1-s2.0-S2772442523000527-main

Healthcare Analytics 3 (2023) 100185
Contents lists available at ScienceDirect
Healthcare Analytics
journal homepage: www.elsevier.com/locate/health
A hybrid mental health prediction model using Support Vector Machine,

Multilayer Perceptron, and Random Forest algorithms
E. Syed Mohamed a , Tawseef Ahmad Naqishbandi a ,∗, Syed Ahmad Chan Bukhari b , Insha Rauf c ,
Vilas Sawrikar d , Arshad Hussain c
a
Department of Computer Science and Engineering, B.S. Abdur Rahman Crescent Institute of Science and Technology, Chennai 600048, India
b
Division of Computer Science, Mathematics and Science, Collins College of Professional Studies St. John’s University, NY, United States of America
c Department of Psychiatry, Government Medical College Srinagar, Jammu and Kashmir 190010, India
d Department of Clinical and Health Psychology, University of Edinburgh, Edinburgh, United Kingdom
ARTICLE INFO ABSTRACT

Keywords: The prevalence and burden of mental health disorders are on the rise in conflict zones, and if left untreated,
Mental health they can lead to considerable lifetime disability. Following the repeal of Article 370, political unrest spread
Machine learning quickly, forcing the Indian government to impose safety precautions such as lockdowns and communication
Prediction
ban. Consequently, the region of Kashmir experienced a marked rise in anxiety as a result of these lifestyle
Support Vector Machine
changes. Machine learning has proven useful in the early diagnosis and prognosis of certain diseases. Therefore,
Multilayer Perceptron
Random Forest
this study aims to classify anxiety problems early by utilising a pre-clinical mental health dataset collected
after the abrogation of article 370 in Kashmir. The first part of the paper aims at developing and implementing
a prediction model based on classification into one of the five pre-clinical anxiety stages, i.e., Stage 1: minimal
anxiety, Stage 2: mild anxiety, Stage 3: moderate anxiety, Stage 4: severe anxiety, and Stage 5: very severe
anxiety. The second part offers recommendations for those suffering from anxiety disorders. Feature selection
and prediction are used to predict the correct stage of anxiety for best possible medical intervention. Three
different algorithms: Support Vector Machine(SVM), Multilayer Perceptron (MLP), and Random Forest (RF),
are employed for predicting anxiety stages. Among them, random forest (RF) achieved 98.13% accuracy. A
forecasted likelihood condition was assessed to provide a suitable recommendation. Further, accuracy and
kappa statistics are used to assess the performance of the suggested model, which offers a significant addition
to predicting anxiety early, and exhibits high prediction and recommendation accuracy. This study aims to
assist mental health professionals and experts in making quick and accurate choices.
1. Introduction to describe the factors behind mental health suffering [3–5]. These
factors vary from one situation to another or one area to another
Anxiety and Depression affect an estimated 264 million individuals e.g., conflict or war zone give rise to totally different sort of factors
globally, making it one of the top causes of disability.1 According to the as compared to non-conflict areas. Mental health state is affected by
new Lancet Committee report, mental health disorders will upsurge in day-to-day environment, the place you live, living conditions, con-
every country without exception and it will cost the world’s economy
flict, war, lifestyle, family history, and mostly socio-economic factors.
$16 trillion by 2030.2 and the world economy loses $1trillion per year
Traditionally mental health professionals use face-to-face interviews,
due to anxiety and depression alone.3 In 2018, the World Health Orga-
self-reporting and questionnaire distribution to distinguish between
nization (WHO) published a guideline on managing physical problems
of Individuals with severe mental disorders, which if left untreated the mental health status of individuals. However, in clinical settings,
can result in a significant lifetime burden of the disease [1,2]. The disorders are identified using the statistical manual of mental disorders
rise in the prevalence and global burden of mental illness, prevention (DSM–5) [6]. The clinical staging paradigm allows that different levels
and treatment has become a public health priority and it is difficult of care to be assigned and early intervention to be delivered to slow
∗ Corresponding author.
E-mail addresses: syedmohamed@crescent.education (E.S. Mohamed), mtechnb@gmail.com (T.A. Naqishbandi), Bukharis@stjohns.edu (S.A.C. Bukhari),
insha.rauf@gmail.com (I. Rauf), vilas.sawrikar@ed.ac.uk (V. Sawrikar), arshadtina@gmail.com (A. Hussain).
1
https://www.who.int/news/item/13-04-2016-investing-in-treatment-for-depression-and-anxiety-leads-to-fourfold-return (accessed on 20th December 2021)
2
https://www.who.int/teams/mental-health-and-substance-use/mental-health-in-the-workplace. (accessed on23 rd Dec 2021)
3
https://www.who.int/teams/mental-health-and-substance-use/promotion-prevention/mental-health-in-the-workplace. (accessed on23 rd July 2022)
https://doi.org/10.1016/j.health.2023.100185
Received 17 November 2022; Received in revised form 30 March 2023; Accepted 24 April 2023
2772-4425/© 2023 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
E.S. Mohamed, T.A. Naqishbandi, S.A.C. Bukhari et al. Healthcare Analytics 3 (2023) 100185
down participation or prevent the emergence and recurrence of certain in the psychological field is still relatively underdeveloped [28]. In
diseases [7]. general, AI-based innovative diagnostic, monitoring, and therapeutic
A significant shortage of specialised mental health practitioners methods in psychiatry may enhance mental health outcomes, reduce
exists, such as psychiatrists and psychiatric nurses, for example, 8.59 the treatment gap, reduce inequality, public health administration and
and 29.15 per 100,000 persons in HICs, respectively. LICs, on the other minimise practitioner workload [29–33].
hand, have 0.5 psychiatrists and .42 psychiatric nurses per 100,000 Therefore, we propose a risk prediction model based on Pre-clinical
residents4 . The consequences are far-reaching, affecting people in prac- multistage classification strategy to classify individuals early before the
tically every aspect of their lives, including academics, careers, fam- onset of disorders based on severity level, i.e., Stage_1: Minimal Anxiety
ily, friends, and community participation. Conflicts are detrimental to (MA); Stage_2: Mild Anxiety (MIA); Stage_3: Moderate Anxiety(MDA),
the population. Destabilisation, aggression, and pervasive insecurity Stage_4: Severe Anxiety(SA) and Very Severe Anxiety(VSA). A risk
have well-documented ramifications for the day-to-day operation of a prediction model that analyses data to predict an individual’s risk
healthcare system [8]. Considering the interdependence of vulnerable of experiencing a mental health crisis is suggested. Individual risk
communities, and frequently declining public services, healthcare in prediction models are significant in two ways, firstly short-term risk
these situations is inevitably tricky [9]. Mental health needs are high forecasts can significantly improve outpatient psychiatric management,
in conflict settings with the meagre system, exacerbated by the crisis, such as in the detection of moderate anxiety in generalised anxiety
and mental health professionals’ shortfalls complicate mental health disorder. Long-term risk prediction, on the other hand, would allow
services [10]. Moreover, there is a serious lack of psychiatrists in India, for the focused administration of preventative actions in the early
with just 0.75 psychiatrists for every 1 lakh people. Along with an in- stages of a disorder or even before the start of symptoms. Individual
crease in drug abuse, substance use, loneliness, and suicidal tendencies risk prediction, on the other hand, could considerably improve the
during the double lock down (i.e. Shutdown after abrogation of Arti- efficiency of the development and evaluation of preventive treatments
cle 370 & COVID-19) mental health issues have become increasingly by focusing research efforts, especially on at-risk individuals. Once
prevalent [11–13]. the proposed prediction model’s results are evaluated, individuals are
A revolution in medicine and health care has been spawned by given recommendations based on risk evaluation and level of sever-
AI concurrently. AI is viewed as an innovative tool for organising ity score. The accuracy and kappa statistics are used to assess the
mental health treatments, as well as for identifying and keeping tabs performance of the suggested prediction system. Experimental and
on mental health issues in both individuals and groups. AI-driven clinical recommendations regarding severity assessment is carried and
solutions can employ digital health-care data, which is available in a established based upon a knowledge database developed with the help
variety of formats including electronic health records, medical pictures, of professional assistance from mental health specialists. Moreover, risk
and handwritten clinical notes, to automate jobs, support doctors, and assessment analysis is carried out to provide correct advice to patients
further our understanding of the underlying mechanisms of compli- or individuals.
cated disorders. The potential for artificial intelligence to help with So far, very few researchers have adapted machine learning tech-
this shortage. Artificial intelligence (AI) systems have the ability to niques to identify and predict pre-clinical anxiety stages in conflict
eliminate unjustified variation in clinical practise, enhance efficiency, settings. In an effort to fill this gap, this research has put forth a
and prevent avoidable medical errors, which may eventually prove to cutting-edge answer. The purpose of this study is to pinpoint the stage
be cost effective despite the high cost of the technologies utilised in of anxiety a person is experiencing and to identify the major triggers
these systems [14].
of anxiety. Also, this study aims to shorten the screening period for
While AI technology is becoming more prevalent in medicine for
anxiety stages. The following are the main contributions of this work:
physical health applications, the discipline of mental health has been
slower to adopt AI particularly in conflict settings [15–17]. Addition- • Identifying the key sociodemographic and psychological elements
ally, compared to most non-psychiatric professionals, mental health that contribute to various anxiety stages.
experts focus mostly on ‘‘smoother’’ abilities such as developing con- • Classifying anxiety stages early by utilising a pre-clinical mental
nections with patients and closely watching patient behaviours and health dataset.
emotions. They are also more hands-on and patient-centred in their • Diagnosing anxiety stage disorders using feature selection and
clinical practice [18]. Pre-clinical information on mental health is prediction.
frequently collected through patient interviews, written notes, and • Providing appropriate mental health recommendations based on
subjective, qualitative patient remarks. AI has great potential to re- the prevalence of suspected anxiety stage disorder.
define clinical intervention and understanding of mental illnesses [19]. • Confirming the fairness and correctness of the recommender sys-
Utilising AI techniques enables the creation of improved pre-diagnosis tems as suggested by professional mental health experts.
screening tools and the creation of risk models to assess a person’s
propensity for or risk of developing mental illness [20]. To implement The sections of the manuscript are as follows: - Section 2 discusses
personalised mental healthcare as a long-term goal, we need to harness related research. Section 3 talks about proposed system and is further
computational approaches best suited for conflict settings. Although halved into two sections i.e. classification and prediction of various
much work has been done to predict mental health in clinical and anxiety stages, Section 4 delves deeper and talks about Experimental
non-clinical settings [21–23], there is still a considerable need to syn- Outcomes and discussion, and finally, Section 5 concludes the article.
chronise machine learning predictions with pre-clinical mental health
anxiety stages in order to deliver individually tailored mental health 2. Related research
recommendations, as cultural context of mental health reporting varies
in Asian subcontinent as compared to western subcontinent [24,25]. The components of this type of study problem are extremely com-
Further, finding hidden Patterns can be challenging which can oth- plicated, necessitating a thorough investigation. In order to determine
erwise simplify pre-clinical mental health decision making and can the methods and approaches employed in the previous works and the
make life better [7,26,27]. Though machine learning algorithms are research gaps, this portion reviewed a number of related research arti-
frequently utilised in the medical and health fields, their application cles. Recent technological breakthroughs and advancements in AI/ML
approaches have enabled the development of more effective prediction
4
World Health Organization. (2011) Mental health atlas 2011. and decision-making tools [34,35]. These tools enable medical prac-
World Health Organization. https://www.who.int/publications/i/item/ titioners to choose an effective treatment strategy for mental health
9799241564359 (Accessed on 10th November 2022). disorders at early stage. Data analysis in medical field has numerous
2
problems, including low quality data, scare resources of data particu- The depression range in this instance was determined from the socio-
larly in case of mental health due to stigma association and the majority demographic information’s properties as the predictor factors. The
of the data is unstructured in the form of clinical notes which makes desired variable was the BDI test. This study contrasted the effective-
analysing difficult for its intended use. Furthermore, the inclusion of ness consisting of three distinct algorithms: Artificial Neural Network
unlabelled data, bias, as well as inconsistent data makes mental health (ANN) [50] utilising backpropagation learning and an extreme learning
prediction challenging [36,37]. However, machine learning algorithms method genetic algorithm and fuzzy with it.
have recently gathered considerable amount of attention in healthcare In a relevant study by Hassan et al. [51], they used unsupervised
analytics ranging from classification to revealing hidden patterns in techniques to create a model for diabetes detection. The constructed
large amount of data [23,38]. Various types of machine learning ap- cluster-based dataset and whole dataset were then subjected to the
proaches and algorithms have been employed in the past in various application of multiple algorithms (Multilayer Perceptron, RF, DT,
field of healthcare for predicting the risk of disease in patients to predict SVM, and KNN), and assessment results reveal that RF attained the best
illness intensity in patient [39]. Moreover, various research articles accuracy of 99.57% on the cluster-based dataset.
provide a discussion of the model’s development, validity, and utility Charlson et al. [52] estimated that depression and post-traumatic
of clinical staging for mental health [40–43]. stress disorder in conflict-affected environments are five times higher
In a thorough review of machine learning prediction approaches for than average estimates worldwide. Further, Conflicts have repercus-
anxiety disorders, Pintelas and colleagues., [44] conducted research. sions that have been extensively researched using various mental health
They arrived at the conclusion that the accuracy is determined by the evaluation techniques. Ashley Moore et al. [53] evaluated 27 mental
type of prediction techniques and data acquisitions, such as clinical health measures often used in conflict settings to assess the population’s
data, self-reporting data, or screening tools. They found that, out of incidence and mental health service demands to swiftly recognise indi-
the 16 research, hybrid approaches and support vector machines(SVM). viduals who need assistance during distress, focusing on anxiety and
were the most widely used methods for forecasting post-traumatic PTSD.
stress disorder (PTSD) and seasonal affective disorder (SAD). By applying the gradient boosting method, Chekroud., et al. [54]
The researchers Attaran and Gokhan [45] suggest that with the constructed a machine learning algorithm to predict the clinical remis-
development of digital twin technology a lot of interest have been sion from (n=1949) patients who had experienced stage 1 depression
received from both business and academics. The authors are of the and attained an accuracy of 64.6 percent. Ahmed et al. [55] have
view that clinical research could be completely changed by digital twin
presented a paradigm for the early detection of depression and anx-
technology as, Digital twins can be utilised to obtain better answers,
iety. The machine learning techniques convolutional neural network,
gain actionable insights, and forecast how experimental treatments will
support vector machine, linear discriminant analysis, and K-nearest
affect a patient without endangering their lives. They are further of
neighbour have been used to categorise the intensity level of the
the view that technology also aids in the examination of the infected
anxiety and depression, which consists of two data sets. According to
patient by medical personnel.
the data, the convolutional neural network had the highest accuracy,
R. Garriga et al. [46] are of the view that,it is possible to achieve
achieving 96.8% for depression and 96.6% for anxiety. The support
better results and reduce burdens and expenses by quickly identifying
vector machine also performed well, achieving accuracy of 95.6% for
individuals who are at danger of a mental health crisis. Unfortunately,
anxiety and 95.8% for depression. In addition, the accuracy of the
because of the elevated incidence of mental health issues, it is not
linear discriminant analysis was 93 percent for anxiety and 87.9 percent
practical in reality to manually evaluate complex patient records in
for depression.
order to make proactive care decisions. In order to continually monitor
In a recent survey by Doraiswamy et al. [56] about 48.7% of
patients for the likelihood of experiencing a mental health crisis over
respondents felt that AI/ML would have no influence or only minimal
the course of 28 days, the researchers constructed a machine learning
influence on the future work of psychiatrists over the next 25 years, and
model that uses electronic health records.
47% of respondents felt that their jobs would be moderately changed
Pintelas and colleagues., [44] conducted a systematic review of
machine learning prediction methods for anxiety disorders. They came by AI/ML over the next 25 years as there are not enough mental health
to the conclusion that the type of prediction methods and data acquis- professionals to treat the overwhelming mental health issues among
itions—such as clinical data, self-reporting data, or screening tool people. At the same time, many psychiatrists believe that Artificial
determine the accuracy. They discovered that, out of the 16 studies, the intelligence (AI) and related technologies are the need of the hour to
most popular technique for forecasting post-traumatic stress disorder be adapted for mental healthcare to minimise burnout while working
(PTSD) and seasonal affective disorder(SAD) were Hybrid methods and in conflict settings.
Support Vector Machine (SVM). Ganie and Malik [57] used Ensemble classifiers to diagnose diabetes
In order to diagnose CKD in patients Sawhney and colleagues [47], in its early stages. They obtained inpatient, outpatient, and emergency
suggested that by employing a deep neural network-based multi-layer department data from hospitals in Jammu and Kashmir-UT, India. This
perceptron classifier. The suggested model achieves 100% testing accu- study compared the performance of many traditional machine learning
racy in classification tasks, according to experiments. The main contri- algorithms to the effectiveness of three ensemble learning techniques,
bution of the research is a Deep Neural Network model for chronic kid- namely Voting, Boosting, and Bagging. The results showed that Bagged
ney disease diagnosis that outperforms conventional machine learning Decision Tree surpassed the others with an accuracy rate of 99.14%.
models and achieves 100% accuracy. Talha et al. [58] underline the serious negative side effect of anxi-
D. Brathwaite et al. [48] findings confirm that age and sex have ety, which include emotional imbalance, unhappiness, stress, and alien-
a role in the psychological needs of a patient in the emergency de- ation The information was obtained in three different forms: physically,
partment. All age groups, including teenagers with suicidal thoughts, digitally, and through health records. In this study the Naive Bayes 71%
middle-aged men with alcohol misuse, and elderly women with demen- optimistic argument explains how human behaviour has a negative
