A Social Media Study on Mental Health Status Transitions Surrounding Psychiatric Hospitalizations

2.1. Clinical Studies of Psychiatric Hospitalization Experiences

Since the deinstitutionalization movement [10], the role of the psychiatric hospital has shifted from a place for long-term stay and treatment to a community-based system of care emphasizing reducing feelings of dependency, and supporting integration [139]. Today, when patients are admitted to a psychiatric bed, the goals of clinical care revolve around crisis stabilization, diagnosis, and initiation of appropriate treatment [139]. Once patients’ symptoms stabilize, this model encourages rapid discharge from the hospital so that individuals may continue receiving care in outpatient settings.

However, psychiatric hospitalizations are still life-altering, as the admission often implies that individuals are unequipped to manage their psychiatric needs and require removal from their existing environment to receive appropriate, urgent care [32]. When an individual is admitted to a hospital for mental illness, it is often due to an adverse event like self-harm or suicidal ideation, lack of insight or denial of illness, social crises such as relationship problems, or non-compliance with medication [89]. Based on the symptoms, treatment during hospitalization involves individual and group therapy, psychotherapy, pharmacotherapy, or other standing medication [162]. Such circumstances increase stress experienced by the patient and their caregivers, leading to perceptions of hospitalization-related anxiety, and fear of confinement [120]. Recovery journeys of people post-hospitalization are similarly challenging due to the high likelihood of re-hospitalization [153], lack of support, non-adherence and side effects to medication, and difficulty in managing the condition and reintegrating back to social life and roles [89]. Paksarian et al. [120] found 69% of participants reporting at least one of their hospitalizations as traumatic, with the most common experiences related to rigidity and involuntary hospitalization, being put on restraints, and being forced to take medications. Ådnanes et al. [3], on the other hand, found participants reporting re-hospitalizations to be less traumatizing than the first hospitalization, a necessary relief, as occurring by default and without progress, and as part of the recovery process.

Whether positive or negative, while re-hospitalizations reflect the clinical aspects of recovery journeys, social processes are found to be equally important. Based on a systematic review and narrative synthesis of the literature, Leamy et al. [94] identified five categories (CHIME): connectedness, hope and optimism about the future, identity, meaning in life, and empowerment to engender recovery processes. Hope, agency, opportunity for purposeful activity, and social inclusion are measured as outcomes of recovery in mental health [117]. In fact, successful social reintegration involving resuming “age, gender, and culture appropriated roles, statuses and activities” [47] and community participation [2] can help reduce stigma and improve overall well-being [155].

Underpinning this interplay of clinical and social processes is the recovery approach [9], as introduced in Section 1. Notable here is the work of Corey Keyes, who has advocated viewing recovery as flourishing in life despite having a mental illness and relying on two complementary reintegration experiences: the restoration from mental illness and the optimization of positive mental health [129]. These perspectives put the individual in the center and highlight the importance of understanding interrelationships over time and episodes like hospitalizations as part of “life course experiences with health” [146]. Combining medical records information (e.g., hospitalizations) and social media data of people suffering from mental distress, in this paper, we adopt the recovery and person-centered care approach in our characterization of mental health status transitions to capture social and clinical aspects of peoples’ hospitalization experiences.

2.2. Study of Mental Health Experiences in the Social Computing Literature

In the last decade, social media has emerged as an important tool in mental health, meeting various social, technical, public health, and clinical goals. The first category of this research includes computational and predictive approaches using social media data to infer peoples’ mental health states, targeting different conditions [15, 27, 35, 134], platforms [27, 141], and disciplines [16, 17, 34, 72]. Interest in social media data stems from the fact that it is readily available and can be unobtrusively gathered, overcoming challenges in state of-the-art clinical assessments of mental health that involves subjective recollection of historical facts – a method prone to retrospective recall bias [38]. We refer to these works appropriating social media data to inform clinical efforts around early diagnosis, tailoring treatment, or delivering patient-provider interventions as perspectives on clinical care in mental health.

In a second category, we find approaches to study people’s motivations, goals, and practices in using social media for support and help-seeking behaviors, often outside of clinical treatment. For instance, Andalibi et al. found that individuals struggling with negative emotions, such as that related to depression or self-harm, use Instagram to self-disclose and engage in storytelling about their stigmatized experiences [8]. Further relating to stigmatized experiences around mental health, both qualitative and quantitative studies have identified how platform interactions lead to self-disclosure, social capital and social support goals and outcomes [7, 44, 115, 119, 163]. We consider this body of research to provide perspectives on social care in mental health.

