AI framework for DRIVE model based mental health detection in text: a case study on how coping strategies are expressed during COVID-19

Research questions and contribution

This study seeks to create an artificial intelligence framework that helps to understand and classify the stressors, resources, and human response and choices of coping strategies. The study demonstrates the framework partially by identifying the coping strategies that individuals are using under long stressful conditions (COVID-19 lockdown).

Specifically, the research questions and contributions can be summarized as follows:

• 1.

  Define an artificial intelligence framework that can identify and classify individual demands, resources, and coping strategies from text posted on social media.

• The article provides a theoretical framework to observe individual’s mental status expressed in text. The framework is designed to extract the features required to detect individual’s resources, demands, and coping strategies based on the DRIVE-model. The experiments also demonstrate the predictive reliability of the coping strategy classification component of the framework.

• 2.

  Identify the coping strategies that are used by individuals during stressful situations.

• The article provides a quantitative analysis and distribution of the coping strategies that were found in individuals’ posts on the platform X during COVID-19.

• 3.

  Provide evidence that in-text coping analysis is more effective in detecting the mental status and performs better than pure SA.

• The article analyzed the distribution of SA labels within each coping category and showed the limitation of SA to detect a number of positive and negative coping strategies. The work also provides examples of tweets with negative or neutral SA labels that are found to represent positive coping categories (and vice versa). This demonstrated that in-text coping analysis is better than pure SA.

• 4.

  Find the set of words and topics that people have used on social media to express their coping strategies with COVID-19-related stress.

• The article lists original examples of tweets for each coping strategy which reflect a set of expressions of their emotions and coping strategies. The word bigram analysis also provides a visual view of the most frequent word structure of each coping category.

The rest of this article is organized as follows. The following section provides a theoretical background of the DRIVE model and the text-based social media mining. Next, a threefold approach to the literature review of related previous work is provided. First, the literature that applies the DRIVE model to evaluate MH status is given. The main research that applies text-based social media mining to MH-related problems is discussed. This is followed by a review of machine-learning approaches that specifically analyze coping strategies expressed in social media posts in terms of descriptive models, statistical analysis, or qualitative data analysis. Then, the proposed framework for the DRIVE-Coping machine learning system is explained, followed by the experimental design. Experimental results and discussions are given before concluding with the main findings and listing any limitations.

DRIVE mental health model

Various indicators contribute to well-being, and combining these factors provides the most accurate prediction of outcomes. Well-being encompasses and is informed by diverse theories and models, such as the well-being appraisal framework, which includes multiple measures reflecting its various components (Hart & Wearing, 1995). The DRIVE model, developed by Mark & Smith (2008), offers an enhanced and flexible approach to well-being by incorporating subjective perceptions, resources, and individual differences. This conceptual framework combines previous frameworks, including elements from previous models like the Demand Control Support (DCS) model, the Effort-Reward Imbalance (ERI) model, coping behavior, attributional explanatory styles, and outcomes such as job satisfaction, depression, and anxiety (Mark & Smith, 2008).

The DRIVE model effectively predicts the effects of individual differences and work characteristics, such as coping style, on outcomes like depression, anxiety, and job satisfaction. However, there are uncertainties regarding moderating relationships (Mark & Smith, 2008; Mark & Smith, 2012a, 2012b). The DRIVE model has been supported in various contexts, including studies on the psychosocial effects on migrant workers in Italy (Capasso, Zurlo & Smith, 2016a, 2016b, 2018), UK postgraduate psychology students and nurses (Galvin et al., 2015), and university staff in the UK (Williams & Smith, 2016). Evidence for moderation effects has been limited (Galvin et al., 2015; Williams & Smith, 2016), with few interactions between predictor variables. The strength of the DRIVE model lies in its simplicity and flexibility making it a practical and multi-dimensional tool.

In particular, demands refer to external stressors, challenges, and obstacles that individuals encounter in their environment (Macía et al., 2021). These can include work-related pressures, interpersonal conflicts, health issues, financial difficulties, or major life events. Demands vary in intensity and can trigger coping responses in individuals. On the other hand, resources encompass the internal and external assets that individuals can utilize to cope with demands effectively. These include personal strengths, social support, coping skills, and environmental supports. Resources play a crucial role in buffering the impact of stressors and promoting resilience.