tia, rely on the emergency department for psychiatric care. impact on overall health. Finally, they established a link between the
Jovana [49] conducted research on depression risk among people effects of their proposed specialisation and the outcomes of the three es-
with breast cancer. 84 patients, ranging in age from 30 to 78 years, sential reference points. Numerous researchers [59] have looked at how
were used in this study, and information was gathered by using a well AI models diagnose mental illnesses like depression, anxiety, post-
Two-stage interview. The initial stage involved gathering the patients’ traumatic stress disorder (PTSD), bipolar disorder (BD), schizophrenia,
sociodemographic data. In the second session, the patients underwent and Alzheimer’s disease (AD). The evidence deriving from these in-
the standardised Beck Depression Inventory (BDI) test phase. The pa- vestigations has been summarised in numerous systematic reviews. To
tients’ true depression thresholds were assessed using the BDI test. our knowledge, no prior studies have been published to summarise
3
the evidence about the diagnosis and prediction of anxiety stages algorithm’s bias. It is based on distance, consistency, dependency,
for mental disorders, despite the fact that conducting a stage based information, and correlation as generic features of the training data.
prediction of anxiety disorder is important to draw more accurate and A knowledge-based personalised recommendation system using con-
thorough conclusions on a specific topic. textual data is used to identify the disease and uses collaborative
For the purpose of using machine learning algorithms to diagnose filtering to treat it. For disease prediction and recommendations, how-
various mental diseases, Priya and her colleagues [60] analysed data ever, all of these systems rely on fuzzy inferences. Recommendations
on anxiety, stress, and depression. They discovered that the Naive are frequently issued to assess diseases without the use of a cutting-
Bayes classifier performed the best for predicting depression, with an edge algorithm. Further, to assess whether a person is sad or satisfied,
accuracy of 85.50%. In order to provide crucial medical advice to in- the researchers employed Random Forest Classifier, Random Forest
dividuals that suffer, from mental health disorders, intelligent machine Regression, Naive Bayes, and K Neighbours Classifier [81]. The re-
learning techniques and strategies play a crucial role. Further, in some searchers subsequently looked at important characteristics that may
research studies, we found the use of regression techniques such as lead to anxiety in order to determine whether or not a person is having
linear regressions, log-linear regression, correlation analyses, and sur- anxiety or not. Finally, investigators were able to predict the percentage
vival analysis/Cox regression [61–63] have been used to predict health of males and females that are depressed or happy.
related issues. Additionally, recent studies has aimed at identifying the From the foregoing discussion, it can be concluded that the majority
various clinical stages of depression and associated phases [64–66] of the research in existence have been carried out to forecast depression
A. Othmani and A. O. Zeghina [67] suggested a paradigm for in a certain context, such as: individuals in a particular age range,
identifying depression and predicting depression relapse are highly individuals with a particular ailment, etc. This study makes an attempt
encouraging. They identified depression with an accuracy of 78.97%. to get over this limitation by considering various Pre-clinical socio-
The high-performing similarity measure-based technique for detecting demographic data of persons living in conflict zone. Further, it is
relapse depression achieved an accuracy of 82.55%, while the proposed noteworthy that the majority of the examined studies either use no
correlation-based anomaly detection framework acquired an accuracy sample size at all or only use a tiny sample size in their experiments.
of 80.99%. There are no studies on the conflict hit regions predicting anxiety
To assess the severity of depression, Zarandi et al. [68] used type-2 stages using ML as, predicting anxiety stages using machine learning
fuzzy logic. To improve the study’s precision and predict the patients’ is challenging especially in conflict areas due to non-availability of
level of depression with fewer questions, they used the Mutual Informa- open source data. Further, this paper intends to use real-world data
tion Feature Selection (MIFS) technique. Their suggested method had collected from Kashmir India to predict various anxiety stages using a
an accuracy of 84.00% with just fifteen questions to predict the severity carefully chosen machine learning approach that comprises SVM, MLP
of depression. and RF. Additionally, a methodical feature selection process was used
Thompson and his peers [44] reportedly looked into why people put to identify the 10 best characteristics out of the 20 available features
off getting help for anxiety and mood disorders. They discovered that that had the highest accuracy. Further, we introduced a risk and
the average person waits 8.2 years before seeking treatment. Addition- probability focused recommender model which is adaptive in nature
ally, they identified two key markers—slower problem awareness and and is driven by machine learning based prediction system.
younger age at onset—associated with this delay. Due to their slower
contact with initial treatment, older patients. Early anxiety prediction 3. Methodology
utilising machine learning models could successfully stop this [8]. Some
recent study for automatic anxiety and depression clinical staging for This section is divided into three subsections. The entire methodol-
mental health disorders by [69,70] encouraged this proposed work to ogy of this study has been described in the following subsections
develop pre-clinical mental health anxiety prediction model, So that
• Sample and Data
individuals can know their mental health status earlier before the onset
The mental health sample dataset in this study consists are 215,
of mental health disorder.
aged between 25–36 years of age. The short discretion of the
Broekharst and colleagues [71] contend that increasing patient
collected data is given in Table 1. Each variable’s potential value,
agility would increase patient value. However, the abilities that allow
variable type, and variable description as displayed in Table 1.
hospitals to recognise the patients’ needs for healthcare services and
One target variable and thirty predictor variables make up the
adapt to their evolving needs are collectively referred to as patient
dataset that was acquired from a well-known psychiatric clinic.
agility. In addition, this study makes the claim that hospitals’ exist-
By administering the General Anxiety Disorder-7 (GAD-7) to each
ing sensing and responding capacities are essentially reactive because
participant, the target variable was created.
they only recognise, respond to, and seek patterns and occurrences in
• Measures of variables
patients’ wants and demands after these have already materialised.
The status of the mental health is measured in terms of anxiety
To avoid a mental health crisis, early planning is essential. Recently,
level a person is suffering form. The collected predictor data
researchers developed a recommendation system (RS) to provide quick
variables are ordinal in nature and target variable is having 4
and timely medical recommendations to cardiovascular and mental
different ordinal levels, i.e., minimal, mild, moderate and severe.
health patients [72–75]. Previous work on building a clinical decision-
The explanatory variables considered are shown in Table 1.
making prediction methodology is based on the theoretical clinical
staging framework [31,69,70,76]. In addition, intelligent approaches
and procedures play an important role in offering appropriate mental 3.1. Proposed model for prediction and recommendation of anxiety pre-
health suggestions. These suggestions can help in improve their quality clinical stages
of life and daily behaviour by minimising the workload and costs as-
sociated with everyday healthcare activities. However, many research The proposed methodology is a hybrid of two models: Mental Health
publications did not seek to harness established theory in clinical Anxiety Stage Disorder Prediction Model (MH_ASD_PM) and Anxiety
psychiatry to map or ground the state they explored. The other feature Stage Mental Health Recommendation Model (AS_MH _RM) based on
selection strategies included manually eliminating features by abla- statistical analysis that can assist psychiatrists and individuals in mak-
tion method [77]. The algorithms of the filter model are Relief [78], ing quick clinical decisions regarding mental health disorder diagnosis
Fisher score [79], and Information Gain based methods [80]. The filter and medical recommendations. Various researchers have reported the
model isolates feature selection from classifier learning, ensuring that notion of anxiety [82] despite the fact that none of them define the
a learning algorithm’s bias does not interfere with a feature selection severity or level of stage nor they know what they mean when they
4
Table 1
Variables for predicting Pre-clinical anxiety stages.
Variable name Variable type Variable description Possible values/feature units
Q1 Predictor Feeling nervous, anxious, or on edge {0 = Not at all; 1 = Several day ; 2 = More than half
the days; 3 = Nearly every day}
Q2 Predictor Not being able to stop or control worrying {0 = Not at all; 1 = Several day ; 2 = More than half
Q3 Predictor Worrying too much about different things {0 = Not at all; 1 = Several day ; 2 = More than half
Q4 Predictor Trouble relaxing {0 = Not at all; 1 = Several day ; 2 = More than half
Q5 Predictor Being so restless that it is hard to sit still {0 = Not at all; 1 = Several day ; 2 = More than half
Q6 Predictor Becoming easily annoyed or irritable {0 = Not at all; 1 = Several day ; 2 = More than half
Q7 Predictor Feeling afraid, as if something awful might {0 = Not at all; 1 = Several day ; 2 = More than half
happen the days; 3 = Nearly every day}
Gender Predictor Gender of the participant Male = 1; Female = 0
Marital status Predictor Marital status of the participant Unmarried = 0; Married = 1; Widowed = 2;
Divorced=3; Separated=4
Age Predictor Age of the participant 25–36
Location Predictor Location of the participant Rural = 0 , Urban = 1
Field Predictor Field of study of the participant Arts = 1, STEM = 2
Phase Predictor Phase of study (Early Phase, Later Phase) EarlyPhase = 1, Later Phase = 2
Income Predictor Income of the participant Number
Anxiety Predictor Total Anxiety score of the participant Number
Anxiety level Target Anxiety level of the participant MA = Minimal Anxiety, MIA = Mild Anxiety, MOD =
Moderate anxiety, SA = Severe anxiety
Fig. 1. Depicting proposed mental health prediction and recommendation methodology.
examine this condition. Anxiety is a symptomatology warning that can stages. The AS_MH_RM then uses the prediction information, as well
impact overall health badly, similar tendencies emerge when it comes as the medical history, to recommend general mental health advice
to the concept of depression, the researchers regularly and discreetly to individuals suffering from a particular predicted anxiety level. The
imply that they are mentioning anxiety and depressive disorder; nev- following subsections describe each phase.
ertheless, these stages are hardly explained. Fig. 1 depicts the steps • Mental Health Anxiety Stage Disorder Prediction Model
involved in the suggested methodology. MH_ASD_PM employs data pre- (MH_ASD_PM)
processing to eliminate missing data and deal with outliers in mental The entire chain of operations is passed down in the mental health
health data obtained from a credible institution. prediction model as depicted in Fig. 2. We collected a Pre-clinical
The mental health dataset is subjected to the feature selection mental health dataset of individuals from a famous psychiatric clinic,
approach, which selects the most essential and informative attributes whose details need to be kept confidential for privacy. The dataset is
to aid the multi-classification process. Finally, one of the five stages split, and the ‘‘synthetic minority over-sampling method’’ is performed
is predicted using the specified features of General Anxiety disorder to mitigate imbalances if present. To eliminate noise and ambiguity
5
Fig. 2. Proposed detailed architecture for Anxiety Stage disorder prediction model.
from the training dataset, it is cleaned. Proposed Algorithm1 below it. As a result, we may conclude that data pre-processing is an
describes the specifics of MH_ASD_PM. essential stage in the development of an efficacious classifica-
tion model [83]. During data pre-processing, duplicate records,
3.2. Data analysis and models spelling errors, and incorrect data if found is removed. Accuracy
of classification is reduced due to significant rate of missing
• Data analysis values [84,85]. Without using any classification techniques, a
The significance of association between the underlined attributes numerical cleaner filter [26] is used to find missing values in the
and dependent variable is highlighted using the concept of cross- data and evaluate features. By changing numeric data to a present
tabulation, and the results drawn are given in Table 2. It is value, this filter cleans numeric data that is either too large or too
clear from Table 2 that there is significant association between little. Further, data Cleaning in machine learning is a crucial step
levels of anxiety and Gender (𝑃 -value = 0.013), Location (𝑃 - in the data pre-processing pipeline that helps to ensure that the
value =0.027), Stream (𝑃 -value = 0.016), and Phase of Ph.D. dataset used to train or evaluate the model is accurate, consistent,
(𝑃 -value =0). Out of total of 215 sample units 41.86% are males, and relevant. Data cleaning helps to reduce bias and variability
58.13% are females, 61.40% are rural, 38.6% are urban, 60.46% in the model, resulting in more accurate and reliable predictions.
are arts studying, 39.53% are science, technology, engineering, • Feature selection
mathematics (STEM) studying, 53.48% are in the first phase of The dimensionality of data used in machine learning tasks has
Ph.D. and 46.51% are in the second phase of Ph.D. 12.25% are erupted in the last four decades, and posing major problems like the
found to be suffering from minimal, 23.25% from mild, 33.48% dimensionality curse [20,86] to present learning methods [33,84].
from moderate and 30.68% from severe anxiety. Among professionals, feature selection is a widely used technique for
• Data pre-processing reducing the dimensionality [87]. It strives to remove a specific subset
Any prediction model’s performance is measured in terms of how of irrelevant features from the original set based on a relevance assess-
well it predicts the future and is acknowledged to be a function ment criterion. Feature selection usually results in improved learning
of a number of parameters. In other words, data is not always performance (e.g., higher classification learning accuracy), lowering
clean there may be redundant data features, inconsistent features, computing cost, and improved model interpretability. Feature selection
noise, and/or missing data in a dataset. Before applying machine aims at identifying a sample of highly distinguishing variables for
learning techniques, the data needs to be processed by removing diverse mental health diagnoses. In this multi-label categorisation pro-
the redundancy, inconsistency, noise and missing values from cedure, every data point could be correlated with many labels. There
6
Table 2
Chi-square test statistic.
Abbreviations:- MA: Minimal Anxiety; MIA: Mild Anxiety; MDA: Moderate Anxiety; SA: Severe Anxiety.
𝑝-value < 0.05 significant.
are a total of 16 features in the dataset, however only a few of them value of 0.1(discarding any features with a score below cut-off value
are important in the classification of anxiety stage disorders. As a result, yet preserving all high-ranking features). A total of ten features were
the dataset is subjected to the feature selection procedure in order to evaluated, with the results being entered into the prediction algorithm
reduce the input vectors size. With, feature selection information gain for a speedy decision.
method was employed as suggested by (Claude Shannon’s [88] because
of its advantage of being fast. For a specific class, the information 3.3. Description of the proposed prediction techniques
gain (entropy) of each feature is determined using this technique. The
entropy value goes from 0 to 1, where the highest number indicates the An overview of the machine learning algorithms used in the pro-
existence of the most relevant information. The highest ratings features posed project for pre-clinical anxiety stages can be found in the sub-
contribute the most to the decision-making process, thus they are sections that follow.
kept, and lower scoring features are removed. Relying on information
gain of a feature X, the most appropriate dividing criterion feature • Support vector machine (SVM)
is determined by Eq. (1). For generating decision tree, the highest SVM algorithms are classifiers that are characterised by a linear
information gain is selected and is calculated as, decision boundary known as a ‘‘hyperplane’’. SVM is one of the
most well-known methods for improving the desired outcome.
𝐼𝑛𝑓 𝑜𝑟𝐺𝑎𝑖𝑛(𝑋) = 𝐼𝑛𝑓 𝑜𝑟 (𝐷) − 𝐼𝑛𝑓 𝑜𝑟𝑋 (𝐷) (1) The fundamental purpose of SVM is to separate classes in the
where 𝑋 is the feature under consideration and 𝐷 represents dataset. training data by using a surface that maximises the margin be-
The Information essential to classify any lone feature can be determined tween them. Initially, support vector machines were used to
by applying the formula given below, classify linear binary data. They can, however, create non-linear
outcomes. Kernel functions can split data into classes by adding
∑
𝑚
𝐼𝑛𝑓 𝑜𝑟(𝐷) = − Pr 𝑖 log2 (Pr 𝑖 ) (2) more features in the higher dimension [90] . SVMs are most
𝑖=1 likely to succeed if they use a data-driven algorithm strategy
In Eq. (2), 𝑚 is the total number of classes and Pr 𝑖 is the probability that when working with datasets with less samples than the normal
an incidental instance in 𝐷 which belongs to class 𝐶. Eq. (2), is used to variables, which is why they are so popular in medicine for
quantify the single-label Information gain attribute ranking approach to the purposes of disease prognosis and diagnosis [91]. An SVM
assess a feature’s capacity to differentiate between distinct class values. improves the nearest distance of the object whether the outcome
The information needed after selecting feature 𝑋 to divide the dataset is positive or negative. Following that, the objects are categorised
𝐷 into distinct segments is computed by: according to which side of the hyper-plane they are on. Because
most real-world situations are not binary, Support Vector Ma-
∑𝑣 ||𝐷 ||
| 𝑗 | 𝐼𝑛𝑓 𝑜𝑟(𝐷 ) chines (SVMs) are designed to work with multiple classes. As
𝐼𝑛𝑓 𝑜𝑟𝐴 (𝐷) = 𝑗 (3)
𝑗=1
|𝐷| a result, many variants of SVM are employed, including one to
one and one v/s all [92]. In this study, the data set is trained
In Eq. (3), |𝐷𝑗 | indicates total records that have a feature value 𝑗 in using a pair wise classification using SVM with each pair of
dataset 𝐷, all records in the dataset are represented by |𝐷|, 𝐷 and given classes [93]. This classification has 𝑁 classes, which yields
𝑣 represents entire set of feature values. For multi-label data [89], 𝑀𝑐 𝑀(𝑀−1)
number of binary classifiers. Where 𝑀 𝑐 2 uses
2 = 2
Some researchers have used the C4.5 algorithm to handle decision binary classification algorithm to count how often point class
tree algorithms with numerous labels at the tree’s leaves, based on an
label is entrusted with’𝑦’. Unknown occurrence ’𝑦’ is allocated to
entropy calculation adaption stated in Eq. (4).
the highest-counting class. The data is translated into hyper plane
∑
𝑙
( ) ( ) ( ) ( ) using a polynomial kernel.
𝐼𝑛𝑓 𝑜𝐴 .𝑀𝐿(𝐷) = − 𝜌 𝜆𝑖 ∗ log 2𝜌 𝜆𝑖 + 𝑞 𝜆𝑖 ∗ log 2𝑞 𝜆𝑖 , (4)
𝑖=1 • Multi-Layer Perceptron (MLP)
( )
In Eq. (4), 𝜌 𝜆𝑖 represents the probability that an erratic instance in Multilayer Perceptron (MLP) is a very useful approach for health-
𝐷 corresponds to class label, 𝑞(𝜆𝑖 ) = 1 − 𝜌(𝜆𝑖 ) and 𝑙 is the number care research [94]. A feed forward Artificial Neural Network is
of labels in total within the data set. To handle multi-label data, a popular term for it. The Multilayer Perceptron is made up of
the Eq. (4), is used to develop an information gain feature selection an input layer and an output layer that are linked together by a
process. Using this strategy, feature selection can be utilised with any series of hidden layers. The data to be processed is fed into the
multi-label classifier. It uses a ranker search algorithm to rank features input layer. The outcome layer performing tasks like prediction
evaluated in the light of class labels that have been assigned. Features and categorisation. The designed computational structure of the
are chosen and eliminated in the ranker approach based on a cut-off MLP is made up of an erratic number of hidden layers that reside
7
account for the suggested model. Final, recommendations are drawn

