A Study of Machine Learning Models for

Predicting Mental Health through Text

Analysis
2023 1st DMIHER International Conference on Artificial Intelligence in Education and Industry 4.0 (IDICAIEI) | 979-8-3503-3842-3/23/$31.00 ©2023 IEEE | DOI: 10.1109/IDICAIEI58380.2023.10406845

Deepali P. Kadam K. T. V. Reddy

Faculty of Engineering and Technology (FEAT), Faculty of Engineering and Technology (FEAT),
Datta Meghe Institute of Higher Education & Research, Datta Meghe Institute of Higher Education & Research,
India India
[email protected] [email protected]

Abstract - Since forever, there has been a severe

might be suffering from into somatization, substance
problem with mental health illnesses. Most often, people based disorder or postpartum depression through textual
think that a mental health issue exclusively refers to analysis using deep learning.
anxiety or depression. Other psychiatric diseases
A major global public health concern is the incidence
including digital addiction, substance use disorder etc. are
mostly unknown to them. Predicting such diseases can be of mental health conditions. The negative perception
done by observing and studying behavioural patterns, surrounding mental health creates obstacles in seeking
facial expressions, speech patterns, heartbeats, and other timely assistance, limits individuals' access to healthcare
biological markers. We are choosing to go ahead with services, hinders the delivery of optimal treatment,
prediction using text. By analysing the keywords and their contributes to unfavorable outcomes, and raises the
polarity in a person's social media posts, many studies likelihood of human rights infringements on individuals
have attempted to determine whether or not the person is [11]. Research reveals a significant underdiagnosis of
depressed. For digital addiction, substance-based depression in clinics, primarily attributed to patients'
disorders, etc. very few people have chosen automatic text
limited comprehension of the condition. Additionally,
extraction of keywords representic symptoms. Moreover,
deep learning has been proved better than machine inadequate availability of medical resources leads to
learning. Thus, we studied research papers related to delayed treatment. Conventional diagnostic approaches
mental health prediction using text with the help of ML mainly identify individuals who actively seek medical
and DL. assistance, making it challenging to identify patients who
are unaware of their own depression [12].
Keywords – Mental Health Prediction, Text Analysis,
somatization, substance-based disorders, postpartum Even within affluent societies, a substantial treatment
depression gap persists, indicating a disparity between the count of
persons with mental health issues and those who actually
I. INTRODUCTION
receive appropriate care [13]. As demonstrated by ongoing
The WHO defines mental health as a state of mental studies and initiatives, continued attempts are being
health that enables individuals to efficiently manage life's undertaken to fill in this void by incorporating mental
stresses, realise their potential, study, work, and volunteer health treatments into initial care. [14]. Through the
in their communities. Mental disorders or illnesses affect strategies of task sharing and capacity building, endeavors
our emotions, thinking, emotions and actions. They could are made to address this gap [15].
be momentary or ongoing (chronic). It may affect our
Internet counseling serves as a valuable resource for
ability to engage with others and perform daily duties. We
individuals who face barriers to accessing mental health
do express our thoughts with the help of words, let it be in
care due to diverse circumstances [16]. There is evidence
written format or in audio format.
from earlier studies that a person's psychological condition
Mental health disorders have been a serious issue and their linguistic and conversational patterns are related.
since ages now. Usually people have a misconception that [18].
mental health disorder means depression or anxiety only.
Numerous approaches exist for the automatic
They are not aware of other forms psychiatric disorders
detection of emotions in text, with machine learning being
namely somatization, substance based disorder,
one of the well-established methods. Additionally, several
postpartum depression, etc. Numerous methods can be
studies have put forth different deep learning models to
employed to forecast various disorders, including the
address this issue [1,2,3,4,5,6,7,8,9,10]. Traditional
examination and assessment of behavioral patterns, facial
machine learning models are often deemed incomplete as
expressions, speech patterns, heart rate, and other
they primarily concentrate on specific features and fail to
biological indicators. Several researchers have made
capture the contextual information embedded within the
efforts to ascertain if an individual is experiencing
text. Text analysis studies in the field of mental health
depression by examining the keywords they use in their
tend to adopt a generalized approach, primarily focusing
social media posts and analyzing their sentiment. Only a
on identifying the presence or absence of depression.
