may ensure the deployment of EEG wearable sensors in mental healthcare applications [34].
In fact, EEG data is more hard to obtain than video data.The EEG data scarcity could be a limitation to the development of further advanced machine learning-based approaches for PTSD diagnosis and assessment.However, EEG data is less sensitive than video data as patient's identity can be hidden more easily.We strongly believe that brain activity measures can improve not only PTSD, but other mental disorders diagnosis with a more precise, more automatic, more quantitative and less biased diagnosis.It can gives a complementary information to the DSM-5 questionnaires.It can also gives a screen capture about the brain activity and its evolution in time and help the doctor to better follow-up patients and to better prescribe medication No multi-modal approach that fuses EEG and video data has been yet proposed for PTSD diagnosis and assessment.As discussed in previous sections and as seen in the literature, multi-modal approaches have better performances than single modality-based approaches.We are interested in our future works to propose a deep learning-based approach for EEG and video learning of PTSD patterns recognition and assessment.We hope also that this survey paper will give innovative ideas to AI researchers interested on developing smart systems for mental health in general and for PTSD more precisely.

V. CONCLUSION AND FUTURE WORK
Post-Traumatic Stress Disorder is common and disabling mental disorder that can affect the person's life.Computer vision and artificial intelligence community starts recently proposing smart systems for the diagnosis, the assessment, the assistance and the follow up of patients with PTSD.In this survey paper, we gave an overview about video and EEG sensors-based approaches for PTSD detection and assessment using machine learning.To the best of our knowledge, this is the first survey paper that investigate this research topic and it will be of great help to researchers that will start developing artificial intelligence based approaches for PTSD or in general for psychiatry and mental disorders study based on external symptoms, brain activity, textual information from video and EEG signals.It is a recent research field with great potential to help clinicians in their diagnosis and to better assist PTSD patients.Despite the limitations of the proposed approaches, we expect that in future works, several technical, ethical and innovative solutions will be proposed.The use of video and EEG sensors is new revolution in telemedecine that can bring new interesting information in addition to the DSM-5 questionnaires, improve the medical diagnosis with automatic quantitative and less biased measures, to ensure the follow-up of patients at home and between the medical visits at hospitals and to alert the doctor and/or the patient in case of deterioration of the patient's mental health condition.

Fig. 1 :
Fig. 1: The figure shows a typical machine learning pipeline, which includes pre-processing, feature extraction, feature fusion and selection, and model training.

Fig. 2 :
Fig. 2: A pipeline for searching electronic databases and filtering papers.

Fig. 3 :
Fig.3: Pipeline of machine learning-based approaches for Automatic PTSD detection using video of clinical interviews.This figures shows a veteran suffering from PTSD after being exposed to war trauma.The clinical interview is performed by a virtual interviewer instead of the clinician.The audio and images frames of the veteran are processed by a machine learning-based approach to detect PTSD.The different possible steps of the machine learning based approach are described in Section.III.

Fig. 4 :
Fig. 4: Hamming windowing multiplies the first and second signals as inputs to generate the third signal.
(a) A list of the most commonly used AUs for diagnosing PTSD.(b) Facial landmarks used for PTSD diagnosis.

Fig. 5 :
Fig. 5: The main low-level features extracted from facial images for PTSD diagnosis and assessment.(a) the most used AUs for PTSD diagnosis and assessment, (b) the used Facial landmarks for PTSD diagnosis and assessment.

Fig. 6 :
Fig. 6: Triplet Loss architecture implemented in Openface.theanchor is compared to a matching input (positive) and a non-matching input (negative), the purpose is to reduce the distance/deviation between the anchor and the positive and increase the same between the anchor and the negative.
ence are correlated with a high PTSD symptoms severity.In addition, lexical diversity, pauses between words and the sentiments emanating from speech content are considered as unique patterns and indicators of PTSD.Vocal features are the most used features and concerns all studies [25] [39] [32] [28] [73][67][66][35], they are very accurate in characterizing the vocal sound.The most used vocal features are:

Fig. 8 :
Fig. 8: Audio, visual and EEG based single and multi-modal based approaches for PTSD diagnosis and assessment.Deep encoding is learned from each modality separately for PTSD detection and assessment.The learned high-level features are fused in multi-modal approaches for PTSD detection and assessment.