However, pathways to social care (relating to support, disclosure, stigma, etc.) and clinical care (relating to symptom management, etc.) are not dichotomous. Both social relationships and clinical care are required to improve the overall well-being of those with mental illnesses [155]. Despite this, current approaches in social computing to study mental health are heavily focused either on diagnostic interventions or management of the illness and social support, as noted above.

Ernala et al. [54] critiqued this disjoint consideration of the mental health experience in social computing literature. This work found that social media-based “proxy diagnostic signals” used to characterize people’s mental illness diagnoses on social media do not generalize to patient populations; thus revealing significant behavioral differences between patients with a valid clinical diagnosis who use social media, and those seeking help for their mental distress online. To realize the potential of social media in mental health, we argue that it is crucial to consider the mental health experience of an individual as a whole, intertwining their clinical and social recovery journeys. In this paper, we synthesize viewpoints from both clinical and social care perspectives to study how mental health related major life transitions – hospitalizations – are exhibited on social media. For this, we combine medical records comprising clinical information on diagnosis and hospitalizations with social information about online self-presentation and social interactions.

2.3. Social Media, Health Transitions and Liminality

“Transition” is a concept widely used in the social science literature and is most commonly defined as “a process of convoluted passage during which people redefine their sense of self and redevelop self-agency in response to disruptive life events [93].” While some conceptualize transitions as linear processes with a clear beginning and end [154], others note that transitions can be “complex, nonlinear, sometimes cyclical and potentially recurring” [65, 92].

3.1. Participant Recruitment and Data Collection

For this study, we utilized Facebook data of consented participants with a medical history of at least one hospitalization due to a mental health condition. This data was collected as a part of a broader research study involving the paper’s authors, aimed at identifying technology-based mental health information to provide early identification, intervention and treatment to patients with psychiatric disorders. The research protocol was approved by the Institutional Review Board (IRB) of the coordinating institution managing patient recruitment – a large healthcare organization in the north-east of the United States, as well as the local IRBs at collaborating sites.

Individuals over 15 years of age were recruited from various inpatient and outpatient psychiatric departments at the coordinating and partner institutions. Participants were eligible if they were diagnosed with a schizophrenia spectrum disorder, mood disorder with and without psychotic features, borderline personality disorder or anxiety disorder based on clinical assessment scales (e.g., the Psychiatric Diagnostic Screening Questionnaire or PDSQ [164]) and formal clinical examination conducted by a licensed clinical psychologist, and facilitated by the Structured Clinical Interview for DSM-5 (SCID) [145]. Participants also experienced at least one hospitalization for the mental health condition. Informed consent was obtained from participants after describing to them the research study, type of data to be collected, policies for storage and use, clinical risk mitigation protocols, and clarifying that their relationship with the medical institution would remain unaltered whether they chose to participate in the study or not. Consented participants included 142 psychiatric patients diagnosed with schizophrenia, 106 diagnosed with mood disorders, and 6 with other mental health conditions (Table 1). To answer the RQs, in downstream analyses, we combine participants with schizophrenia, mood disorders, and other mental health conditions into a single study population. While the objective outcomes of recovery, such as time taken for remission of symptoms and probability of relapse, vary across these conditions, clinical literature suggests commonalities in subjective recovery and reintegration experiences like a growing sense of agency and autonomy, quality of life, peer support, greater participation in normative activities, etc. [40]. As a formative investigation into transitions around hospitalizations, we focus on understanding the subjective experiences of recovery and reintegration transitions, thus, we combined analysis across conditions. All data collected from these participants were de-identified and stored in HIPAA compliant secure databases and servers, which were located at the coordinating institution with access privileges limited to only the core project personnel.

Table 1.

Descriptive statistics of participants.