small number of researchers have explored the automated
Therefore, our aim is to focus on specific mental disorders
extraction of symptoms related to somatization, substance-
and extract relevant information from text data.
based disorders, or postpartum depression. Furthermore,
contemporary findings indicate that deep learning II. MOTIVATION
outperforms traditional machine learning in terms of
achieving superior results. In our study, we will be The stigma associated with mental illness, which
studying the classified mental disorders an individual permeates society, is absorbed by people. Many people
have been prevented from receiving mental health care
Authorized licensed use limited to: Indian Institute of Technology Indore. Downloaded on February 14,2024 at 15:45:49 UTC from IEEE Xplore. Restrictions apply.
because of the stigma. There are several impacts on the available : SD/A/T. Logistic Regression was 2nd with
workplace and society from the level of stigma 72% AUC and Decision tree wass worst with 63%
surrounding mental illness. AUC.(SD: Sociodemographic data,A : Average,T :
Trends) Utilizing the Predictive Modelling Approach on
One way that mental health issues have an effect on
an extensive dataset can enhance the discovery of resilient
society is through maternal depression. It might raise a predictors. There is potential to broaden the scope of
baby's chance of low birth weight, childhood health issues, information derived from emails by exploring diverse
and "incomplete immunisation," all of which are warning avenues, such as employing graphical frameworks, to
flags for an early death. Mental health illnesses are "very uncover additional semantic elements. Further
high," according to the WHO, especially among female investigation is warranted to ascertain the comprehensive
prisoners. intricacy of the text. In [2], the data set (such as data
It advises to countries, both developing and advanced to sources, keywords, and geographic areas), method of data
execute more extensive, preventative and pertaining analysis, ML or DL approach, classifier performance, and
mental health programmes in order to lessen the mental feature extraction method were all coded for the articles
illness's negative effects on people and their local and by Rohizah Abd Rahman et al. Researchers applied SVM
international societies. for mental health problem detection and found accuracy
approx 70%. This study acknowledges limitations in its
This is the reason we are going for prediction of mental article selection process, as it exclusively relied on four
health status of a person in its initial stage only. We are journal databases (PubMed, IEEE Xplore, ScienceDirect,
doing this with the help of digital transcript. Scopus) and included only English language articles. Sole
III. LITERATURE REVIEW reliance on Online Social Network (OSN) for data
extraction pertaining to mental health detection was
Searching by keywords for our topic, we found 50 employed in this study. Various factors, including but not
papers overall from IEEE, ACL, Elsevier, MDPI, ACM, limited to locations, data extraction methods, language
Scopus, other research sources. After removing duplicate considerations, and multiple data sources, were not taken
papers, 48 papers were left. As we examined the abstract into account in this study and should be considered in
and conclusion, only 21 papers were found as most future research. In [3], Computational methods were used
relevant papers for our research. Out of 21, 18 articles are by the Norah Saleh Alghamdi et al. to make use of an
original articles and 03 articles are review articles. No. of Arabic online platform where people talk over and persue
studies included in quantitative analysis are 10. Fig. 1 guidance on a range of mental health conditions.
shows prisma chart for the same. Information is obtained from the aforementioned forum
and is either automatically or manually marked. For the
prediction of sadness from postings, a semi-supervised
technique (lexicon-based approach) and a supervised
approach (machine-learning-based approach) are applied.