based upon knowledge base created with the help of a clinical psy-
chiatrist. It assists in identifying the importance and primary concern
of environmental factors that plays a significant role in predicting
anxiety stage disorders. This recommendation model would benefit
both individuals and psychiatric practitioners by reducing workload
and time costs and will significantly help low-income individuals from
huge financial loss. Additionally, Fig. 3 depicts a high-level overview
of anxiety stage General recommendation model based on statistical
analysis, which is described in the next subsections that follow.
• Risk Exposure Analysis
Risk factors is a crucial factor associated with an individual. These
might be the features, variables, or hazards that, if present in a specific
Fig. 3. A Comprehensive general recommendation model (AS_MH _RM) architecture individual, increase the likelihood of that individual developing a
based on statistical analysis.
disorder. The signs and symptoms of psychological disorder vary, based
on the population studied, the circumstances, and a variety of other
between the input and output layers. It consists of perceptrons, variables. Mental illness symptoms can affect emotions, thoughts and
which are simple nodes that can produce a lone output from behaviours. Depending on the level of exposure anxiety, it can lead to a
numerous inputs by first allocating weights and then constructing variety of diseases as well as long-term disabilities. When an individual
a linear model of weighted input’s. A non-linear kernel function is diagnosed with panic attack, the severity of the risk exposure related
is used to calculate the results, expressed in Eq. (5), below
( 𝑛 ) to panic attacks should be understood because it can progress to a
∑ ( ) severe mental health disorder. The objective is to generate a model by
𝑦=𝛷 𝑤𝑖 𝑥𝑖 + 𝑏 = 𝛷 𝑤𝑇 𝑥𝑖 + 𝑏 (5)
estimating the risk with each variable against a set of vulnerabilities.
𝑖=1
The risk factor is calculated using the following formula:
In Eq. (5), ‘𝑤’ represents vector weights as input,‘𝑦’ is the input com-
𝑃 (𝑑)
bination, ‘𝑦’the bias and the kernel function is represented by ‘𝛷’. To 𝑅𝑖𝑠𝑘 = (6)
train the MLP, back-propagation approach is used which consists of two 𝑃 (𝑑)
phases. Forward phase is the first phase, which uses Eq. (5) to evaluate Where 𝑃 (𝑑) and 𝑃 (𝑑) denotes the probability or likelihood of having a
the categorised outputs for the provided input data. The backward certain disorder respectively. 𝑃 (𝑑) is calculated by,
phase produces a partial derivative of a kernel function associated with
parameter changes and propagates it back into the network. After that, 𝑁𝑒
𝑃 (𝑑) = (7)
any gradient boosting algorithm can be used to update the network’s 𝑁𝑝
weights, and the procedure is performed till the network’s weights Where 𝑁𝑒 reflects the number consecutive vulnerabilities occurrences
converge. within class and 𝑁𝑝 represents the total number of individuals inside a
• Random Forest Algorithm class. Therefore, 𝑃 (𝑑) is derived by,
The Random Forest (RF) algorithm is a supervised training algo-
rithm. It creates a ‘‘forest’’ from a group of decision trees trained by the 𝑁𝑑
𝑃 (𝑑) = (8)
‘‘bagging’’ method. The bagging method’s key notion is that combining 𝑁𝑝
multiple learning models improves overall capacity [95]. It interrupts Where 𝑁𝑑 denotes the total number of non-vulnerable occurrences, and
the greedy splitting algorithm during tree construction, RF outperforms
𝑁𝑝 represents the number of individuals within a class.
bagged decision trees by allowing only a random percentage of the
• Multi-Classification VS Severity
input attributes to be used as split points. While the classification tree
evolves, RF evaluates only the bifurcation breakdown on an erratically Severity plays a vital role in identifying the specific mental health
picked sub-set which is similar to bagging. This approach uses "simul- disorder; hence multi-classification of anxiety disorder levels is based
taneously ensembling’’, wherein multiple decision tree models are run on exposure levels. The objective of this segmentation is to gain a
in parallel on various data set sub-samples, with voting determining deeper understanding of each stage for a variety of exposures. Eq. (9),
the conclusion or final result. The size of the random variable is is used to assess the likelihood or probability of all exposed occurrences
predetermined. As an outcome, the problem of overfitting is minimised, in comparison to every other class:
and prediction accuracy with control is enhanced. The RF adaptive ( ) 𝑓 (𝐷𝑅,𝑦 )
learning with several decision trees is significantly more reliable and 𝑃 𝐷𝑅,𝑦 = (9)
𝑁
accurate than a model based on a single decision tree [96]. It may be ( )
used to solve both classification and regression issues, and it works well Here, 𝑃 𝐷𝑅,𝑦 depicts disorder class occurring in the event of exposure
with both continuous and categorical data. levels, (R,Y) depicts two different levels of exposure (red and yellow),
𝑓 (𝐷𝑅,𝑦 ) where 𝑁 represents the number of class recordings.
3.4. Anxiety stage mental health recommendation model (AS_MH_RM)
based on statistical analysis
3.5. KBS -Knowledgebase system
The recommendations model is an important aspect of the proposed