However, individual effects acknowledge that people vary in their coping styles, preferences, and responses to stress (Johnson & Smith, 2021). Factors such as personality traits, cultural backgrounds, past experiences, and coping strategies influence how individuals navigate and cope with stressful situations. Effects encompass the subjective perceptions, interpretations, and meaning-making processes that individuals engage in when encountering stressors (Lakey & Cronin, 2017). How individuals appraise and interpret stressors influences their coping strategies, emotional responses, and well-being outcomes.

Recent research has focused on several concentrated coping theories, including Lazarus and Folkman’s Transactional Model of Stress and Coping (Lazarus & Folkman, 1984), which emphasizes the dynamic interaction between individuals and their environment. It also explores its application to cultural influences on coping strategies (Matsumoto, 2010; Lazarus, 1993). This has contributed to the field with the Cognitive-Behavioral Coping Theory, highlighting the role of cognitive appraisal in choosing coping strategies. Research the effectiveness of cognitive-behavioral coping interventions in various populations (Hofmann et al., 2012). Folkman & Lazarus (1980) delineated between problem-focused coping (addressing stressors directly) and emotion-focused coping (managing emotional reactions), with recent investigations exploring the balance between these strategies and their impact on well-being (Zimmer-Gembeck & Skinner, 2023).

Additionally, transactional Model of Coping provided by Skinner et al. (2003) expanded on these theories by integrating cognitive appraisal, coping strategies, and outcomes into a dynamic framework, applications in understanding coping in diverse contexts (Carver & Connor-Smith, 2010). These theories provide valuable insights into how individuals navigate stressors and manage their responses, contributing to the development of effective coping interventions in psychology.

Coping strategies are categorized in various ways across the literature, reflecting different dimensions of coping behavior (Skinner et al., 2003; Carver & Connor-Smith, 2010). Problem-focused coping strategies refer to active engagement aimed at addressing the stressor and finding a solution; meanwhile, emotion-focused coping strategies aim at managing the emotional reactions to a stressor. Adaptive coping strategies lead to positive outcomes and well-being, whereas maladaptive coping strategies can exacerbate the stress-related difficulties. Mindfulness practices involve focusing on the present moment, without any judgement, which helps in increasing the self-awareness and reducing the stress reactivity (Keng, Smoski & Robins, 2011). Problem-solving refers to actively addressing the stressors and finding a solution via a systematic process of goal-setting and seeking problem-solving steps (Nezu, Nezu & D’Zurilla, 2012). Emotion regulation strategies include cognitive reappraisal (changing one’s cognition to change one’s emotions), expressive suppression (literally suppressing or controlling one’s emotions) and emotion acceptance (accepting one’s feelings and not trying to control them) (Gross, 2015). Social support is having the needed psychological and material support to help and individual to cope with stress. It plays a vital role in resilience and moderately influences coping effectiveness (Lakey & Cronin, 2017). Positive reappraisal, which refers to changing the way one looks at a stressful situation in a positive light, fosters resilience and psychological well-being and coping efficacy (task-related coping) (Tugade & Fredrickson, 2004). These coping strategies are an integral part of managing stress and one’s reaction to stressors, psychological well-being and resilience in individuals.

Text-based social media mining

The expansion of social media has led to the development of major techniques in the field of text mining (Karami et al., 2020). The main text-based techniques include: SA or opinion mining, ED, topic modeling and content analysis, event or trend detection, and human behavior analysis.

The SA methods, most notably VADER (Hutto & Gilbert, 2014), Sentiwordnet (Sebastiani & Esuli, 2006), and TextBlob (Loria, 2018), are lexicon-based NLP techniques that aim to classify the emotional value of a text expression into a multi class classification of emotions (positive, negative, and neutral). To improve its performance, the SA has been expanded by deploying supervised machine learning models and by improving its lexicon-base dictionary. The ED is an extension of the SA that is also a keyword-search based approach. It searches for emotion keywords that are assigned to psychological states according to specific psychological models such as sadness, joy, fear, disgust, anger, and so on. Examples of classical ED models and emotion lexicon dictionaries can be found in numerous studies, e.g., Roberts et al. (2012), Mohammad & Turney (2013).