By achieving an accuracy of more than 80%, the
dictionary-based technique surpassed the ML-based
approach and approaches from comparable studies. The
absence of considerations for user locations and genders in
the design and development of the prediction solution
hampers its overall efficiency.Other potential
classifications of depression, such as seasonal affective
disorder, bipolar disorder, and babyblues, which may be
investigated for a more thorough analysis, are ignored in
this study, which only focuses on the level of sadness. In
[4], Yang Ding et al. Introduces the DISVM classification
model, which makes use of a deep neural network and the
AdaBoost integration technique. Sina Weibo is used as the
original dataset in Chinese Language. The features include
language choice, emoji usage, Weibo followers, etc. The
findings from this paper is DISVM gives the best result :
88% on training dataset and 86% on testing dataset, when
compared with other algorithms. Accuracy was ascending
from 58% for 3 months and 85% peak for 24 months. In
order to attain the highest recognition accuracy, it is
imperative to employ a more objective approach to
Fig. 1 : PRISMA diagram identify and select the most appropriate data features. As
per Kamil Zeberga et al., in [6], BERT uses textual data to
Several different ML methods to predict the mental represent mental health issues while keeping contextual
health were evaluated. In [1], Mark Hoogendoorn et al., and semantic meanings. As per their findings, SVM works
They investigate social anxiety symptom predictive relatively well among existing classical ML algorithms
modelling and make therapeutic predictions based on with accuracy 84%. Higher classification accuracies of
sociodemographic information and the content of emails 94%, 96%, and 98% were attained with LSTM, Bi-LSTM,
the patient sent while receiving therapy. The dataset and Knowledge Distilled BERT, respectively. To produce
comes from a 69 patient clinical experiment. The good findings, multimodel depression detection systems
fundamentals of mailing behaviour, word choice, writing can be created to make use of more divergent data, such as
style, and sentiment issue are taken into consideration as text, image, and behavioural aspects. In [7], Adenrele
predictors. In their findings, Top performing model was Oduntan et al.use eight machine learning (ML) algorithms
Random Forest with AUC as 78% when exploring all data
Authorized licensed use limited to: Indian Institute of Technology Indore. Downloaded on February 14,2024 at 15:45:49 UTC from IEEE Xplore. Restrictions apply.
and two lexicon-based strategies to categorise diary 81% and Logistic Regression being the least with 79%.
entries as having a positive or negative sentiment polarity. S.A.Farooq et. al [24] propose a Machine Learning based
6023 journal entries from 755 users were examined. They model approach to detect the early onset of Post-traumatic
combined sentiment analysis and thematic analysis Stress Disorder (PTSD) in COVID-19 survivors . The
methods. SVM is the best classifier having a high F1- model utilizes a reduced questionnaire based on
score of 89.7%, accuracy of 89.7% and specificity 88.9%. Diagnostic and Statistical Manual of Mental Disorders ,
By using machine learning techniques to automatically fifth edition ( DSM-5).The paper is based on data
identify individuals' thematic issues from their journal collected from the Kashmir valley where the COVID-19
entries and then suggest suitable interventions. pandemic caused havoc during the pandemic and became
Conducting field research to see whether the application more severe wave after wave. The decision tree (CART)
actually improves users' resilience. In [8], Waleed Ragheb classifier was found to be the most effective and accurate
et al., It has been pondered about how to identify at-risk (97.4%) among the six machine learning algorithms tested
users who exhibit overt symptoms of despair, anorexia, while Logistic Regression classifier was found to be the
self-harm, and suicide ideation. NCNL, a novel deep least accurate with 83.57 % . Shi-Heng Wang et al. [25]
learning ensemble architecture is introduced. Cutting-edge discusses how text mining methodologies and algorithms
Transformer-based backbone models were used. NCNL are applied to a collection of research papers, articles, or
model achieves new state-of-the-art results with F1 as other textual sources in order to discover common themes,
75%. Study is limited to depression, anorexia, self-harm patterns, or topics related to adolescent substance use and
and suicide. Rupali S. Patil presents a study of sentiment depression using Latent Dirichlet allocation (LDA).
analysis of Marathi tweets using supervised classifiers and Latent Dirichlet allocation (LDA) was applied to perform
TF-IDF features in her paper. The main goal of the text mining on the dataset. Word clouds were used to
research is to find a solution to the Sentiment Analysis visually display the content of topics and demonstrate the
problem, which deals with categorising tweets as positive distribution of vocabularies over each topic. The
or negative. Tweets are experimented with several preprocessing was performed by using MALLET
supervised classifiers, including NB, DT, RF and SVF, (MAchine Learning for LanguagE Toolkit) , which is an
using TF-IDF features for feature extraction. When open-source Java-based package for statistical natural
combined with the TF-IDF feature extraction model, language processing, topic modeling, and other machine
MNB Algorithm surpasses all other classifiers with a learning applications to text. Brian E. Perron et al. [26]
highest accuracy of 87.29%. The limited availability of discusses the feasibility of using text mining with
training data for Marathi tweets, in comparison to English Traditional statistical procedures like linear and logistic
tweets, poses a challenge. Enhancing performance regression, cluster analysis to better understand substance-
necessitates the balancing of data through techniques like related problems among families investigated for abuse or
under-sampling or over-sampling, along with the neglect. It employed 2,956 human-coded investigation
incorporation of Twitter-specific features like emoticons summaries from a state child welfare agency to develop
and hashtags. To effectively handle implicit opinions such automated coding models. The study revealed computer
as sarcasm and irony, a fine-grained analysis at the entity models achieving over 90% accuracy in identifying
and aspect level is necessary, moving beyond sentence- substance-related problems (SRPs) in child welfare
level analysis. Tanya Nijhawan1 et al. [10] use deep investigations. Three classification algorithms were used:
learning model BERT for sentiment categorization along a rule-based model, logistic regression, and random forest.