system. Based on the severity of the Anxiety condition and the risk Building a knowledge data base is the initial stage in AS_MH_RM
to one’s mental health, the goal is to give an appropriate and timely as shown in Fig. 3. The knowledge base contains information on self-
advice. During the prediction, an intuitive and user-friendly recommen- reporting characteristics and their importance in diagnosing the clinical
dation model examines the data of the individual/patient to calculate stage of anxiety disorder. The knowledge base system is built with
the risk. Depending upon the severity and likelihood of disorder, accu- the help of psychiatric experts and the data already gathered. Fig. 4
rate clinical psychiatric suggestions for future course of action are then depicts the procedures needed in developing a knowledge database set
made. Proposed Algorithm 2 below depicts the procedures taken into up. Ten important indicators have been identified that aid in the early
8
detection of a particular anxiety level/stage. These indicators include The information of severity ranges is used to spot and classify one
some socio demographic and stages adapted from [71]. Each of the anxiety stage from another (Refer Fig. 7). Using a colour code, if the
specified parameters are given a weight based on its importance in total value is greater than 21, the patient/individual will be labelled
the diagnosis. As a result, the cumulative score defines the individual’s ‘‘white’’ for very severe anxiety, and if the value is 15–21, the patient
critical importance to be dealt with for example an individual with will be labelled as ‘‘red’’ for severe anxiety. Likewise, for moderate
a score of 15 should be given higher prime concern and medical anxiety if the severity score is in the range of 10–14 it will be colour
observation than rest individuals. Further, severity score of greater than code in ‘‘yellow’’, and for mild anxiety if the severity score is in the
20 can have serious traumatic psychiatric issues and can be integrated range of 5–9 it will be colour code in ‘‘green’’, while values below 5
within higher-tier mental health services. i.e. 1–4 are colour coded in ‘‘blue’’ for minimal anxiety (Refer Fig. 7).
9
Fig. 4. Knowledge database for recommendation.
A clinical mental health expert can forecast future mental health events For general recommendations in the current study, we have proposed
by combining the importance of weight and abnormality range. A the following stages for recommendation, which comprises of four
patient living in a non-conflict zone, in contrast with a patient living in stages. Table 3. shows the result for mental health recommendations
a conflict zone may be less exposed to mental health issues. Individuals of various anxiety stages.
with several irregularities in their specified ranges can be addressed
more effectively with time to time mental health recommendations 4. Experimental outcomes and discussion
and treatments. A mental health professional labels the dataset records
with the selected parameters for clinical recommendations. The expert The data and experimental arrangement utilised to assess the pro-
interventions are then used in the evaluation step to validate proposed posed predictions and recommendations model are presented in the
AS_MH_RM’s outcomes. Risk factor analysis can be computed against subsequent sections.
each condition category for the selected and exposed psychological
parameters when the knowledge base is created. • Mental Health Dataset
A novel real world pre-clinical mental health dataset was taken
3.6. Mental health anxiety stage recommendations from a reputed psychiatric clinic for multi-classification predic-
tion and recommendation of various anxiety stages. In terms of
Once computational statistical analytics is over, recommendations patterns and sources of variability, real-world data are typically
for individuals are developed by merging the findings of the analy- significantly more complex than simulated data [97]. No formal
sis. A predefined rule has been established from the knowledge base ethical approval was necessary because this study involves the
while taking into account risk components and probability estimations, retrospective study of anonymised data taken from third party
which are then used to develop an inferential protocol for propagat- and gathered as part of routine clinical operations, as required
ing recommendations from Algorithm2. The rule predefined stipulates by the declaration of Helsinki (Principal 1)5 .
that whenever a new individual/patient seeks mental health treatment • A total of 215 sampling units are present in the data set (which
arrives, tests are performed to determine the anxiety stage based on is far better in terms of dataset size than the previous studies
the severity score. Following computational predictions, spectrum of [98–107]) entailing those who may be suffering from minimal
exposure parameters is established, and calculations are determined to mild or very severe mental problems. For efficient decision
based on the spectrum, risk factor, and likelihood. Lastly, the acquired and support system the mental health professional tagged the
data is compared to the rules predefined to determine the severity training data to minimise major flaw. The dataset consists both
stage. The final score is obtained using the severity values and score male and female from 25 to 35 years of age, with about 65 percent
from the base of knowledge. The Eq. (10) for finding the final score is being male and the remaining being female. The patient’s age, as
as follows: well as its other features plays an important part in identifying
∑
𝑚
𝐹 𝑖𝑛𝑎𝑙𝑆𝑐𝑜𝑟𝑒(𝐹 𝑆) = 𝑆𝑖 (𝑤𝑖 ) (10) 5
https://www.wma.net/policies-post/wma-declaration-
𝑖=1
of-helsinki-ethical-principles-for-medical-research-involving-
In Eq. (10) 𝑚 is the number of vulnerabilities taken into account, 𝑆𝑖 de- human-subjects/#:∼:text=It%20is%20the%20duty%20of,
notes the severity, and 𝑊𝑖 denotes the importance of the vulnerability. the%20fulfilment%20of%20this%20duty.
10
Table 3
Mental health recommendation for anxiety severity stages.
Severity stages Labels Recommended interventions (RI) Severity score
range
Stage 1: Minimal Level 1: Self-directed e-health based anxiety management ≤4
anxiety monitoring and management
Stage 2: Mild Level 2: Low intensity care Level 1 and e-health or low intensity treatment of anxiety management, ≤9
anxiety plus exposure therapy with active general practitioner management
Stage 3: Moderate Level 3: Moderate intensity Level 2 and pharmacotherapy with active general practitioner ≤14
anxiety care management and support from mental health professionals
Stage 4: Severe Level 4: High intensity care Level3 and multidisciplinary care or multiagency support model ≤21
anxiety involving active general practitioner management, specialist physicians,
allied health professionals, and possible case manager
Stage 5: Very severe Level 5: Acute and specialist High intensity care with possible involvement of higher-tier mental ≥22
anxiety care health services if indicated by comorbid mental disorders (e.g., major
mood disorder, psychosis) or risk of self-harm/suicide
mental illness. Table 2 shows the dataset features (which includes outcome would be positive if and only if the disorder exists (com-
mental health preclinical assessment questions and demographic monly referred to as true positive results). Otherwise result will
features). The data is captured under the observation of a men- be negative (referred to as true negative values). Percentages are
tal health professional and is based on information obtained used to show specificity and sensitivity. The accuracy measure is
during the patient–psychiatrist interaction. Each response record used to assess AS_MH_RM’s performance. The AS_MH_RM makes
includes demographic information such as name, age, and gender, use of information from a knowledge database that was developed
as well as responses recorded as per validated scale [108]. It is in collaboration with mental health experts. The outputs of the
generally impracticable to evaluate the entire population during MH_ASD_PM are forwarded to the AS_MH_RM, which generates
the data gathering and analysis phase. As a result, sampling recommendations.
approaches are employed to collect without having to investi-
gate each one separately. The cost and workload of a study can
be lowered by limiting the number of population samples so 4.1. Findings and analysis
that high-quality data will become easier to collect. It is mostly
employed in situations when we expect the interest points vary The findings of the suggested system are summed up in this section.
between separated groups. After that, samples from each sub- The ‘‘accuracy’’, ‘‘kappa’’, and ‘‘root mean square(RMS)’’ metrics are
group are taken to create the study sample. The likelihood of used to evaluate the prediction performance algorithms on the acquired
an occurrence being incorporated in a stratified sample changes dataset. The efficacy of MH_ASD _PM is assessed, firstly taking feature
depending on known variables like gender, age, and study phase. selection technique into account and then without feature selection.
The proposed system classifies individuals into four classes based The findings of the MH _ASD_PM are first evaluated using a ROC curve
on the scores they exhibit as a result of their characteristics. The (in terms of true positive and false positive rates) and then by the
output set consists of minimal anxiety (MA), mild anxiety (MIA), validation of AS_MH_RM. To demonstrate the model’s usefulness, a test
moderate anxiety (MOD), severe anxiety(SA), and very severe case scenario is employed. Finally, in order to assess the proposed sys-
anxiety(VSA) which represent the various anxiety stage of indi- tem’s validity, the generated recommendations were compared with the
viduals. Moreover, this study has limited participants as mental labels provided by mental health professionals. The results were com-
health study depends on several factors, such as the research puted with the help of confusion matrix, by using subjective analysis
question, study design, statistical power, and available resources. of mental health professionals to validate the results.
Studies in conflict zones face additional challenges, such as se- • MH_ASD_PM Results
curity concerns, difficulty accessing participants, and ensuring The accessed dataset contains a variety of attributes, a few of which
participant safety. These factors can impact the feasibility and are worthwhile and play an important role in early diagnosis, while
sample size of a study. Some studies [109] in conflict zones some are merely raw data with no information. As, a result, information
have reported using smaller sample sizes due to the challenges gain based feature selection approach is taken into account to choose
and constraints of working in such environments. Overall, the the meaningful and relevant data besides eight remarkable features are
decision to have only 215 participants in a mental health study chosen for prediction in this work. The default values of the prediction
in conflict zones would depend on a range of factors, including method were applied to the full dataset, with and without feature
the research question, available resources, ethical considerations, selection and results were measured in terms of accuracy, kappa statis-
and statistical power. tics, and root mean square error (RMSE) as listed in Table 4, and the
• Experimental evaluation detailed performance metrics are depicted in Table 5. The experiments
The 215 Sample data units are split into two groups: 150 par- were performed on WEKA [110] which demonstrate that the random
ticipants (70 percent) for training and 64 subjects (30 percent) forest ensemble model performs well and achieves high accuracy of
for testing. To ensure the correctness of the system’s outputs, 98.14% with feature selection and 97.67% without feature selection
the system’s predictions and recommendations are compared to in comparison with other algorithms. During the training and testing
the real predictions and recommendations labelled by a medical phases Random Forest takes care of both missing data and outliers.
expert. Our suggested prediction model is evaluated using three The Random Forest(RF) classifier dominates other approaches due to its
performance metrics: ‘‘accuracy metric’’, ‘‘kappa statistics met- ensemble based classification methodology. The RF method eliminates
ric’’, and ‘‘root mean square error metric’’ (RMSE). Each class the excessive variance and bias concerns that decision trees are famous
is utilised to evaluate the data, and the prediction accuracy is for. It grows by making a huge number of classification decision trees;
analysed to see which classification method has been the most which results a new record in classifying a new class by storing it in a
promising. Correspondingly, the model’s performance is assessed vector and comparing it to every other tree. Each tree picks one of the
by plotting a receiver operating characteristic (ROC) curve or available classes, and the class with the highest votes becomes the final
sensitivity v/s specificity curve. The likelihood that a test record’s class for the record. Due to the complicated nature of random forest
11
Table 4
Two stage assessment for predicting anxiety stages.
Feature selection Prediction algorithms Evaluating metrics
(FS)
Kappa statistics Root Mean Square Accuracy (%)
(KS) Error (RMSE)
Support Vector 0.96 0.32 96.7442
Without feature
Machine (SVM)
selection
Multi-Layer Perceptron 0.89 0.18 92.093
(MLP)
Random Forest (RF) 0.96 0.15 97.6744
Support Vector 0.96 0.31 97.6744
With feature
Machine (SVM)
selection
Multi-Layer Perceptron 0.94 0.12 96.2791
(MLP)
Random Forest (RF) 0.97 0.11 98.1395
under investigation the classifier produced similar findings with regard Table 5
Polynomial and RBF kernels SVM performance evaluation.
to mean absolute error (MSE) and root mean square error (RMSE).
Additionally, SVM achieved the second highest accuracy of 97.68%. Kernels Filter type accuracy
Table 5 displays the accuracy results, which indicate that the poly- Normalisation Data standardisation
nomial kernel with data standardisation yields the best result, and Polynomial 97.2093 97.6744
Table 6 presents the detailed performance using different classifiers. RBF 35.814 94.8837
To determine the similarity between distinct data features, the poly- RBF: Radial Basis Function.