Topic modeling, most notably the Latent Dirichlet Allocation (LDA) (Blei, Ng & Jordan, 2003) and the Structural Topic Model (STP) (Roberts et al., 2014), is a major application of text mining that is used to analyze the text content of social media posts. The aim is to learn the main topics of discussion by social media users (Laureate, Buntine & Linger, 2023). By analyzing the frequency of words and their co-occurrences in documents (social media posts), the topic modeling technique is able to identify the underlying themes of the set of documents. Another application of text mining in social media stems from the need to detect a topic or a theme from the text such as emerging events or trends (Afyouni, Al Aghbari & Razack, 2022), rumors and false information (Laureate, Buntine & Linger, 2023), or people’s stance detection (AlDayel & Magdy, 2021). The last related technique of text-based social media mining is the human behavior detection in which the emotional, social, and cognitive behaviors of humans are analyzed in the context of text mining (Gutierrez et al., 2021). The upcoming section provides a brief explanation of the most relevant work in which text-based social media mining techniques were applied in the context of MH.

Related work

The related work is reviewed based on three main themes. First, the conceptual foundation of applying the DRIVE model to understand the definition of the well-being within the field of psychology is given. Second, the related work of applying text-based social media mining to mental health problems is explained. The third subsection is focused on reviewing the previous work on analyzing and detecting coping strategies in social media.

Framework evaluation

The feasibility of the framework was studied by developing and evaluating the proposed predictive coping classification model. The objective of the model is to take user posts as input features and to predict the type of the coping strategy (positive or negative). To train the model, the data was divided into 80% and 20% for training and testing, respectively. Each combination of features in Table 2 was fitted into one pipeline and used as input for the classifier models. Table 4 shows the prediction accuracy results averaged over the classifier model (row) and over the feature set (column). On average, Feature Set II, which included the text features only, had the greatest accuracy, followed by Feature Sets IV, V, and VI which included the text features, the user’s number of following, and different subsets of SA features. K-nearest neighbor performed better than the other classifiers followed by the naïve Bayes. The greatest accuracy was achieved by K-nearest neighbor in Feature Set V. Since the objective was to demonstrate the effectiveness of the coping framework, future work is left to explore other feature engineering approaches, learning models, and the use of LDA topics or other word embedding techniques to enhance model performance.

To further investigate performance, the dataset with the feature subset No. II in Table 2 was selected to demonstrate the effectiveness of the classifier, given that it resulted in the best accuracy on average among all datasets. A summary of the performance results for this case is given in Table 5. For each model, the weighted averages of accuracy, precision, recall, and F-1 score are listed. Among the models, logistic regression, naïve Bayes, and SVM had consistently higher accuracy scores of 79%. Naïve Bayes and SVM stand out for the F-1 score, achieving 77% on average. This highlighted their balance between precision and recall. Conversely, decision tree and logistic regression exhibited the lowest F-1 scores with accuracy percentages of 67% and 69% respectively.

Table 6 shows a comparison of the two models that had the highest accuracy scores among all datasets. These were the K-nearest neighbor trained on the dataset with the feature subset V and the naïve Bayes trained on the dataset with the feature subset II. The confusion matrices of the models are shown in Fig. 3. The class “True” corresponds to the negative coping class. For the K-nearest neighbor, there were 128 true negatives (TN) and 33 true positives (TP). The precision for the “False” class (i.e., positive coping) was 86%, denoting a high proportion of correctly identified positive coping strategies used by individuals among all instances that were classified as positive coping. The recall was slightly higher with 89% for the recalled positive coping strategies. However, for the “True” class, i.e., negative coping, the precision was lower, at 67%, while the recall was 62%.

For the naïve Bayes model trained on dataset II, the accuracy was 79%, which was slightly lower than the K-nearest neighbor model trained on Dataset V. The precision and recall for identifying instances of the positive coping were 81% and 93%, respectively. However, the precision and recall measures were lower for the negative coping instances, reaching 68% and 40% respectively.

The receiver operating characteristic (ROC) curve plots the true positive rate against the false positive rate at various threshold settings. The closer the ROC curve is to the upper left corner of the graph, the better performance of the model, which is represented as the area under the curve (AUC). An AUC greater than 0.5 implies that the classifier has a better ability to distinguish between classes than a random guess, demonstrating acceptable predictive reliability (Fawcett, 2006). Figure 4 illustrates the ROC curves of the K-nearest neighbor model for Dataset V (left) and the naïve Bayes model of Dataset II (right). By approximating the AUC, the KNN model achieves a score of approximately 0.7, whereas the naïve Bayes model yields an AUC of around 0.6. The performance of both the K-nearest neighbor and the naïve Bayes demonstrate acceptable predictive reliability as both of the curves were above the 0.5 threshold line of accidental methods. The K-nearest neighbor model has predicted the coping category better than the naïve Bayes.