Variable	Statistics	Distribution
Demographics
Age	Mean = 24.0, Median = 22.4, Range = (15.1, 60.7)
Gender	Male | Female (55% Female)
Race	Native American/Indian, African American, White, Other
Ethnicity	Hispanic/Latino | Non-Hispanic/Latino (85% Non-Hispanic/Latino)
Diagnosis	Psychosis, Mood disorders, Other disorders
Facebook data
#Posts	Mean = 2596.3, Median = 1471.0, Range = (0, 11018.0)
Duration (days)	Mean = 1025.16, Median = 730.0, Range = (1, 3287)
Medical data
#Hospitalizations	Mean = 2.3, Median = 2.0, Std = 2.14, Range = (1, 17)
Duration (days)	Mean = 15.5, Median = 11, Std = 13.7, Range = (1, 104)
Gaps (days)	Mean = 365.2, Median = 175.5, Std = 513.9, Range = (2, 4179)

Open in a new tab

3.2. Curating Facebook Data around Hospitalization Events

The goal of this paper is to identify different journeys of people during life transitions around a psychiatric hospitalization, as observed and expressed on Facebook. In doing so, we note that not all hospitalization events will have similar experiences, even for the same individual. If psychiatric hospitalizations are considered major life transitions, the effects of each hospitalization as experienced by individuals might be different. Therefore, in this work, we consider each hospitalization event per consented participant as a different observation.

To gather Facebook data surrounding each hospitalization, we first collate the admission dates for all hospitalization events in each participant’s medical history. Using the hospitalization admission dates as temporal markers, we center the participant’s Facebook data such that the day of admission to the hospital is treated as day 0. Then we extend the temporal window before and after the hospitalization (day 0) and stop only when we reach another hospitalization admission for the same individual. Note that for a participant with only one hospitalization, we include their entire Facebook archive from the earliest to most recent date. We repeat this process for every hospitalization event recorded per participant. Across all 254 participants, we have a total of 584 recorded hospitalization events; however, only 372 hospitalizations had some digital traces from Facebook archives during that time. Corresponding to each of the 372 hospitalizations, we gather and center the participant’s Facebook data around their admission date. The distribution of the number of posts surrounding each hospitalization is shown in Figure 1. The average number of posts surrounding each hospitalization event is 3162.44 (median = 670.0, std = 8503.81). Across all hospitalization events, the minimum and maximum number of Facebook posts are 1 and 86,465 respectively.

Next, to analyze data across all hospitalization events (which ranged from 2009–2019), we identify a fixed time period preceding and succeeding each hospitalization. We adopt an empirical approach by generating cumulative density functions (CDFs) of the number of posts¹ before and after each hospitalization event. These CDFs are shown in Fig. 1. Based on these figures, we observe that most users (around 70–80%) have posts for at least 230 days before and 160 days after each hospitalization event. Therefore, we choose Facebook timeline data spanning 230 days before and 160 days after each hospitalization event as a fixed-length time period for downstream analyses.

4.1. Theoretical Framework: Possible Selves

Corresponding to RQ1, our goal is to identify and understand patterns on Facebook that characterize peoples’ individualized mental health states around psychiatric hospitalizations. To do so, we adopt the Possible Selves framework [103], as introduced earlier. The possible selves concept is used in psychology to complement current conceptions of self-knowledge and to capture cognitive representations of alternative versions of the self in the future. In more detail, possible selves represent “individuals’ ideas of what they might become, what they would like to become, and what they are afraid of becoming, and thus provide a conceptual link between cognition and motivation” [103]. Possible selves are largely based on past experiences, but their essence lies in clear references to the future [103] – they are cognitive representations of hopes, fears, and fantasies regarding oneself. Thus, per the Possible Selves framework “an individual’s collection of self-conceptions and self-images can include the good selves (the ones we remember fondly), the bad selves (the ones we would just as soon forget), the hoped-for selves, the feared selves, the not-me selves, the ideal selves, the ought selves. They can vary dramatically in their degree of affective, cognitive and behavioral elaboration” [103].

As described in Section 2, we consider psychiatric hospitalizations to be major life transitions in an individual’s life. Psychotherapy research has argued that surrounding this liminality, it is important to consider the patient’s “possibilities” (i.e., the possible future states of the self) as an important instrument of effecting a change [80]. These possible selves can be understood as a kind of positive resource that the patient draws on when making desirable changes in their behavior or self-regulation to manage their underlying mental illness, around the hospitalizations. Prior clinical literature has studied individuals’ possible selves in relationship to their diagnosed mental health condition itself [31, 84, 87]. For instance, Janis et al. found that participants with borderline personality disorder were less likely than controls to endorse positive possible selves as current, but more likely to endorse negative possible selves as “current, probable, desired, and important” [87]. In [31], Clarke found that being positive about achieving possible selves was positively related to functional outcomes in first episode psychosis. We hypothesize that people’s language, behaviors and self-presentation signals surrounding hospitalization, as expressed on Facebook, may represent the various possible selves or “future-projected” aspects of self-knowledge, that they perceive as potentially possible.