nomial kernel considers the combination of supplied features because
MLP delivers somewhat less accurate outcomes because it is not a
knowledge-based technique. However, when combined with feature (Where TPR=True Positive Rate; FNR= (False Negative Rate);
selection, it beats other ultra-modern prediction models. The kappa FPR=False Positive Rate; TNR=True Negative Rate; TP=True Positive;
statistics reveals that when kappa is near to 1 how effectively the clas- FP=False Positive; TN=True Negative; FN=False Negative).
sifiers performs when it comes to data allocation and data distributions
which are appropriate for each class. Table 4 depicts that RF yielded a 4.1.1. AS_MH_RM Results
kappa score approximately 0.97, that is high and very close to 1. Low The proposed AS_MH_RM model’s results summary and the findings
RMSE score, likewise, indicate good categorisation ability. Whenever are presented here in this section. The procedure followed is shown
these findings are contrasted, it is clear that proposed approach, when in Fig. 4 Step 1 displays the information received from mental health
used in conjunction with the feature selection technique, produces professionals and patient records. The stages are assigned weights based
outstanding prediction results. on their importance in the knowledge base. Numerical range is assigned
The prediction algorithm outcomes are improved and accurate dis- based on the severity of these parameters; with the white range being
order stage detection is achieved by assessing entropy values pro- ignored none of the individuals was found in this category. Step 2 dis-
duced by information gain evaluator. In order to rank the information plays the effectiveness of the risk assessment on the data. The formula
acquired, the ranking search algorithm (RSA) identifies the entropy in Eq. (6) is used to calculate the risk of the specified exposures in rela-
values of the feature sets. Findings of the entropy values are depicted tion to their corresponding classrooms. The results of a risk assessment
in Table 7 and are depicted in Fig. 6. Where a cut-off point is chosen on a data suggest that SA individuals are more vulnerable. Patients
at ≥0.1 and the values beneath cut-off are eliminated as they do limit with a higher overall severity score are more likely to be labelled
in providing useful information. as Depressive, which falls under the category of severe anxiety and
The Receiver Operating Characteristic (ROC) analysis of all the contributes to increased impairment and needs to act accordingly. Step
three algorithms for every anxiety stage which enables unbiased and 3 uses Eq. (8) to display the estimated frequency findings. Exposure
comprehensive assessment of classifier performance than simply util- towards a certain anxiety stage disorder can be analysed depending
ising accuracy. Fig. 5(a), 7(b), and 7(c) shows ROC curves of three upon severity levels. It has been found that individuals with a severity
classifiers for four different anxiety stages (MA, MIA, MOD, SA). The score of 14 are more inclined to be labelled with severe anxiety rather
ROC plots illustrate the prediction algorithm’s true positive and false than moderate anxiety.
positive rates, and the points on the curve reflect sensitivity and speci- Step 4 combines the aggregate results from the previous three
ficity of threshold values. Fig. 5(b) shows that the ROC plot curve drives steps to construct a severity scoring rule system. The probability cut-
through the left upper corner, meaning that the curve is near to 1, off is kept at 0.45, obtained by averaging the probability for the
indicating accurate performance. Additionally, the curves for other two two intervals. The final score is calculated using Eq. (9). A patient’s
classifiers, SVM, as depicted Fig. 5(a), and MLP, as depicted Fig. 5(c), obtained score is matched to the severity range levels indicated in
shows excellent performance. A ROC does not depend on the class Table 1 to recommend mental health treatment. Fig. 8. shows an
distribution which makes it best for predicting rare events such as example of AS_MH_RM scenario for 8 distinct patients. From step 1 of
mental health disorders. The ROC Analysis is carried out using the Fig. 7 overall severity range is detected and classified as red, yellow,
following ratios shown in Eqs. (11)–(14), green, blue and white. Moreover, none of the individuals was classified
𝑇𝑃 in Stage 5 i.e. very severe anxiety (VSA), however for scalability in
𝑇𝑃𝑅 = (11)
(𝑇 𝑃 + 𝐹 𝑁) future recommendations were depicted in Table 1. Each parameter’s
𝐹𝑁 relevance weight is mentioned next to it. Steps 2 and 3 calculate risk
𝐹 𝑁𝑅 = (12)
(𝑇 𝑃 + 𝐹 𝑁) and likelihood. Lastly, both severity and final scores were calculated to
𝐹𝑃 determine the recommendation class. Every record’s suggestion classes
𝐹𝑃𝑅 = (13)
(𝐹 𝑃 + 𝑇 𝑁) are displayed on the left side. A confusion matrix is generated to mea-
𝑇𝑁 sure the overall outcomes of the recommendations, as shown in Table 8.
𝑇 𝑁𝑅 = (14)
(𝐹 𝑃 + 𝑇 𝑁) The proposed system’s achieved an average accuracy of 96.75% and
12
Fig. 5. ROC Curves (a) Shows the ROC for all four stages of anxiety in case of SVM (b) Shows the ROC for all four stages of anxiety in case of RF (c) Shows the ROC for all
four stages of anxiety in case of MLP.
Abbreviations: - FPR False Positive Rate; TPR True Positive Rate; ROC Receiving Operating Characteristics, SVM Support Vector Machine, RF Random Forest, MLP Multi-Layer Perceptron.
Table 6
Detailed performance of algorithms under consideration using different classifiers with feature selection.
Classifier Precision (%) Recall (%) F-measure (%) ROC area (%) Class
SVM 0.932 1.000 0.965 0.983 MOD
1.000 1.000 1.000 1.000 MA
1.000 1.000 1.000 1.000 SA
1.000 0.907 0.951 0.969 MIA
0.978 0.977 0.977 0.987 Weighted average
MLP 0.941 0.941 0.941 0.987 MOD
1.000 1.000 1.000 1.000 MA
1.000 1.000 1.000 1.000 SA
0.926 0.926 0.926 0.979 MIA
RF 0.944 1.000 0.971 0.990 MOD
1.000 1.000 1.000 1.000 MA
1.000 1.000 1.000 1.000 SA
1.000 0.926 0.962 0.982 MIA
Abbreviations:- SVM: Support Vector Machine, RF: Random Forest, MLP: Multi-Layer Perceptron, MA: Minimal Anxiety, MIA: Mild Anxiety,
MOD: Moderate Anxiety, SA: Severe Anxiety.
4.2. Discussion
Table 7
Attribute Evaluator (supervised, Class (nominal): Using Information Gain Ranking Filter
(Entropy values for various features are utilised to pick features, and all values above The World Psychiatric Association (WPA) and the Lancet Psychiatry
0.1 are chosen). Commission(LPC) recently emphasised that a significant technological
Entropy values Features shift in mental healthcare is required, and they propose Digital Psy-
0.7542 Trouble relaxing chiatry as one possible answer [111]. In this situation, prediction and
0.7058 Being so restless that it is hard to sit still
recommender systems should be adapted to reduce these disparities
0.5545 Becoming easily annoyed or irritable
0.5027 Worrying too much about different things and enable both individuals and mental health professionals in mak-
0.4241 Feeling nervous, anxious, or on edge ing mental healthcare decisions better and affordable. It also enables
0.4057 Feeling afraid, as if something awful might happen decision makers to spotlight on individuals who are most likely to
0.3473 Not being able to stop or control worrying
benefit from first-line treatment while allocating other resources to
0.082 Age
0 Phase individuals who will require second-line or additional treatments. A
0 Field of study better patient experience will lead to improved health and motivation
0 Location to live a healthier lifestyle with pre-clinical prediction and recom-
0 Income
mendation systems. Additionally, they aid healthcare practitioners in
0 Marital status
illness prediction and to bridge the mental health treatment gap [40–
42]. In this research, a new adaptive prediction and recommendation
models have been proposed for diagnosing and treating various stages
is determined by reviewing the outcomes of the actually suggested of anxiety based on severity score. Further, recent advancements in the
classes, as assessed by the mental health expert, with the AS_MH_RM field of machine learning for mental health can help us to better under-
recommendations. stand individuals in terms of behaviour, surroundings, panic, feelings,
13
Fig. 6. Visualisation of different Entropy values for feature selection, values ≥0.1 is selected. (Note: -Q1–Q4 abbreviations represents the features depicted in Table 7. Here 𝑋-axis
represents the number of features and 𝑌 -axis represents the feature ranking i.e entropy values).
Table 8
The cluster based confusion matrix for AS_MH_RM.
Stages Prediction labels Stage Stage Stage Stage Stage
1: Level 1 2: Level 3: Level 3 4: Level 4 5: Level 5
Stage 1: Minimal Level 1: E-health based anxiety management 26 1 0 1 0
anxiety
Stage 2: Mild Level 2: Level 1 + Exposure therapy with active general 1 5 2 0 0
anxiety
Stage 3: Moderate Level 3: Level 2 + Pharmaco therapy with active general 1 20 4 0
anxiety
Stage 4: Severe Level 4: Level 3 + multidisciplinary care or multiagency 1 1 1 10 0
anxiety support model involving active general practitioner
management, specialist physicians, allied health
professionals, and possible case manager
Stage 5: Very severe Level 5: Level 4 and further to be integrated within 0 0 0 0 0
anxiety higher-tier mental health services
Accuracy 98% 97% 97% 95% 0%
anger, happiness, uncertainty [112]. With smart technological devices and most of the works lack adequate psychiatric expert supervision
and social networking in place the amount of health related data has and guidelines. To reduce potential hazards and enhance benefits for
increased dramatically in recent years, posing certain challenges for users, the idea of ‘‘first do no harm’’ must be kept in mind when
both health care providers and for patients as well. creating a recommendation system. When it comes to creating a new
Additionally, individuals are most concerned about privacy and are product, the health of the individuals is the most important factor to
looking forward to locate best mental providers, in order to resolve consider recommendations, even if they go against the wishes of the
their mental health problems without any shame and stigma, which fur- patients [120].
ther presents the issue of individual-mental health professional match- The proposed model is validated using a data set of size 215
making, in which individuals might find the ideal psychiatrists with individuals to develop a reliable and rapid mental health prediction
whom they can form a trusting connection [113–115]. However, it is
and recommendation methodology. The sample dataset size used is
time consuming and is difficult to adopt in India particularly in Kashmir
much better than the earlier studies utilised to classify the various
due to a high level of commitment on the part of both the therapist
mental health illnesses were typically modest, with fewer than 100
and the patient, lengthy sessions at exorbitant prices and due to huge
sample [98–100,100–107]. However, few studies included more than
shortage of mental health professionals [116].
100 samples to predict the presence of mental health disorders [121–
Table 8 reflects an important Comparative analysis of this work
127], and have focused on different stages of mental health disor-
with other existing works in classifying and predicting the depression
and anxiety cases with machine learning. Most of the research articles ders [23,69,70], while the proposed model is a multi-label prediction
show that machine learning models have obtained the accuracy of and classification of five well known anxiety stage levels i.e. MA, MIA,
above 70%. Meanwhile, the convolutional neural network has obtained MD, SA and VSA, and we achieved the highest level of accuracy of
excellent performance with an accuracy of 96.0% for anxiety classi- 98.13% which is greater than other similar models and is fair and
fication and 96.8% for the depression classification as stated in the promising as shown in Table 9. The research done for this study
article by Ahmed et al. [55]. Besides, the random forest and support substantially filled in the gaps left by the earlier efforts. The suggested
vector machine perform very well in classifying the depression and technique is adopted in accordance with mental health processes which
anxiety cases as stated in the research articles by Sau and Bhakta [117], makes use of psychiatric expert supervision and professional experience
Katsis et al. [109], Sau and Bhakta [118], and Hilbert et al. [119]. that, is more precise, reliable and trustworthy. Although the approach
Additionally, related works have worked upon open source datasets has a restriction in terms of its data set size and heavy reliance on
14
Fig. 7. A full illustration of the steps for (AS_MH _RM). In collaboration with mental health experts, (Part A) A features important weight is assigned, (Part B) an exposure analysis
is conducted, (Part C) a severity analysis is conducted, and (Part D) a severity level and score is assigned, all of which are utilised to develop recommendations.
Fig. 8. Different AS_MH_RM test case situations with anticipated anxiety stage disorder and recommendation (For recommendation please refer (Table 3)).
mental health experts as clinical validation is critical for any decision the quality of the study. The majority of earlier research has focused
support system’s efficacy. on predicting depression in persons who fall into a particular age
Further, it is vital to contrast the findings and contribution of category, profession, or state of health. A select handful of them have
this study with those of other published works in order to assess identified the most important socio-demographic and psychological
15
Table 9
Important comparative analysis of this work with other existing relevant works.
S. no Authors Year Mental health issue Data size Classifiers Overall accuracy
1. Katsis 2011 Anxiety Not specified RF, SVM, 84.3%
et al. [109]
2. Sau and 2019 Depression and anxiety 470 LR, SVM,NB,RF 89.3%
Bhakta, [117]
3. Marmar 2019 PTSD 129 RF 89.1%
et al. [121]
4. Hilbert 2017 Anxiety Not specified SVM 90.10%
et al. [119]
5. Ahmed et 2019 Depression and anxiety Not specified CN,SVM , LDA, KNN 96.0% for anxiety and
al., [55] 96.8% for depression
6. Jerry et al. [33] 2019 Depression Not specified LR, NN, RF, SVM, F1 score 0.73
XGBoost, K-nearest
neighbours