Identifying coping strategies

Examining the labeling of the random sample set revealed a majority of positive coping tweets, making nearly 70% of the sample. Examples of posts that represented positive and negative coping strategies based on the manual labeling by the team are given in Tables 7 and 8 respectively. Any user or personal name, brand name, company name, commercial channel or show are shown as XYZ or ellipsis (…) to maintain anonymity and ensure impartiality. The text of the tweets is kept in its original form except for a few spelling errors. As the users are mostly non-native English speakers, there are some structural and grammatical issues.

The distribution of the coping categories is given in Fig. 5. It can be seen that the most used coping strategy is the positive “mindfulness practices” strategy, followed by “positive reframing” and “engaging in relaxation techniques”. The negative coping strategies (“venting of emotions”, “avoidance”, and “denial”) come next in frequency. The strategies that people used the least were “self-blame” and “wishful thinking” (negative coping) followed by “seeking social support” and “problem-solving” (positive coping).

The main techniques that are linked to the positive strategy labeled “mindfulness practices” were reciting prayers, meditation practices, expressing gratitude and blessings, and being sarcastic and humorous. The “positive reframing” strategy included practices related to work adaptation, spreading positive news, and having a positive perception of the events. Similarly, the “engaging in relaxation techniques” strategy involved self-care activities like spa and body treatments and participating in sports or entertainment activities. The negative “venting of emotions” strategy was connected to the spread of negative news or thoughts, while the expressions of hopelessness were considered part of the negative “avoidance” strategy. Conspiracy theory talk was mainly connected to the negative coping strategy labeled “denial”. In parallel, any wishing for how it was in the past (future) or for impossible activities or events were linked to the negative strategy labeled “wishful thinking”. Any self-blaming statements that considered the pandemic a punishment was labeled as negative “self-blame” coping mechanism.

Comparison between in-text coping analysis and SA

Comparing the coping mechanisms with the SA scores, Fig. 6 illustrates the frequency distribution of the VADER (left) and TextBlob (right) labels grouped by positive and negative coping categories. Assuming that positive coping and positive sentiment analysis represent a good mental health outcome (and vice versa), then the SAs that have more positive labels under the positive coping category would be a better indicator of coping. The figures show that VADER is more aligned with the coping categories than TextBlob, with 400 tweets labeled positive which are considered positive coping compared to nearly 340 tweets for TextBlob. Approximately 150 tweets which represented negative coping were labeled negative using VADER sentiment analysis, compared to 100 negative coping tweets that were labeled negative by the TextBlob sentiment analysis. There are a lower number of neutral-labeled tweets under VADER than under TextBlob. Nonetheless, both SA methods could not recognize a number of the expressed coping strategies, labeling them positive when they were negative coping and vice versa. Examples of tweets for which the VADER and/or TextBlob categories have miscategorized their coping category are listed in Tables 9 and 10 for the positive coping strategies and for the negative coping strategies, respectively. The examples provide evidence of how some coping strategies that involve the use of negative terminology, given their context, could confuse the SA analysis.

To look deeper into the performance of the SA within each coping strategy, a drill-down examination of the SA categories was done to identify the tweets for which the coping type and the SA category disagreed. Table 11 provides the number and percentage of tweets that the TextBlob or VADER classified as a negative or neutral sentiment while they were a positive coping and vice versa. A visual chart of the miscategorizaed statistics is given in Fig. 7. The VADER SA had, on average, a lower average miscategorized percentage for both positive and negative coping strategies than the TextBlob SA. The highest miscategorized percentage for TextBlob was found in the “wishful thinking” strategy (80%) and the lowest (37.31%) for VADER was found in the “positive reframing” strategy. The TextBlob had, on average, more miscategorization of the coping strategies than VADER did, with most of the disagreements stemming from the larger neutral class. Both SA methods had the largest miscategorization in the negative coping which exceeded 50% of the negative data. This use of negation, subjectivity, and sarcastic expressions may have influenced the SA performance.

Linking words and text to coping strategies

An analysis of co-occurring words using bigrams NLP models was conducted to gain insights into the relationships between words in the positive coping and negative coping text datasets independently. Figure 8 shows a network of the top 50 occurring bigrams in the positive coping text corpus. The bigram “oh god” appeared in the top ten bigrams in the positive coping data which was related to the mindfulness practice of praying. The “wear mask” and “covid vaccine” bigrams, which are specific to COVID-19-related problem-solving techniques, also appeared in the top ten positive bigrams. Figure 8 shows more bigrams that can be linked to positive coping strategies. Examples included: “stay safe,” “social distancing,” “oh bless,” “return normal,” “live long,” and “thank god.”