Our rationale is grounded in the fact that existing behavior change theories (such as the Transtheoretical Model [128] or the Theory of Planned Behavior [4]), often used to capture health transitions (refer Section 2.3), have been criticized for their lack of adaptive capabilities as well as their inability to take into account an individual’s unique psychological state, social context, activity, and behavior patterns [77]. Furthermore, these models rarely match the reality of health transitions, due to differences between actual and perceived behaviors and assumptions related to homogeneity within behavioral change [127]. In contrast, our conceptualization of mental health statuses using the Possible Selves framework allows us to consider that the same individual can exhibit multiple possible selves on Facebook during a certain period before or after a hospitalization. Also, different individuals may express different possible selves on Facebook at any given time preceding or succeeding psychiatric hospitalizations because of their contrasting life situations. Finally, as one navigates the transition caused by the psychiatric hospitalization, these possible selves may evolve and change over time. In this paper, we refer to an individual’s diverse mental health statuses surrounding hospitalizations as a collection of their possible selves.

To operationalize and capture possible selves surrounding hospitalizations guided by this framework, we adopt the following empirical approach described in the remainder of this section.

4.2. Identifying Temporal Phases around Hospitalizations

Studying life transitions surrounding psychiatric hospitalizations necessitates identifying data spanning pre- and post-hospitalization phases where the self-presentation and behavioral changes are most likely to be manifested. Prior work in CSCW and HCI on behavioral changes during major life transitions found abrupt declines in posting activity on social media and noted it as a sign of social withdrawal [55]. For instance, in individuals challenged with postpartum depression, changes in sociality, and behavior on social media manifested through patterns of posting volume [41]. Other work has also noted a sudden increase in posting volume referred to as a “rush” of excitement for the future in the case of life transitions like engagement, starting a new job, or having a child [43]. Also, per clinical literature, we expect people to show markers of social withdrawal, which is known to be a notable risk marker around hospitalizations [97]. Therefore, we use measures of changes in posting volume of an individual (normalized number of posts per day) to identify temporal phases around psychiatric hospitalizations. Our phase identification approach is adopted from Ernala et al. [55] and includes the following two steps:

1. First, we calculate the daily posting volumes on Facebook timeline around each of the 372 hospitalization events spanning 230 days before and 160 days after the hospitalization. We computed the rates of change throughout the before and after hospitalization periods by employing a weekly moving average model on this posting volume time series data. This would allow us to smooth out local fluctuations and seasonality while allowing comparison between the posting volume at day t and that during the seven days preceding it.

2. Next, we compute the medians of the weekly rates of changes during the periods before and after hospitalization and use a median split method to define phase boundaries. Specifically, the first time point (day) in the pre-hospitalization data when the rate of change of posting volume becomes higher than the pre-hospitalization median rate of change is taken as a cutoff. Similarly, the first time point (day) in post-hospitalization data when the rate of change of posting volume becomes lower than the post-hospitalization median rate of change is taken to indicate another cutoff.

As shown in Figure 1, the median rate of change in posting volume pre-hospitalization was 0.0007 and 165 days prior to hospitalization (day 0) is the first time the rate of change surpassed this median. Similarly, the median rate of change in posting volume post-hospitalization is 0.0005 and 23 days after the hospital admission is when this rate is lower than the post-hospitalization median. Since the rates of changes in posting volume are computed weekly, we adopt the following day demarcations to define four temporal phases around hospitalization:

• Bf_long or Long before hospitalization (N = 214): 223 to 165 days prior to hospitalization.

• Bf_hosp or Before hospitalization (N = 271): 165 days prior to the hospitalization day.

• Af_hosp or After hospitalization (N = 130): 23 days after the hospitalization admission.

• Af_long or Long after hospitalization (N = 166): 23 to 160 days after the hospitalization.

Thus, segmenting each of the 372 hospitalization events into four phases, we obtain 781 phases.