7. This work 2022 Pre-clinical Anxiety 215 SVM.RF.MLP 98.13%
Stage Prediction and
Recommendation
Abbreviations:- RF: random forest; SVM: support vector machine; MLP:multi-layer Perceptron; LR: logistic regression; NB: naive bayes; NN: neural network; KNN: K-nearest neighbour,
LDA: linear discriminant analysis.
elements that contribute to anxiety and depression. But the goal of this In this work, a prediction model that divides anxiety into pre-
study was to forecast the stages of anxiety in persons from a variety of clinical stages was created for the first time using data from Kash-
socioeconomic, cultural, and professional backgrounds and age ranges. mir(India). The findings of the proposed system demonstrate high
Additionally, the most important sociodemographic characteristics that prediction and recommendation accuracy, implying that the system
contribute to anxiety have been found by this study. Mental health will be effective in predicting various pre-clinical anxiety stages based
facilities, funding, and capability issues lead to reduced access and on severity level. The proposed research shows the development of a
difficulty in accessing appropriate care throughout many conflict set- smart, adaptable prediction and recommendation model for individuals
tings. Moreover, given that the Diagnostic and Statistical Manual of who might be suffering from Anxiety stage disorders. By measuring the
Mental Disorders, Fifth Edition diagnostic criteria are occasionally used risk of disorder and predicting the probability of disorder occurrence,
differently depending on the patient’s age, there might be biases based the model predicts the presence of various anxiety stage disorder in
on age which limits its applicability to both infants and adults [6] individuals and suggests the psychiatric help. The criticality of variables
Furthermore, these issues have an impact on the organisation and based under consideration as defined by the mental health expert,
planning of mental health services, resulting in more gaps in referral are used to determine risk and probability estimates. By developing a
systems, a decrease in evidence-based treatment standards, and in- hybrid prediction and recommendation model, this proposed work con-
tegrated offers with public administration and other vital sectors in tributes to mental health informatics by assisting mental health experts
policy and operational plans [8,128–133]. Moreover, robust mental in making timely diagnostic decisions. The empirical results attested to
health information systems could support effective treatment research the fact that RF outperformed the SVM with 98.13% accuracy against
capabilities and emphasise data quality. From the literature, it is clear 97.6744% for the multi class problem when predicting anxiety stages
that AI technology risks widening the gap between conflict and non- for earlier diagnosis and timely intervention using ten features after
conflict countries by shifting more focus and investment on research
feature selection. Therefore, the researchers recommend that decision
for mental healthcare in already established High Income countries
maker in Kashmir(India) adopt the prediction model produced by this
(HICs). To prevent this growing disparity, researchers and policymakers
study to strategically plan the distribution of both preventative and
in conflict economies will need to take action to raise mental healthcare
curative mental health care services. Moreover, the proposed method-
service standards. Moreover, a significant problem is the scarcity of re-
ology is not supposed to replace Clinical diagnostic made by a mental
search studies using AI/ML in conflict settings and the lack of data sets.
health expert; alternatively, it is supposed to enhance it. As a result,
However, the use of small samples is widespread in the field of mental
the automated anxiety clinical stage classification technique given here
health Vabalas et al. [134], because it is expensive to acquire data from
should not (and is not intended to) be used in any way as the primary
human subjects, and the guidelines for conducting experiments under
measure for individual’s mental health.
various settings are still being developed.
Although it is doubtful that artificial intelligence will overtake psy-
5. Conclusion chiatrists in the coming years, mental health practitioners must become
familiar with the basics of AI technology and how AI-based solutions
An efficient implementation of AI/ML in mental health care par- might support them with their daily job to improve patient outcomes.
ticularly in conflict zones can significantly impact service quality, Current estimates warrant more investment in preventing and treat-
especially in conflict settings. However, any new technology in health ing mental disorders in conflict-affected populations. Future research
care will inevitably face obstacles, but an emergent strategy incorpo- should focus on adapting AI to improve diagnostic outcomes using mul-
rating user feedback can help make the technology more economical, timodal data from conflict zones. The fascinating challenge is to help
accessible, and efficient. A significant and under-recognised component combine these diversified and possibly multidimensional approaches
of global health is mental health issues. AI in mental healthcare has a into a unified analytical technique that considers a large amount of
bright future. We must actively participate in guiding the implemen- high-dimensional data and accurately captures any underlying cross
tation of AI in mental health care particularly in conflict settings. As variation. Using AI and associated technologies in conflict-affected
researchers, and mental health professionals are committed to improv- populations to improve mental health will highlight necessary evidence
ing mental healthcare, we have contributed our expertise of AI and and bring efficiency and effectiveness to mental healthcare due to data-
collaborated with data scientists, computer scientists, and other experts driven outcomes. One such most significant impact will be the ability of
to transform the field of mental health in conflict settings so as to early detection and prediction of various internal and external conflict-
improve mental health patient care. Currently, AI for mental health is related health determinants, which might be responsible for chronic
only used in HICs; this should be expanded in conflict situations. mental health disorders among people living in conflict zones.
16
A drawback of this research is that firstly, it relied on pre-clinical • Cultural sensitivity: Machine learning algorithms may not be able
data as the gold standard for diagnosing anxiety stages. The dataset to accurately capture cultural differences and nuances in mental
did not contain any biological markers that may be used to forecast health. It is essential to develop algorithms that are culturally
various anxiety stages. In order to anticipate anxiety stages in an sensitive to ensure that mental health interventions are effective
individual, certain biological characteristics are important. The model and culturally appropriate.
may more accurately predict anxiety stages if these biological factors • Training and capacity building: The use of machine learning
were included as different biological characteristics of the participants in mental health care requires specialised skills and expertise.
can be incorporated in the latter study because diverse biological Training and capacity building programs must be developed to
factors have a notable impact on how anxiety develops in people. ensure that mental health professionals are equipped with the
Numerous studies show that applying various dimensionality reduction necessary skills to use machine learning effectively.
algorithms throughout the data pre-treatment processes might improve
performance. Many researchers are of the view that there may be Overall, the use of ML in mental healthcare in conflict zones has
response bias present in the pre-clinical self-reported data and should the potential to improve access care and early interventions. To fully
not be taken alone into consideration. Secondly, due to the high cost of fulfil the promise of AI, a diverse community of professionals in-
data collection that necessitates human engagement, small sample size terested in psychiatric treatment and research, including scientists,
has been used which are however common and many ML models may clinicians, politicians, and consumers, need to interact and collaborate.
show resilience when trained on a small sample size of data without Further, we pitch for large-scale field studies to substantiate the ef-
affecting performance accuracy. Thirdly, one of the largest obstacles is fectiveness of AI for mental health in conflict zones like Iraq, Syria,
the accurate diagnosis of mental health disorders is that they constantly Kashmir, Yemen, and Afghanistan, among others. The broader scien-
fluctuate overtime. Researchers can investigate creating efficient mod- tific community and international organisations are urged to endorse
els that recognise various symptom intensities and take into account technology-based mental health in conflict zones. Attempting to learn
the various scenarios in these disorders that alter over time. Fourthly, how to better navigate mental health in locations with extreme adverse
there is lack of transparency as ML algorithms are often viewed as situations and emergencies can benefit other populations facing similar
"black boxes’’, meaning that it can be challenging to understand how challenges.
the algorithm arrived at its recommendations or predictions. This lack
of transparency can lead to a lack of trust in the algorithm’s results, Declaration of competing interest
especially when it comes to sensitive mental health issues. Fifthly, due
to privacy and security concerns mental health data is highly sensitive The authors declare that they have no known competing finan-
and should be protected to ensure patient privacy. Machine learning cial interests or personal relationships that could have appeared to
algorithms must be designed to protect the confidentiality and security influence the work reported in this paper.
of patient data. Last but not the least, mental health sector faces a Ethical aspects
huge issue in accessing relevant, top-notch, large-scale data. This is No formal ethical approval was necessary as required by the decla-
because of worries about subject recruitment, costs, and the nature of ration of Helsinki (Principal 1)6 . The study participants were handled
data gathering, which typically calls for multidisciplinary collaboration as per APA/ICMR ethical guidelines.
with healthcare professionals. In order to increase user confidence and
informed consent, further steps may be done before obtaining data from Data availability
individuals. Building anonymous, trustworthy repositories for mental
health data where individuals can freely share details about their men- Data will be made available on request.
tal health conditions for research purposes may also boost participant
confidence. To ensure improved mental healthcare in conflict settings References
policy makers must bear in mind the following points in terms of health
policy implications: [1] N. Votruba, G. Thornicroft, Global mental health, 2022, http://dx.doi.org/10.
1017/gmh.2016.20.
• Access to mental health care: Conflict zones often lack the re- [2] F. Bertini, D. Allevi, G. Lutero, D. Montesi, L. Calzà, Automatic speech classifier
sources and infrastructure to provide adequate mental health for mild cognitive impairment and early dementia, ACM Trans. Comput.
Healthc. 3 (1) (2022) 1–11, http://dx.doi.org/10.1145/3469089.
care. Machine learning can help identify individuals who are at
[3] R.C. Kessler, et al., Individual and societal effects of mental disorders on
high risk of developing mental health problems and provide early earnings in the United States: Results from the national comorbidity survey
interventions, even in areas where there are no mental health replication, Am. J. Psychiatry 165 (6) (2008) 703–711, http://dx.doi.org/10.
professionals available. 1176/appi.ajp.2008.08010126.
• Increased Funding: To address the limitations of machine learning [4] H.A. Whiteford, et al., Lobal burden of disease attributable to mental and
substance use disorders: Findings from the global burden of disease study
in mental health, policymakers should consider increasing fund-
2010, Lancet 382 (9904) (2013) 1575–1586, http://dx.doi.org/10.1016/S0140-
ing for mental health research and data collection especially in 6736(13)61611-6.
conflict settings. [5] C.W. Woo, L.J. Chang, M.A. Lindquist, T.D. Wager, Building better biomarkers:
• Standards and Regulations: Policymakers should develop stan- Brain models in translational neuroimaging, Nature Neurosci. 20 (3) (2017)
365–377, http://dx.doi.org/10.1038/nn.4478.
dards and regulations for the use of machine learning in mental
[6] F. Ma, in: D. Gu, M.E. Dupre (Eds.), Diagnostic and Statistical Manual of Mental
health to ensure that algorithms are transparent, unbiased, and Disorders-5 (DSM-5) BT - Encyclopedia of Gerontology and Population Aging,
protect patient privacy and security. Springer International Publishing, Cham, 2020, pp. 1–12.
• Education and Training: Healthcare professionals should be edu- [7] I.B. Hickie, et al., Applying clinical staging to young people who present
cated and trained on the use of machine learning algorithms in for mental health care, Early Interv. Psychiatry 7 (1) (2013) 31–43, http:
//dx.doi.org/10.1111/j.1751-7893.2012.00366.x.
mental health. This training should include information on the [8] T. Martineau, et al., Leaving no one behind: Lessons on rebuilding health
limitations and ethical considerations associated with the use of systems in conflict- and crisis-affected states, BMJ Glob. Heal. 2 (2) (2017)
machine learning. 1–6, http://dx.doi.org/10.1136/bmjgh-2017-000327.
• Collaboration and Interdisciplinary Approaches: Policymakers
should encourage collaboration between mental health profes- 6
https://www.wma.net/policies-post/wma-declaration-
sionals, data scientists, and other stakeholders to develop more of-helsinki-ethical-principles-for-medical-research-involving-
effective machine learning algorithms for mental health that human-subjects/#:∼:text=It%20is%20the%20duty%20of,
address the limitations mentioned above. the%20fulfilment%20of%20this%20duty.