with machine learning methods to perform sentiment Accuracy scores ranged from 78.3% to 93.5%, with
analysis on massive datasets of tweets. They used Latent random forest models excelling in sensitivity and
Dirichlet Allocation, an unsupervised machine learning specificity. These results highlight the cost-effective
technique, to scan a collection of papers and recognize potential of text mining for extracting valuable insights
word and phrase patterns. Random Forest Classifier is from child welfare data. T. H. K. R. Prabhashwaree and N.
with an accuracy of 97.78%. The suggested model lacks M. Wagarachchi [27] predicted mothers's postpartum
consideration for topic specification, which could be an depression using ML approaches. They found that
important factor to incorporate for improved performance. FFANN model (95% accuracy) and the ANFIS - GA
V. Lajjawala et.al [22] compare performances of three model (testing error: 0.0600) have good performance as
classification algorithms , Random Forest , Decision Tree classification and regression types of models, respectively.
and naive bayes , against each other in detecting P. Mazumder and S. Baruah[28] aims to develop a
depression and build a model using the more accurate predictive model for early detection of Postpartum
algorithm . Dataset used is provided by OSMI ( Open Depression (PPD) among women who have given birth
Source Mental Illness ) survey and consists of working within a year. It included 96 participants surveyed through
individuals . Selected features for prediction are : Age , interviews and online forms, focusing on social, mental,
gender , distance of workplace from home , previous and quality of life attributes along with EPDS scores. The
mental illness , family history , work interference . WEKA model achieved accuracy rates of 84% to 94%,
(Waikato Environment for Knowledge Analysis ) is used significantly improved by ensemble techniques. Findings
to test the algorithms and compare the results. Decision showed correlations between PPD and factors like
Tree yielded most accurate results with the accuracy of unwanted pregnancy, relationship with spouse, delivery
82.2 % and also took the least processing time while type, and urban/rural residence. S. Natarajan et al. [29]
Naive Bayes was the least accurate with 78.7 % accuracy aims to predict Postpartum Depression (PPD) using
and most time taken to process resuls with 0.5 seconds . survey data about demographics, depression, and
Work Interefernce was found to be the root node for the pregnancy. Sample size is of 173 responses. Functional-
decision tree. N.Kaur , S.Juliet [23] examine the use of Gradient Boosting (FGB) algorithm showed promising
ML algorithm to predict mental health disorder on a results in predicting PPD and outperformed other
dataset of self reported information Algorithm used in this classifiers such as SVM, Naive baiyes decision tree. M. B.
project are Random Forest, Decision tree, Logistic Carneiro et al. [30] produced a context-aware solution for
Regression, KNN. Random forest was most accurate with gestational depression prevention using text mining. They
Authorized licensed use limited to: Indian Institute of Technology Indore. Downloaded on February 14,2024 at 15:45:49 UTC from IEEE Xplore. Restrictions apply.
 Analyzed pregnant women's documents for sentiment and scientific literature.
identification. Evaluated model performance using
VI. CONCLUSION AND FUTURE WORK
confusion matrix metrics. Proposed model achieved
reliable performance, aiding health professionals in high- Many of the studies used comparatively less datasets,
risk pregnancy monitoring. Text file containing 297 which have limited the broadness of their findings.
phrases, which 165 phrases belong to the happiness class Additionally, many of the studies only used data from
and 132 phrases belong to the sadness class. English-speaking populations, which limits their
applicability to other languages and cultures. Some of the
IV. METHODOLOGY studies used methods for selecting and extracting features
The main initial task of this research will be from the data, which may have introduced bias into the
building a dataset for this research, as it is not readily or results. Additionally, some of the studies did not consider
otherwise available anywhere. As per our best knowledge the impact of different feature types (e.g., text, image,
this is the first time that we are considering the behavioral) on the performance of the detection models.