Figure 9 shows the top 50 occurring bigrams with respect to the negative coping text corpus. The bigram “swear god” appeared as the top bigram in the negative coping category. The bigrams “must solved,” “month isolation,” “leave family,” and “isolation must” also appeared in the top 12 bigrams in the negative coping category. Other bigrams included: “biological war,” “god deconfinement,” “god forbidden,” “go mall,” “swear suffocating,” and “mentally ill.

What was also interesting is the emojis within each coping category. Since social media users depend heavily on emojis for expressing their feelings, the bigrams that defined the emojis were prevalent for both the positive and negative categories. The list of the top bigrams in the positive coping dataset included the “tear joy,” “folded hand,” “red heart,” “check mark,” “smiling eye,” “victory, hand,” and “rolling laughing” emojis. The list of the top bigrams in the negative coping dataset included the “broken heart,” “loudly cry,” and “double exclamation” emojis. Figure 10 lists graphically the top Emojis used in the positive and negative coping categories based on the bigram analysis. Moreover, the analysis reveals that positive coping tweets contain a wider variety of emojis compared to the negative coping tweets.

In addition, LDA analysis was applied to the positive and negative coping datasets separately to identify their main themes and terminologies. Table 12 lists the topics found by LDA. In a psychological context, applying LDA to the positive coping dataset identified latent themes that emerged from individuals’ experiences with effective coping strategies. Within the negative coping dataset, the LDA analysis helped uncover topics related to barriers to coping and revealed challenges in managing crises.

Feasibility study of the coping detection

The classification models performed reasonably well with a simple feature set that consisted of text related features only. This provides promising evidence of the underlying structures for the coping strategies. Additionally, adding a user related feature (for example the number of following) along with the SA score had an impact on the classification performance. However, the inclusion of the SA features, as in Dataset V, introduced confusion and reduced the model’s ability to detect the positive coping category. The non-parametric (K-nearest neighbor) and the probabilistic (naïve Bayes) models performed better than other learning methods. The K-nearest neighbor algorithm was more effective in capturing instances of individuals who had negative coping experiences. naïve Bayes was better able to identify those instances of individuals who used positive coping strategies. This provides some insight into the diverse nature of negative coping strategies, suggesting that non-parametric based models could be more suitable.

Thus, a number of insights can be inferred given the performance of the classification model. First, the small sample size of the data and the unbalanced distribution of the positive and negative instances of coping strategies have an impact on the way negative coping strategies are classified based on their underlying structure. This makes a non-parametric distance-based algorithm perform better. Secondly, the probabilistic approach performed better in general and in predicting the positive coping strategies. This is expected due to the variety and subjectivity of the expressions which makes it harder to find a predictive function that represents all the possible expressions. To enhance the results, strategies such as optimizing hyperparameters, addressing class imbalances, and refining feature selections or feature engineering could be considered.

Coping strategies analysis

Understanding the major types of positive and negative coping mechanisms and identifying the strategies within each type provided an insightful understanding of people’s behaviors during stressful conditions. It is important to highlight that 70% of the sample identified positive coping strategies, with “mindfulness practices” and “positive reframing” being the most noticeable among the sample in Kuwait during COVID-19 quarantine. This illustrates the importance of these strategies in navigating the challenges posed by the pandemic in this cultural context. The “mindfulness practices”, such as praying, meditation, and spiritual practices have been extensively studied for their effectiveness in reducing stress, enhancing emotional regulation, and promoting overall well-being (Al-Ghazali, 1997; Tugade & Fredrickson, 2004). The “positive reframing” strategy, which involves consciously focusing on the positive elements of a situation or finding meaning in stressful situations, has been scientifically proven to be linked to greater resilience and psychological well-being (Tugade & Fredrickson, 2004).

In the overall context of the COVID-19 quarantine, where individuals faced intensified levels of stress, uncertainty, fear, and social isolation, the implementation of “mindfulness practices” and “positive reframing” served as effective coping strategies. These strategies helped individuals manage the psychological impact of quarantine, maintain a sense of control despite the uncertainty, and promoted adaptive responses to adversity (Tugade & Fredrickson, 2004; Khoury et al., 2015).