17
[9] S. Garry, F. Checchi, Armed conflict and public health: Into the 21st century, [33] C. Mario, B. Igor, Rewards in public administration: A proposed classification,
J. Public Heal. (United Kingdom) 42 (3) (2020) E287–E298, http://dx.doi.org/ J. Soc. Adm. Sci. 5 (2) (2018) 68–80, http://dx.doi.org/10.1453/jsas.v5i2.1648.
10.1093/pubmed/fdz095. [34] F. Thabtah, Machine learning in autistic spectrum disorder behavioral research:
[10] D. Duclos, A. Ekzayez, F. Ghaddar, F. Checchi, K. Blanchet, Localisation and A review and ways forward, Inform. Heal. Soc. Care 44 (3) (2019) 278–297,
cross-border assistance to deliver humanitarian health services in North-West http://dx.doi.org/10.1080/17538157.2017.1399132.
Syria: A qualitative inquiry for the Lancet-AUB Commission on Syria, Confl. [35] I. Kadi, A. Idri, J.L. Fernandez-Aleman, Knowledge discovery in cardiology:
Health 13 (1) (2019) 1–10, http://dx.doi.org/10.1186/s13031-019-0207-z. A systematic literature review, Int. J. Med. Inform. 97 (2017) 12–32, http:
[11] L. Connah, Double lockdown in Kashmir during the Covid-19 pandemic, Peace //dx.doi.org/10.1016/j.ijmedinf.2016.09.005.
Rev. 33 (1) (2021) 33–38, http://dx.doi.org/10.1080/10402659.2021.1956128. [36] C.H. Lee, H.J. Yoon, Medical big data: Promise and challenges, Kidney Res. Clin.
[12] I. Georgieva, et al., Prevalence, new incidence, course, and risk factors of PTSD, Pract. 36 (1) (2017) 3–11, http://dx.doi.org/10.23876/j.krcp.2017.36.1.3.
depression, anxiety, and panic disorder during the Covid-19 pandemic in 11 [37] E. Vayena, A. Blasimme, I.G. Cohen, Machine learning in medicine: Addressing
countries, Healthc. (Basel, Switzerland) 9 (6) (2021) http://dx.doi.org/10.3390/ ethical challenges, PLoS Med. 15 (11) (2018) 21–23, http://dx.doi.org/10.
healthcare9060664. 1371/journal.pmed.1002689.
[13] M. Coccia, Sources and diffusion and prediction in COVID-19 pandemic : lessons [38] H. Das, B. Naik, H.S. Behera, Medical disease analysis using neuro-fuzzy with
learned to face next health emergency, 10, (February) 2023, pp. 145–168, feature extraction model for classification, Inform. Med. Unlocked 18 (2020)
http://dx.doi.org/10.3934/publichealth.2023012. 100288, http://dx.doi.org/10.1016/j.imu.2019.100288.
[14] A. Ray, A. Bhardwaj, Y.K. Malik, S. Singh, R. Gupta, Artificial intelligence [39] N.C. Caballé, J.L. Castillo-Sequera, J.A. Gómez-Pulido, J.M. Gómez-Pulido,
and psychiatry: An overview, Asian J. Psychiatry 70 (2022) 103021, http: M.L. Polo-Luque, Machine learning applied to diagnosis of human diseases: A
//dx.doi.org/10.1016/j.ajp.2022.103021. systematic review, Appl. Sci. 10 (15) (2020) 1–27, http://dx.doi.org/10.3390/
[15] G. Bowsher, et al., eHealth for service delivery in conflict : a narrative review app10155135.
of the application of eHealth technologies in contemporary conflict settings, [40] J. Pincay, L. Teran, E. Portmann, Health recommender systems: A state-of-the-
(April) 2021, pp. 974–981, http://dx.doi.org/10.1093/heapol/czab042. art review, in: 2019 6th Int Conf. EDemocracy EGovernment, ICEDEG 2019,
[16] F. Jiang, et al., Artificial intelligence in healthcare: Past, present and future, 2019, pp. 47–55, http://dx.doi.org/10.1109/ICEDEG.2019.8734362.
Stroke Vasc. Neurol. 2 (4) (2017) 230–243, http://dx.doi.org/10.1136/svn- [41] H. Schäfer, et al., Towards health (aware) recommender systems, ACM Int.
2017-000101. Conf. Proceeding Ser. Part F 1286 (2017) 157–161, http://dx.doi.org/10.1145/
[17] D.D. Miller, C.M. Facp, E.W. Brown, Artificial intelligence in medical practice 3079452.3079499.
: The question to the answer? Am. J. Med. 131 (2) (2018) 129–133, http: [42] A.K. Sahoo, C. Pradhan, R.K. Barik, H. Dubey, DeepReco: Deep learning based
//dx.doi.org/10.1016/j.amjmed.2017.10.035. health recommender system using collaborative filtering, Computation 7 (2)
(2019) http://dx.doi.org/10.3390/computation7020025.
[18] G.O. Gabbard, H. Crisp-han, The early career psychiatrist and the psychother-
[43] B. Rekabdar, D.L. Albright, J.T. McDaniel, S. Talafha, H. Jeong, From machine
apeutic identity, Acad. Psychiatry (2017) 30–34, http://dx.doi.org/10.1007/
learning to deep learning: A comprehensive study of alcohol and drug use
s40596-016-0627-7.
disorder, Healthc. Anal. 2 (2022) 100104, http://dx.doi.org/10.1016/j.health.
[19] D. Bzdok, A. Meyer-Lindenberg, Machine learning for precision psychiatry:
2022.100104.
Opportunities and challenges, Biol. Psychiatry Cogn. Neurosci. Neuroimaging
[44] A. Thompson, C. Issakidis, C. Hunt, Delay to seek treatment for anxiety and
3 (3) (2018) 223–230, http://dx.doi.org/10.1016/j.bpsc.2017.11.007.
mood disorders in an Australian clinical sample, Behav. Chang. 25 (2) (2008)
[20] A.B.R. Shatte, D.M. Hutchinson, S.J. Teague, Machine learning in mental health:
71–84, http://dx.doi.org/10.1375/bech.25.2.71.
a scoping review of methods and applications, Psychol. Med. 49 (9) (2019)
[45] M. Attaran, B. Gokhan, Digital Twin : Benefits, use cases, challenges, and
1426–1448, http://dx.doi.org/10.1017/S0033291719000151.
opportunities, Decis. Anal. J. 6 (2022) 100165, http://dx.doi.org/10.1016/j.
[21] A. Goodman, R. Joyce, J.P. Smith, The long shadow cast by childhood physical
dajour.2023.100165.
and mental problems on adult life, Proc. Natl. Acad. Sci. USA 108 (15) (2011)
[46] R. Garriga, et al., Crises from electronic health records, 28, (June) 2022,
6032–6037, http://dx.doi.org/10.1073/pnas.1016970108.
http://dx.doi.org/10.1038/s41591-022-01811-5.
[22] A. Coyle, C. Thieme, M. Linehan, J. Balaam, D. Wallace, S. Lindley, Emotional
[47] R. Sawhney, A. Malik, S. Sharma, V. Narayan, A comparative assessment of
wellbeing, Int. J. Hum. Comput. Stud. 72 (8–9) (2014) 627–628, http://dx.doi.
artificial intelligence models used for early prediction and evaluation of chronic
org/10.1016/j.ijhcs.2014.05.008.
kidney disease, Decis. Anal. J. 6 (2023) 100169, http://dx.doi.org/10.1016/j.
[23] I. Pavlova, D. Zikrach, D. Mosler, D. Ortenburger, T. Gora, J. Wasik, De-
dajour.2023.100169.
terminants of anxiety levels among young males in a threat of experiencing
[48] D. Brathwaite, et al., Age and sex trends among mental health-related emer-
military conflict-Applying a machine-learning algorithm in a psychosociological
gency department visits in North Carolina, Healthc. Anal. 2 (2022) 100056,
study, PLoS One 15 (10) (2020) 1–24, http://dx.doi.org/10.1371/journal.pone.
http://dx.doi.org/10.1016/j.health.2022.100056.
0239749.
[49] J. Cvetković, Breast cancer patients’ depression prediction by machine learning
[24] S. Naveed, et al., Prevalence of common mental disorders in south Asia: A sys- approach, Cancer Invest. 35 (8) (2017) 569–572, http://dx.doi.org/10.1080/
tematic review and meta-regression analysis, Front. Psychiatry 11 (September) 07357907.2017.1363892.
(2020) 1–8, http://dx.doi.org/10.3389/fpsyt.2020.573150. [50] I.A. Ahmed, et al., Eye tracking-based diagnosis and early detection of autism
[25] M.M. Hossain, et al., Prevalence of anxiety and depression in south Asia spectrum disorder using machine learning and deep learning techniques,
during COVID-19: A systematic review and meta-analysis, Heliyon 7 (4) (2021) Electron 11 (4) (2022) http://dx.doi.org/10.3390/electronics11040530.
e06677, http://dx.doi.org/10.1016/j.heliyon.2021.e06677. [51] M. Mehedi Hassan, S. Mollick, F. Yasmin, An unsupervised cluster-based feature
[26] B. Stark, C. Knahl, M. Aydin, K. Elish, A literature review on medicine grouping model for early diabetes detection, Healthc. Anal. 2 (2022) 100112,
recommender systems, Int. J. Adv. Comput. Sci. Appl. 10 (8) (2019) 6–13, http://dx.doi.org/10.1016/j.health.2022.100112.
http://dx.doi.org/10.14569/ijacsa.2019.0100802. [52] F. Charlson, M. van Ommeren, A. Flaxman, J. Cornett, H. Whiteford, S. Saxena,
[27] M. Wiesner, D. Pfeifer, Health recommender systems: Concepts, requirements, New WHO prevalence estimates of mental disorders in conflict settings: a
technical basics and challenges, Int. J. Environ. Res. Public Health 11 (3) (2014) systematic review and meta-analysis, Lancet 394 (10194) (2019) 240–248,
2580–2607, http://dx.doi.org/10.3390/ijerph110302580. http://dx.doi.org/10.1016/S0140-6736(19)30934-1.
[28] G. Orrù, M. Monaro, C. Conversano, A. Gemignani, G. Sartori, Machine learning [53] A. Moore, J.A.F. van Loenhout, M.M. de Almeida, P. Smith, D. Guha-Sapir,
in psychometrics and psychological research, Front. Psychol. 10 (January) Measuring mental health burden in humanitarian settings: a critical review of
(2020) 1–10, http://dx.doi.org/10.3389/fpsyg.2019.02970. assessment tools, Glob. Health Action 13 (1) (2020) http://dx.doi.org/10.1080/
[29] S. Srivastava, N. Purkayastha, H. Chaurasia, T. Muhammad, Socioeconomic in- 16549716.2020.1783957.
equality in psychological distress among older adults in India: a decomposition [54] A.M. Chekroud, et al., Cross-trial prediction of treatment outcome in depression:
analysis, BMC Psychiatry 21 (1) (2021) 179, http://dx.doi.org/10.1186/s12888- a machine learning approach., Lancet Psychiatry 3 (3) (2016) 243–250, http:
021-03192-4. //dx.doi.org/10.1016/S2215-0366(15)00471-X.
[30] S. León-Giraldo, et al., Health in conflict zones: Analyzing inequalities in mental [55] A. Ahmed, R. Sultana, M.T.R. Ullas, M. Begom, M.M.I. Rahi, M.A. Alam, A
health in Colombian conflict-affected territories, Int. J. Public Health 66 (2021) machine learning approach to detect depression and anxiety using supervised
http://dx.doi.org/10.3389/ijph.2021.595311. learning, in: 2020 IEEE Asia-Pacific Conference on Computer Science and
[31] M. Coccia, Comparative critical decisions in management, in: A. Farazmand Data Engineering, CSDE, 2020, pp. 1–6, http://dx.doi.org/10.1109/CSDE50874.
(Ed.), Global Encyclopedia of Public Administration, Public Policy, and Gover- 2020.9411642.
nance, Springer International Publishing, Cham, 2020, pp. 1–10, http://dx.doi. [56] P.M. Doraiswamy, C. Blease, K. Bodner, Artificial intelligence and the future
org/10.1007/978-3-319-31816-5_3969-1. of psychiatry: Insights from a global physician survey, Artif. Intell. Med. 2019
[32] M. Coccia, I. Benati, Comparative models of inquiry, in: A. Farazmand (Ed.), (2020) 101753, http://dx.doi.org/10.1016/j.artmed.2019.101753.
Global Encyclopedia of Public Administration, Public Policy, and Governance, [57] S.M. Ganie, M.B. Malik, An ensemble machine learning approach for predicting
Springer International Publishing, Cham, 2018, pp. 1–6, http://dx.doi.org/10. Type-II diabetes mellitus based on lifestyle indicators, Healthc. Anal. 2 (2022)
1007/978-3-319-31816-5_1199-1. 100092, http://dx.doi.org/10.1016/j.health.2022.100092.
18
[58] I.M. Talha, I. Salehin, S.C. Debnath, M. Saifuzzaman, M.N.N. Moon, F.N. Nur, [80] S. Lefkovits, L. Lefkovits, Gabor feature selection based on information gain,
Human behaviour impact to use of smartphones with the python implementa- Procedia Eng. 181 (2017) 892–898, http://dx.doi.org/10.1016/j.proeng.2017.
tion using naive Bayesian, in: 2020 11th Int. Conf. Comput. Commun. Netw. 02.482.
Technol. ICCCNT 2020, 2020, http://dx.doi.org/10.1109/ICCCNT49239.2020. [81] M. Vartak, S. Huang, T. Siddiqui, S. Madden, A. Parameswaran, Towards
9225620. visualization recommendation systems, SIGMOD Rec. 45 (4) (2016) 34–39,
[59] A. Abd-Alrazaq, et al., The performance of artificial intelligence-driven tech- http://dx.doi.org/10.1145/3092931.3092937.
nologies in diagnosing mental disorders: an umbrella review, NPJ Digit. Med. [82] E.G. Pintelas, T. Kotsilieris, I.E. Livieris, P. Pintelas, A review of machine
5 (1) (2022) 87, http://dx.doi.org/10.1038/s41746-022-00631-8. learning prediction methods for anxiety disorders, ACM Int. Conf. Proceeding
[60] A. Priya, S. Garg, N.P. Tigga, Predicting anxiety, depression and stress in Ser. (2018) 8–15, http://dx.doi.org/10.1145/3218585.3218587.
modern life using machine learning algorithms, Procedia Comput. Sci. 167 [83] E. Alshdaifat, D. Alshdaifat, A. Alsarhan, F. Hussein, S.M.F.S. El-Salhi, The
(2019) (2020) 1258–1267, http://dx.doi.org/10.1016/j.procs.2020.03.442. effect of preprocessing techniques, applied to numeric features, on classification
[61] S. Kendale, P. Kulkarni, A.D. Rosenberg, J. Wang, Supervised machine-learning algorithms’ performance, Data 6 (2) (2021) 1–23, http://dx.doi.org/10.3390/
predictive analytics for prediction of postinduction hypotension, Anesthesiology data6020011.
129 (4) (2018) 675–688, http://dx.doi.org/10.1097/ALN.0000000000002374. [84] C.H. Liu, C.F. Tsai, K.L. Sue, M.W. Huang, The feature selection effect on
[62] S. Chancellor, M. De Choudhury, Methods in predictive techniques for mental missing value imputation of medical datasets, Appl. Sci. 10 (7) (2020) 1–12,
health status on social media: a critical review, Npj Digit. Med. 3 (1) (2020) http://dx.doi.org/10.3390/app10072344.
[85] E.R. Ziegel, The elements of statistical learning, Technometrics 45 (3) (2003)
http://dx.doi.org/10.1038/s41746-020-0233-7.
267–268, http://dx.doi.org/10.1198/tech.2003.s770.
[63] A.M. Darcy, A.K. Louie, L.W. Roberts, Machine learning and the profession
[86] P. Indyk, R. Motwani, Approximate nearest neighbors: Towards removing the
of medicine, JAMA - J. Am. Med. Assoc. 315 (6) (2016) 551–552, http:
curse of dimensionality, Conf. Proc. Annu. ACM Symp. Theory Comput. 8
//dx.doi.org/10.1001/jama.2015.18421.
(1998) 604–613, http://dx.doi.org/10.4086/toc.2012.v008a014.
[64] S. Rennick-Egglestone, S. Knowles, G. Toms, P. Bee, K. Lovell, P. Bower,
[87] R. Zebari, A. Abdulazeez, D. Zeebaree, D. Zebari, J. Saeed, A comprehensive
Health technologies ‘in the wild’: Experiences of engagement with computerised
review of dimensionality reduction techniques for feature selection and feature
CBT, in: Conf. Hum. Factors Comput. Syst. - Proc., 2016, pp. 2124–2135,
extraction, J. Appl. Sci. Technol. Trends 1 (2) (2020) 56–70, http://dx.doi.org/
http://dx.doi.org/10.1145/2858036.2858128.
10.38094/jastt1224.
[65] A. Thieme, et al., Challenges for designing new technology for health and [88] R.B. Pereira, A. Plastino, B. Zadrozny, L.H.C. Merschmann, Information gain
wellbeing in a complex mental healthcare context, in: Conf. Hum. Factors Com- feature selection for multi-label classification, J. Inf. Data Manag. 6 (1) (2015)
put. Syst. - Proc., 2016, pp. 2136–2149, http://dx.doi.org/10.1145/2858036. 48–58.
2858182. [89] A. Clare, R.D. King, Knowledge discovery in multi-label phenotype data, Lect.
[66] T.N.T. Tran, A. Felfernig, C. Trattner, A. Holzinger, Recommender systems in Notes Comput. Sci. (Including Subser. Lect. Notes Artif. Intell. Lect. Notes