somatization, substance based disorder and postpartum Some of the studies used traditional machine learning
depression. Thus we will be retrospectively collecting that approaches, while others used deep learning approaches. It
data from patients by taking their clinical interviews or is not clear which approach is more effective for mental
through chatbot. health detection, and more research is needed to compare
the two approaches. A system tends to detects stress only
and not to analyzes the topic of a discussion in a particular
text. Also, The studies are limited to certain mental health
issues and many other disorders are yet to be explored.
VII. REFERENCES
[1] Mark Hoogendoorn et al., “Predicting Social Anxiety Treatment
Outcome Based on Therapeutic Email Conversations”, IEEE
JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS,
2017
[2] Rohizah Abd Rahman et al., “Application of Machine Learning
Methods in Mental Health Detection: A Systematic Review”, IEEE
ACCESS,2020
[3] Norah Saleh Alghamdi et al., “Predicting Depression Symptoms in
an Arabic Psychological Forum” , IEEE ACCESS,2020
[4] Yang Ding et al., “A Depression Recognition Method for College
Students Using Deep Integrated Support Vector Algorithm”, IEEE
ACCESS,2020
[5] Iqra Ameer et al., “Multi-Label Emotion Classification on Code-
Mixed Text: Data and Methods” , IEEE ACCESS,2022
[6] Kamil Zeberga et al., “A Novel Text Mining Approach for Mental
Health Prediction Using Bi-LSTM and BERT Model” , Hindawi
Computational Intelligence and Neuroscience, 2022
[7] Adenrele Oduntan et al., “I Let Depression and Anxiety Drown Me
…: Identifying Factors Associated With Resilience Based on
Fig. 2 : Proposed Methodology Journaling Using Machine Learning and Thematic Analysis” , IEEE
JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS,
The dataset will be annotated manually by therapist / 2022
psychiatrist. Thus we will be basically working with the
[8] Waleed Ragheb et al. , “Negatively Correlated Noisy Learners for
help of supervised approach. Balancing the dataset is the At-Risk User Detection on Social Networks: A Study on Depression,
most essential stage when it comes to any machine Anorexia, Self-Harm, and Suicide” , IEEE TRANSACTIONS ON
learning. So the under-sampling and oversampling of the KNOWLEDGE AND DATA ENGINEERING,2023
dataset will be performed to balance the dataset. As we are [9] Rupali S. Patil, Satish Kolhe, “Supervised classifiers with TF IDF
going for categorization of disorder, the dataset will be features for sentiment analysis of Marathi tweets”, Social Network
normalized as per the categories of disorders. Then the Analysis and Mining, 2022
data will be sent as an input to the classifier i.e. our [10] Tanya Nijhawan1 et al., “Stress detection using natural language
proposed deep learning model. The model will be trained processing and machine learning over social interactions”, Journal of
until we get the desired output. Once the model is trained Medical Systems , 2022
and tested, the result will be evaluated. [11] Afzal Javed etc al., “Reducing the stigma of mental health disorders
with a focus on low- and middle-income countries”, Asian Journal
V. RESULT of Psychiatry, Elsevier, 2021.
[12] Ran Li etc al., “EmoMix+: An Approach of Depression Detection
Our study paper reports research findings, Based on Emotion Lexicon for Mobile Application”, Security and
analysis, and conclusions without providing an actual Privacy in Ubiquitous Mobile Computing, 2022
implementation, as data generation is under process. It [13] E. Kazdin, “Addressing the treatment gap: A key challenge for
presents a problem statement, research methodologies used extending evidence-based psychosocial interventions,” Behav. Res.
upto now and we are planning to use in future, data Therapy, vol. 88, pp. 7–18, 2017.
analysis, and result discussions. While lacking practical [14] T. Rebello, A. Marques, O. Gureje, and K. Pike, “Innovative strate-
application, it offers valuable insights and contributes to gies for closing the mental health treatment gap globally,” Current
theoretical knowledge in a field of mental health prediction Opinion Psychiatry, vol. 27, pp. 308–314, Jul. 2014.