4.3. Modeling Possible Selves Around Psychiatric Hospitalizations

Next, to define and identify people’s individualized mental health states, or their possible selves during the above identified four temporal phases surrounding psychiatric hospitalizations, we build a Gaussian Mixture Model [108]. This approach has been used in prior HCI research to capture the heterogeneity in people’s social roles and their evolution [159].

4.4. Clinical Validation and Grounding of the Possible Self Statuses: A Taxonomy

Validating the output of generative models like Gaussian Mixture Model components or Latent Dirichlet Allocation based topic models is typically done by human coding tasks [55], goodness of fit or predictive likelihood measures, performance on external tasks or validation of coherence [71]. However, qualitatively interpreting the derived components is challenging due to lack of contextual knowledge, difficulty in understanding the operationalization of features and researcher bias. Recent work has reflected on the convergence and divergence between statistical machine learning methods (especially unsupervised approaches) and grounded theory method [13] and suggests hybrid, iterative approaches that combine the two [13, 113] as possible alternatives. We take inspiration from this literature and prior work interpreting GMM components [159] to finalize the final setting of number of GMM components and their labels.

Using the approach in Section 4.3, we fit Gaussian Mixture Models for values of K that showed optimal fit based on the AIC, BIC scores. Then for each model’s extracted components, we find the top behavioral features that are representative of the component. To get the top features per component, we use two measures: 1) We take the feature means per component, which in the form of z-scores, shows whether the behavior was performed more or less than its average value. Based on the magnitude of z-scores, we filter the top features and their mean values per component. 2) We build a linear regression model on the probability of a phase belonging to the component and extract features that are statistically significant and predictive of membership to the component.

We presented the top features per component for each of the GMM models to 4 annotators and gathered their input to help interpret and name the extracted components as possible self statuses. To provide additional context for interpretation, we also provided them with the temporal periods during which the component was most prominent (for instance, whether most phases predicted to belong to a component were right before hospitalization). Using this information, we conducted individual semi-structured interviews and follow-up sessions with annotators to iterate over the component labels and reach the most theoretically grounded and clinically informed GMM setting that would capture the PSS surrounding the hospitalizations. The four annotators included one psychiatrist and two computer science researchers who were domain experts in mental health and social media studies and a graduate student familiar with the Facebook data. The first author conducted the sessions to iteratively build a shared vocabulary combining the annotators’ expertise in clinical care and behavioral analysis of social media data. Annotators compared different GMM model outputs based on the discernability of the components and whether the extracted components were comprehensive. Based on feedback from these sessions, we chose K = 6 as the final GMM configuration. Table 2 shows the names of the six PSS alongside the top representative features per PSS as well as example paraphrased Facebook posts, explaining the behaviors within the PSS. We used moderate levels of disguise [20] while paraphrasing posts in Table 2, i.e. identifying details (such as places) were changed and verbatim quotes were modified grammatically to safeguard privacy of participants. Fig. 2 shows the frequency of occurrence of each PSS in our data.

1. Self-regulation focused PSS: involving lower engagement, reduced posting, and activity signaling a detachment from the online social network with no indications of disclosure about their mental health status (indicated by lower use of pronouns). This PSS also suggests boundary regulation strategies [150] while only sharing positive content on Facebook. This PSS is seen most frequently during Bf_hosp and is presented with lower frequency during other periods.

2. Self-awareness focused PSS: involving high self-referential thinking and pre-occupation indicated by the increased usage of pronouns [1] and posting of content related to mental health symptoms and conditions. The higher use of words related to anger, work, and money and showing a focus on the present and future in language demonstrate an understanding of the transition and subsequent consequences. This PSS also has reduced social interactions and posting about emotional content. This PSS is most salient during Bf_hosp but also persists during Bf_long and Af_long indicating continued self-reflection surrounding psychiatric hospitalization events [126].

3. Sociality focused PSS: involving heightened sociality demonstrated by the greater volumes in posting status updates and photos, sending messages, one-on-one interactions, and other activities on Facebook. This PSS only appears during Bf_hosp.

4. Withdrawal focused PSS: involving an overall reduction in posting and activities on Facebook. As part of this PSS, people use significantly fewer function words and personal pronouns and reduce posting content indicative of mental health symptoms and experiences. The overall lack of activity and engagement shows that people are possibly withdrawing from the active use of social media. This PSS is most salient during Af_hosp.