the healthcare domain: state-of-the-art and research issues, J. Intell. Inf. Syst. Bioinformatics) 2168 (2001) 42–53, http://dx.doi.org/10.1007/3-540-44794-
57 (1) (2021) 171–201, http://dx.doi.org/10.1007/s10844-020-00633-6. 6_4.
[67] A. Othmani, A.O. Zeghina, A multimodal computer-aided diagnostic system [90] S. Huang, C.A.I. Nianguang, P. Penzuti Pacheco, S. Narandes, Y. Wang, X.U.
for depression relapse prediction using audiovisual cues: A proof of con- Wayne, Applications of support vector machine (SVM) learning in cancer
cept, Healthc. Anal. 2 (2022) 100090, http://dx.doi.org/10.1016/j.health.2022. genomics, Cancer Genom. Proteom. 15 (1) (2018) 41–51, http://dx.doi.org/
100090. 10.21873/cgp.20063.
[68] M.H. Fazel Zarandi, S. Soltanzadeh, A. Mohammadi, O. Castillo, Designing [91] A.L. Dallora, S. Eivazzadeh, E. Mendes, J. Berglund, P. Anderberg, Machine
a general type-2 fuzzy expert system for diagnosis of depression, Appl. Soft learning and microsimulation techniques on the prognosis of dementia: A
Comput. 80 (2019) 329–341, http://dx.doi.org/10.1016/j.asoc.2019.03.027. systematic literature review, PLoS One 12 (6) (2017) 1–23, http://dx.doi.org/
[69] F. Iorfino, et al., Clinical stage transitions in persons aged 12 to 25 years 10.1371/journal.pone.0179804.
presenting to early intervention mental health services with anxiety, mood, [92] J. Cervantes, F. Garcia-Lamont, L. Rodríguez-Mazahua, A. Lopez, A comprehen-
and psychotic disorders, JAMA Psychiatry 76 (11) (2019) 1167–1175, http: sive survey on support vector machine classification: Applications, challenges
//dx.doi.org/10.1001/jamapsychiatry.2019.2360. and trends, Neurocomputing (2020) http://dx.doi.org/10.1016/j.neucom.2019.
[70] V. Sawrikar, et al., Using staged care to provide ‘right care first time’ to people 10.118.
with common affective disorders, Psychiatr. Serv. 72 (6) (2021) 691–703, [93] C. Brunner, A. Fischer, K. Luig, T. Thies, Pairwise support vector machines
http://dx.doi.org/10.1176/appi.ps.202000145. and their application to large scale problems, J. Mach. Learn. Res. 13 (2012)
[71] D.S.E. Broekharst, R. van de Wetering, W. Ooms, R.W. Helms, N. Roijakkers, 2279–2292.
Deploying predictive analytics to enhance patient agility and patient value in [94] F. Javed, S.O. Gilani, S. Latif, A. Waris, M. Jamil, A. Waqas, Predicting
hospitals: A position paper and research proposal, Healthc. Anal. 3 (2023) risk of antenatal depression and anxiety using multi-layer perceptrons and
100141, http://dx.doi.org/10.1016/j.health.2023.100141. support vector machines, J. Pers. Med. 11 (3) (2021) http://dx.doi.org/10.
[72] G. Quer, R. Arnaout, M. Henne, R. Arnaout, Machine learning and the future 3390/jpm11030199.
of cardiovascular care: JACC state-of-the-art review, J. Am. Coll. Cardiol. 77 [95] P.Y. Taser, Application of bagging and boosting approaches using decision tree-
(3) (2021) 300–313, http://dx.doi.org/10.1016/j.jacc.2020.11.030. based algorithms in diabetes risk prediction, in: Proceedings, 74, (1) 2021, p.
[73] H. Jung, K. Chung, Knowledge-based dietary nutrition recommendation for 6, http://dx.doi.org/10.3390/proceedings2021074006.
[96] I.H. Sarker, A.S.M. Kayes, P. Watters, Effectiveness analysis of machine learning
obese management, Inf. Technol. Manag. 17 (1) (2016) 29–42, http://dx.doi.
classification models for predicting personalized context-aware smartphone
org/10.1007/s10799-015-0218-4.
usage, J. Big Data 6 (1) (2019) http://dx.doi.org/10.1186/s40537-019-0219-y.
[74] Z. Fayyaz, M. Ebrahimian, D. Nawara, A. Ibrahim, R. Kashef, Recommendation
[97] M. Khondoker, R. Dobson, C. Skirrow, A. Simmons, D. Stahl, A comparison
systems: Algorithms, challenges, metrics, and business opportunities, Appl. Sci.
of machine learning methods for classification using simulation with multiple
10 (21) (2020) 1–20, http://dx.doi.org/10.3390/app10217748.
real data examples from mental health studies, Stat. Methods Med. Res. 25 (5)
[75] S. Yang, P. Zhou, K. Duan, M.S. Hossain, M.F. Alhamid, emHealth: Towards
(2016) 1804–1823, http://dx.doi.org/10.1177/0962280213502437.
emotion health through depression prediction and intelligent health recom-
[98] Y.T. Jo, S.W. Joo, S.-H. Shon, H. Kim, Y. Kim, J. Lee, Diagnosing schizophrenia
mender system, Mob. Netw. Appl. 23 (2) (2018) 216–226, http://dx.doi.org/
with network analysis and a machine learning method, Int. J. Methods
10.1007/s11036-017-0929-3.
Psychiatr. Res. 29 (1) (2020) e1818, http://dx.doi.org/10.1002/mpr.1818.
[76] F. Iorfino, et al., Right care, first time: developing a theory-based automated [99] H. Yang, J. Liu, J. Sui, G. Pearlson, V. Calhoun, A hybrid machine learning
protocol to help clinically stage young people based on severity and persistence method for fusing fMRI and genetic data: Combining both improves classifi-
of mental illness, Front. Public Heal 9 (August) (2021) 1–8, http://dx.doi.org/ cation of schizophrenia, Front. Hum. Neurosci. 4 (2010) http://dx.doi.org/10.
10.3389/fpubh.2021.621862. 3389/fnhum.2010.00192.
[77] K. Ng, S.R. Steinhubl, C. deFilippi, S. Dey, W.F. Stewart, Early detection [100] V. Rocha-Rego, J. Jogia, A.F. Marquand, J. Mourao-Miranda, A. Simmons, S.
of heart failure using electronic health records: Practical implications for Frangou, Examination of the predictive value of structural magnetic resonance
time before diagnosis, data diversity, data quantity, and data density, Circ. scans in bipolar disorder: a pattern classification approach, Psychol. Med. 44
Cardiovasc. Qual. Outcomes 9 (6) (2016) 649–658, http://dx.doi.org/10.1161/ (3) (2014) 519–532, http://dx.doi.org/10.1017/S0033291713001013.
CIRCOUTCOMES.116.002797. [101] D. Grotegerd, et al., Discriminating unipolar and bipolar depression by means
[78] R.J. Urbanowicz, M. Meeker, W. La Cava, R.S. Olson, J.H. Moore, Relief-based of fMRI and pattern classification: a pilot study, Eur. Arch. Psychiatry Clin.
feature selection: Introduction and review, J. Biomed. Inform. 85 (January) Neurosci. 263 (2) (2013) 119–131, http://dx.doi.org/10.1007/s00406-012-
(2018) 189–203, http://dx.doi.org/10.1016/j.jbi.2018.07.014. 0329-4.
[79] C. Li, J. Xu, Feature selection with the Fisher score followed by the Maximal [102] J. Mourão-Miranda, et al., Pattern recognition analyses of brain activation
Clique Centrality algorithm can accurately identify the hub genes of hepato- elicited by happy and neutral faces in unipolar and bipolar depression, Bipolar
cellular carcinoma, Sci. Rep. 9 (1) (2019) 1–11, http://dx.doi.org/10.1038/ Disord. 14 (4) (2012) 451–460, http://dx.doi.org/10.1111/j.1399-5618.2012.
s41598-019-53471-0. 01019.x.
19
[103] G. Akinci, E. Polat, O.M. Koçak, A video based eye detection system for [119] K. Hilbert, U. Lueken, M. Muehlhan, K. Beesdo-Baum, Separating generalized
bipolar disorder diagnosis, in: 2012 20th Signal Processing and Communications anxiety disorder from major depression using clinical, hormonal, and structural
Applications Conference, SIU, 2012, pp. 1–4, http://dx.doi.org/10.1109/SIU. MRI data: A multimodal machine learning study, Brain Behav. 7 (3) (2017)
2012.6204617. 1–11, http://dx.doi.org/10.1002/brb3.633.
[104] M.-J. Wu, et al., Individualized identification of euthymic bipolar disorder [120] T.Y. Tang, P. Winoto, I should not recommend it to you even if you will like
using the cambridge neuropsychological test automated battery (CANTAB) and it: The ethics of recommender systems, New Rev. Hypermedia Multimed. 22
machine learning, J. Affect. Disord. 192 (2016) 219–225, http://dx.doi.org/10. (2016) 111–138, http://dx.doi.org/10.1080/13614568.2015.1052099.
1016/j.jad.2015.12.053. [121] C.R. Marmar, et al., Speech-based markers for posttraumatic stress disorder in
[105] D. Vergyri, et al., Speech-based assessment of PTSD in a military population US veterans, Depress. Anxiety 36 (7) (2019) 607–616, http://dx.doi.org/10.
using diverse feature classes, in: Proc. Annu. Conf. Int. Speech Commun. Assoc. 1002/da.22890.
INTERSPEECH, vol. 2015, 2015, pp. 3729–3733, http://dx.doi.org/10.21437/ [122] D. Greenstein, J.D. Malley, B. Weisinger, L. Clasen, N. Gogtay, Using multivari-
interspeech.2015-740.
ate machine learning methods and structural MRI to classify childhood onset
[106] L.E. Salminen, R.A. Morey, B.C. Riedel, N. Jahanshad, E.L. Dennis, P.M.
schizophrenia and healthy controls, Front. Psychiatry 3 (June) (2012) 1–12,
Thompson, Adaptive identification of cortical and subcortical imaging markers
http://dx.doi.org/10.3389/fpsyt.2012.00053.
of early life stress and posttraumatic stress disorder, J. Neuroimaging Off. J.
[123] S. Srinivasagopalan, J. Barry, V. Gurupur, S. Thankachan, A deep learning
Am. Soc. Neuroimaging 29 (3) (2019) 335–343, http://dx.doi.org/10.1111/jon.
approach for diagnosing schizophrenic patients, J. Exp. Theor. Artif. Intell. 31
12600.
(6) (2019) 803–816, http://dx.doi.org/10.1080/0952813X.2018.1563636.
[107] D. Rangaprakash, et al., Compromised hippocampus-striatum pathway as a
[124] R.N. Pläschke, et al., On the integrity of functional brain networks
potential imaging biomarker of mild-traumatic brain injury and posttraumatic
stress disorder, Hum. Brain Mapp. 38 (6) (2017) 2843–2864, http://dx.doi.org/ in schizophrenia, Parkinson’s disease, and advanced age: Evidence from
10.1002/hbm.23551. connectivity-based single-subject classification, Hum. Brain Mapp. 38 (12)
[108] R.L. Spitzer, K. Kroenke, J.B.W. Williams, B. Lo, A Brief Measure for Assessing (2017) 5845–5858, http://dx.doi.org/10.1002/hbm.23763.
Generalized Anxiety Disorder, vol. 166, 2015, pp. 1092–1097. [125] W.H.L. Pinaya, et al., Using deep belief network modelling to characterize
[109] C.D. Katsis, N.S. Katertsidis, D.I. Fotiadis, An integrated system based on differences in brain morphometry in schizophrenia, Sci. Rep. 6 (December)
physiological signals for the assessment of affective states in patients with (2016) 1–9, http://dx.doi.org/10.1038/srep38897.
anxiety disorders, Biomed. Signal Process. Control 6 (3) (2011) 261–268, [126] G. Roberts, et al., Functional dysconnection of the inferior frontal gyrus in
http://dx.doi.org/10.1016/j.bspc.2010.12.001. Young people with bipolar disorder or at genetic high risk, Biol. Psychiatry 81
[110] E. Frank, M.A. Hall, I.H. Witten, The WEKA workbench, Data Min. (2017) (8) (2017) 718–727, http://dx.doi.org/10.1016/j.biopsych.2016.08.018.
553–571, http://dx.doi.org/10.1016/b978-0-12-804291-5.00024-6. [127] A.G. Reece, A.J. Reagan, K.L.M. Lix, P.S. Dodds, C.M. Danforth, E.J. Langer,
[111] D. Bhugra, et al., The WPA-lancet psychiatry commission on the future of Forecasting the onset and course of mental illness with Twitter data, Sci. Rep.
psychiatry, Lancet Psychiatry 4 (10) (2017) 775–818, http://dx.doi.org/10. 7 (1) (2017) 1–11, http://dx.doi.org/10.1038/s41598-017-12961-9.
1016/S2215-0366(17)30333-4. [128] T.A. Naqishbandi, Z. Gulzar, A.A. Dar, Understanding impact of communication
[112] P. Kaur, M. Sharma, Diagnosis of human psychological disorders using super- ban on mental health in conflict zone : - experiences from Young kashmiri
vised learning and nature-inspired computing techniques: A meta-analysis, J. research scholars understanding impact of communication ban, J. Loss Trauma
Med. Syst. 43 (7) (2019) http://dx.doi.org/10.1007/s10916-019-1341-2. (2021) 1–4, http://dx.doi.org/10.1080/15325024.2021.1928841.
[113] X.X. Zhao, M. Cui, Y.H. Geng, Y.L. Yang, A systematic review and meta-analysis [129] G. Rong, A. Mendez, E. Bou Assi, B. Zhao, M. Sawan, Artificial intelligence
of randomized controlled trials of palliative care for pain among Chinese adults in healthcare: Review and prediction case studies, Engineering 6 (3) (2020)
with cancer, BMC Palliat. Care 18 (1) (2019) 1–10, http://dx.doi.org/10.1186/ 291–301, http://dx.doi.org/10.1016/j.eng.2019.08.015.
s12904-019-0456-z. [130] M.L. Wainberg, et al., Challenges and opportunities in global mental health:
[114] T.R. Hoens, M. Blanton, A. Steele, N.V. Chawla, Reliable medical recommen- a research-to-practice perspective, Curr. Psychiatry Rep. 19 (5) (2017) http:
dation systems with patient privacy, ACM Trans. Intell. Syst. Technol. 4 (4) //dx.doi.org/10.1007/s11920-017-0780-z.
(2013) 173–182, http://dx.doi.org/10.1145/2508037.2508048.
[131] S. Shoib, S.M.Y. Arafat, COVID-19 and the communication blackouts in Kashmir,
[115] F. Narducci, P. Lops, G. Semeraro, Power to the patients: The HealthNetsocial
India, Lancet Psychiatry 7 (9) (2020) 738, http://dx.doi.org/10.1016/S2215-
network, Inf. Syst. 71 (May) (2017) 111–122, http://dx.doi.org/10.1016/j.is.
0366(20)30338-2.
2017.07.005.
[132] W.A. Tol, et al., Relevance or excellence? Setting research priorities for mental
[116] M.K. Hamza, M.H.-R. Hicks, Implementation of mental health services in
health and psychosocial support in humanitarian settings, Harv. Rev. Psychiatry
conflict and post-conflict zones: Lessons from Syria, Avicenna J. Med. 11 (01)
20 (1) (2012) 25–36, http://dx.doi.org/10.3109/10673229.2012.649113.
(2021) 8–14, http://dx.doi.org/10.4103/ajm.ajm_141_20.
[133] S. Rathod, et al., Mental health service provision in low- and middle-
[117] A. Sau, I. Bhakta, Predicting anxiety and depression in elderly patients using
machine learning technology, Healthc. Technol. Lett. 4 (6) (2017) 238–243, income countries, Heal. Serv. Insights 10 (2017) http://dx.doi.org/10.1177/
http://dx.doi.org/10.1049/htl.2016.0096. 1178632917694350.
[118] A. Sau, I. Bhakta, Screening of anxiety and depression among seafarers using [134] A. Vabalas, E. Gowen, E. Poliakoff, A.J. Casson, Machine learning algorithm
machine learning technology, Inform. Med. Unlocked 16 (2019) 100228, http: validation with a limited sample size, PLoS One 14 (11) (2019) 1–20, http:
//dx.doi.org/10.1016/j.imu.2019.100228. //dx.doi.org/10.1371/journal.pone.0224365.
20

1-s2.0-S2772442523000527-main

Uploaded by

Copyright:

Available Formats

1-s2.0-S2772442523000527-main

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1-s2.0-S2772442523000527-main

Uploaded by

Copyright:

Available Formats

Healthcare Analytics 3 (2023) 100185

Contents lists available at ScienceDirect

A hybrid mental health prediction model using Support Vector Machine,

ARTICLE INFO ABSTRACT

Fig. 1. Depicting proposed mental health prediction and recommendation methodology.

account for the suggested model. Final, recommendations are drawn

The recommendations model is an important aspect of the proposed

Fig. 4. Knowledge database for recommendation.

You might also like