using NLP. Researchers, scholars, and professionals can [15] J. O. Abdulmalik et al., “Country contextualization of the men-
use it as a resource to better understand the conceptual tal health gap action programme intervention guide: A case study
aspects of a mental health diversity and ML-DL from nigeria,” PLoS Med., vol. 10, Aug. 2013, Art. no. e1001501.
applications in it, making it an important part of academic [16] Maili Tairel, “When Do People Seek Internet Counseling? Exploring

Authorized licensed use limited to: Indian Institute of Technology Indore. Downloaded on February 14,2024 at 15:45:49 UTC from IEEE Xplore. Restrictions apply.
 the Temporal Patterns of Initial Submissions to Online Counseling
Services”,
[17] Suicidal Ideation Detection: A Review of Machine Learning
Methods and Applications
[18] M. J. Paul and M. Dredze, “You are what you tweet: Analyzing
Twitter for public health,” in Proc. 5th Int. AAAI Conf. Weblogs
Soc. Media, 2011, pp. 265–272.
[19] B. Moulahi, J. Az´e, and S. Bringay, “Dare to care: A context-aware
framework to track suicidal ideation on social media,” in Proc. Int.
Conf. Web Inf. Syst. Eng., 2017, pp. 346–353.
[20] Sarbhan Singh et al., “A systematic review of depression literacy:
Knowledge, help-seeking and stigmatising attitudes among
adolescents” , Journal of Adolescence, 2019
[21] Aziliz Le Glaz et al., “Machine learning and natural language
processing in mental health: Systematic review”, Journal of Medical
Internet Research, 2021.
[22] V.Lajjawala et.al ,”Classification Algorithms based Mental Health
Prediction using Data Mining”, International Conference on
Communication and Electronics System (ICCES 2020)
[23] N.Kaur , S.Juliet , “Comparative Analysis of ML techniques for
mental health prediction” , International Conference on
Communication and Electronics Systems (ICCES 2020)
[24] S.A.Ffarooq , A.M.Wani , S.Rafiq , “Post-Traumatic stress disorder
(PTSD) prediction in Covid-19 survivors - A CLinical Questionnaire
and Machine Learning Based Approach” , 2022 IEEE North
Karnataka
[25] Shi-Heng Wang,Yijun Ding,Weizhong Zhao,Yung,Hsiang
Huang,Roger Perkins,Wen Zou,James J. Chen , "Text mining for
identifying topics in the literatures about adolescent substance use
and depression ", Wang et al. BMC Public Health
[26] Brian E. Perron a, Bryan G. Victor b, Gregory Bushman a, Andrew
Moore a, Joseph P. Ryan a, Alex Jiahong Lu a c, Emily K. Piellusch ,
"Detecting substance-related problems in narrative investigation
summaries of child abuse and neglect using text mining and machine
learning" ,
[27] T. H. K. R. Prabhashwaree and N. M. Wagarachchi, “Predicting
Mothers with Postpartum Depression using Machine Learning
Approaches”, 2022 International Research Conference on Smart
Computing and Systems Engineering (SCSE), Colombo, Sri Lanka,
2022, pp. 28-34
[28] P. Mazumder and S. Baruah, “A Community Based Study for Early
Detection of Postpartum Depression using Improved Data Mining
Techniques”, 2021 IEEE International Conference on Computation
System and Information Technology for Sustainable Solutions
(CSITSS), Bangalore, India, 2021, pp. 1-7
[29] S. Natarajan, A. Prabhakar, N. Ramanan, A. Bagilone, K. Siek and
K. Connelly, “Boosting for Postpartum Depression Prediction”, 2017
IEEE/ACM International Conference on Connected Health:
Applications, Systems and Engineering Technologies (CHASE),
Philadelphia, PA, USA, 2017, pp. 232-240
[30] M. B. Carneiro, M. W. L. Moreira, S. S. L. Pereira, E. L. Gallindo
and J. J. P. C. Rodrigues, Recommender System for Postpartum
Depression Monitoring based on Sentiment Analysis, 2020 IEEE
International Conference on E-health Networking, Application &
Services (HEALTHCOM), Shenzhen, China, 2021, pp. 1-6

Authorized licensed use limited to: Indian Institute of Technology Indore. Downloaded on February 14,2024 at 15:45:49 UTC from IEEE Xplore. Restrictions apply.