1 )
Data Collection Procedure: The experiment's design and execution must avoid noise that can disturb EEG signals, such as power-line interference, bad electrode placement, and poor amplifier quality.While recording EEG signals, electrodes are placed on the scalp of the subject.

Fig. 12 :
Fig. 12: Illustration of the process of extracting the brain network graph.(A) raw EEG signals.(B) correlation matrix generated using functional relation quantifiers (such as phase synchronization methods).(C) connectivity matrix established after threshold the correlation matrix.(D) Brain network graph[49] [75] tested two shallow machinelearning classifiers, Random Forest and AdaBoost, and also three deep learning classifiers based on convolutional neural network (CNN), ConvNet, EEGNet and a thirteen layers-based deep CNN.They have reported only the best performance of the SVM classifier which outperforms the other shallow and deep neural networks in EEG signals classification for PTSD recognition.Deep learning based approaches do not perform well in this study because of the small size of the used EEG dataset.e) Classification metrics.:There are no regression results in the 8 examined papers, meaning that only the classification of PTSD and non-PTSD is reported.The most common metrics used to evaluate the classification are: • Accuracy: the number of correctly classified samples over the whole set Accuracy = TP + TN TP + TN + FP + FN (10) where TP: True positive, TN: True negative, FP: False positive and FN: False negative • Sensitivity: Also known as recall.Defined as the ratio of correctly identified positive classes over all positive classes Sensitivity = TP TP + FN

TABLE I :
An overview of the video datasets reported for PTSD research.
must be accurate to discriminate between faces of healthy and stressorrelated disorder patients such as PTSD or MDD.Developing visual features to discriminate between anxiety and stress disorders is another more and very challenging research topic that is not yet explored in the literature.Two main sets of voice features have been investigated in the literature for PTSD diagnosis and assessment, speech content and vocal features.In fact, the speech content is the most informative features set by far as we can see in the comparative SHAPE d) Low-level Features Extraction from Voice: (Shapely Additive explanations) features ranking in [73], Indeed, for examples a low speech rate and narrative incoher-

•
Classification accuracy (ACC) is a metric that indicates the number of correct predictions divided by the total number of predictions.It is defined as: • AUC (Area under the ROC Curve) represents the measure of the classifier's separability.Hence, AUC is the area under ROC curve, this latter plots the true positive rate (sensitivity) against the false positive rate (FPR) at various threshold settings where sensitivity and FPR are defined as:

TABLE II :
Table comparing studies that used audio and video to diagnose PTSD (*): The tests were conducted on the training set.

TABLE III :
Comparative table of the reported datasets.(*): schizophrenia, mood disorders , obsessive-compulsive disorder, addictive disorders ,anxiety disorders, trauma and stress-related disorders and HCs.(**): major depressive disorder(MDD) 1 Hz to 1 Hz.Very high frequencies are removed because of very little relevance in the PTSD diagnosis.It has been demonstrated [75] that low frequencies can be enhanced the classification results significantly (from 0.7 to 0.86 accuracy).
• Removing gross artifacts.A gross artifact occurs when the electrodes are not properly connected or when the head moves.EEG signals can be visually examined in order to see the irregularities.Pre-processing the EEG [47]48]enotes values on the i-th feature of the nearest instances to x k with the same class.H k denotes values on the i-th feature of the nearest instances to x k with a different class.d(.) used as a distance operator.[76]d)ModelSelectionandTraining:Thealgorithm and the architecture of the model vary from one paper to another.In fact, in machine learning, model selection relies heavily on trial and error.In-depth knowledge of the model internals is helpful in making educated guesses, but generally, the best way to choose a model is to train several models and then choose the one that performs the best based on predefined metrics.Various models were used to diagnose PTSD using machine learning and EEG signals: SVM [61, 58, 75, 43, 76, 74], Random forests[58, 75], logistic regression[58,48], Riemann classifier FgMDM[47]

TABLE IV :
Discussion and Conclusion: A growing interest in computer-aided diagnosis (CAD) of PTSD is evident as seen Comparative table of the reported studies.(Acc): accuracy.(*): method 4 in the feature selection section.(**): accuracy of classifying MDD and PTSD patients in table IV.In 2022, three papers were published and multiple features, models, and datasets were used.It has been demonstrated and discussed in the selected studied papers in this survey that large datasets, multiple classes (not limited to PTSD, and HCs), source localization of features, and low-frequency bands can improve the model's accuracy and generalizations. g)