Similarly, the results showed three main negative coping styles: “venting of emotions”, “avoidance”, and “denial”. It is important to highlight that small business owners and immigrants residing in labor areas with strict lockdown restrictions faced job insecurity and extreme isolation. Consequently, the lack of resources such as social support, job security, and financial aid increased the effects of negative coping styles during times of crisis such as the pandemic. LDA topics and negative bigrams reflected the expressions of such coping strategies. For example, negative feelings, mainly sadness, dispirit, fear, and anger, were present in all the discovered LDA topics including negativity toward the lockdown (first topic), the vaccine (second topic), city isolation (third topic), and job insecurity (fourth topic). Individuals employing these coping strategies may have experienced heightened levels of psychological distress and decreased overall well-being during periods of isolation and uncertainty.

Likewise, “avoidance” strategies were expressed as talks of hopelessness in many tweets. The third and fifth topics in Table 12 and bigrams such as “month isolation”, “swear suffocating” and “mentally ill” reflected the despair felt by users in the isolated cities at not being able to return to normal life. The “denial” coping strategy during the pandemic appeared mainly as conspiracy theories surrounding the source of the virus and doubts about the legitimacy of the virus. The bigram “biological war” provides evidence of such talk, inhibiting individuals from acknowledging and addressing their emotional responses to the pandemic and delaying effective adaptation and use of effective coping strategies.

Comparisons between coping analysis and SA

The results in Tables 10 and 11 show that SA methods, although effective in detecting the polarity and subjectivity of the text, do not capture the outcomes of the MH coping strategies. The strategies: “wishful thinking”, “denial”, and “avoidance” are negative coping strategies. However, these were the most confusing strategies for the SA methods. Expressions of these coping strategies have been categorized as having positive or neutral tones of emotion by the SA. The highest miscategorized percentage was found for the “wishful thinking” given that it usually contained good but unrealistic wishes. The second most strategy miscategorized is the “denial” strategy. In this category, the conspiracy-related discussions contributed significantly to the confusion and misclassification.

Conversely, another miscategorization was detected when exploring the outcome of SA on positive coping expressions. For example, “positive reframing” and “problem-solving” include mentions of the negative aspects of a situation but are presented within a positive context, making the SA analysis vulnerable to categorize the sentiment as negative. In addition, expressions that represented “engaging in relaxation techniques” were found to have neutral or sometimes negative expressions. The highest percent of miscategorized statements was linked to the seeking social help strategy since its expressions usually included an explanation of the problem.

In general, negative coping strategies were miscategorized more often with SA than the positive strategies were, (66.57% vs 51%). One of the main reasons for this is related to the contradiction between the positive expression of the statements and their negative hidden contextual meaning. The VADER sentiment analysis resulted in fewer miscategorizations than the TextBlob method in both the negative and positive categories. This outcome was expected given TextBlob’s simplicity and limitation in analyzing social media posts.

Cultural influence and its effect on resources and stressors

The high rate of positive coping strategies used in Kuwait during COVID-19 quarantine highlighted the influence of cross-cultural influences and resource availability effect on coping strategy (AlSumait, AlHeneidi & Smith, 2021). Kuwait’s socio-economic context, characterized by stable health services, job security, financial stability, and food security for a majority of the population, contrasts with situations in other regions where these resources were limited. As a result, individuals in Kuwait may have exhibited higher levels of adaptive coping strategies, such as seeking in-house social support, “engaging in relaxation techniques”, and maintaining routines, due to the relative abundance of resources available to them. Additionally, the cultural norms and values in Kuwait, such as strong social networks and spiritual practices, contributed to the use of effective coping strategies.

Understanding these cultural and contextual factors is fundamental for developing targeted interventions to support mental health and well-being during COVID-19 quarantine periods in Kuwait and other similar contexts. The LDA and bigram analyses provide evidence for this justification. For example, the first, second, and sixth topics within the LDA topics in Table 12 revolve around prayers, availability of resources such as vaccines, family, and medical care. Further evidence of prayer practices is provided by positive bigrams such as “oh god”, “oh bless”, and “thank god.” The rest of the topics and positive bigrams reflect normal positive responses to stressful situations. This relates to our analysis that the users had good resources to overcome the stressors during COVID-19.

However, the appearance of factors such as spiritual practices and beliefs, family, and individual quarantine experiences in both positive and negative coping data revealed that the relationship between an individual’s psychological state and those factors is complex and highly personalized, given the individual’s experience. While these factors can be sources of strength and resilience for some individuals, they may pose challenges and intensify stress for others. Coping strategies associated with adaptive coping and psychological resilience were identified, as well as factors that contribute to psychological distress and vulnerability.