5. Re-adjustment focused PSS: involving increased sociality, such as adding new friends, sending messages, and co-tagging others. But this PSS also reveals more self-attentional focus and expressive behaviors, shown by usage of words about anger, health, and death without inhibition. This PSS is most salient during Af_hosp and is less visible during other phases.

6. Incorporation focused PSS: involving the inclusion of experiences and narratives related to mental health transition into peoples’ online social lives. The increased use of emotional content and personal pronouns, showing a focus on the present and sharing content about their mental health status indicates self-focus, awareness and disclosure of experiences on Facebook. On the other hand, this PSS also shows signs of increased sociality demonstrated by messages shared and one-on-one interactions with others. This PSS is most salient during Af_hosp and persists with lower frequency before.

Fig. 3 shows a heatmap of membership probabilities of each PSS around the four periods surrounding the hospitalization. In essence, annotators found that the derived PSS captured the heterogeneity in peoples’ experiences along two main themes: symptomatic expression before hospitalization and social reintegration after hospitalization. Self-regulation, self-awareness, and sociality focused PSS that are most salient before hospitalization reveal different manifestations of peoples’ social media use and disclosure levels regarding their mental health condition online. While some people might choose to share their experiences related to mental health on Facebook (as in the self-awareness focused PSS), others might regulate and censor content regarding their mental health status transition (self-regulation focused PSS). This variability in social media use is also noted in prior work in people managing depression [56] and disclosures about schizophrenia diagnosis [54]. Our expert annotators also pointed out that the various PSS that are salient before hospitalization show varying levels of awareness and insights that people might have while experiencing mental health symptoms. For instance, during psychotic and manic episodes, that are most commonly experienced by individuals diagnosed with schizophrenia and mood disorders respectively (the majority of our study population), individuals are known to have distinctive levels of awareness, pre-occupation, and emotion regulation[24]. Annotators also identified the withdrawal, re-adjustment, and incorporation focused PSS that are most salient after hospitalization periods as different trajectories into social reintegration. While some people might withdraw themselves away from social technologies after discharge from the hospital to cope with the stigma and repercussions [69], others might increase activity on these platforms to get back to their online social lives and create a “new normal” [105].

Although five of the six PSS occur with sizable frequencies, the sociality PSS occurs very infrequently. Examining the model, we found that the behaviors in this component were outliers and always captured as a separate cluster is all GMM models. We omit this PSS in subsequent analysis due to the infrequent occurrence and lack of discernable behaviors grounded in literature.

5.1. Using Possible Selves to Understand Recovery Trajectories

Fig. 6(a) shows the 10 most likely transitions seen around recovery trajectories i.e. the PSS exhibited before a hospitalization to the PSS exhibited after the hospitalization. First, we see stability in the presentation of withdrawal and incorporation-focused PSS; those who exhibit these PSS Bf_hosp are likely to maintain it Af_hosp. The stability of this PSS is also confirmed by the fact that withdrawal and incorporation focused PSS are the most common trajectories we see Af_hosp – 48.3% of self-regulation, 31% of incoporation and 28% of self-awareness focused PSS Bf_hosp transition into withdrawal focused PSS, and 56% of re-adjustment and 30% of self-awareness focused PSS transition into the incorporation focused PSS Af_hosp.

However, there is a difference between transitions that end in withdrawal focused PSS and those that end in the incorporation focused PSS. This depends on the whether the event was the first hospitalization experienced by the individual. Considering only those who have portrayed the incorporation focused PSS Bf_hosp, 66.6% of the transitions from incorporation to withdrawal focused PSS are seen during the first hospitalization events experienced by people. On the other hand, the majority of the cases where individuals remain in the incorporation focused PSS before and after the hospitalization (58.3%) are observed during subsequent hospitalizations (like the 2nd, 4th, or 5th hospitalization recorded for the individual). This reveals that while those who have been re-hospitalized can maintain an incorporation focused PSS, the first hospitalization experience leaves most people transitioning into the withdrawal focused PSS. This could be attributed to the stigma, isolation, or “other”-ing experience related to first time psychiatric hospitalizations [112]. This finding is also consistent with prior literature on mental health care that suggests that negative experiences associated with the first psychiatric hospitalisation remains significant even after many years have passed [3]. In a focus group study with participants from 6 countries, Ådnanes et al. note that psychiatric rehospitalizations encompassed some amount of familiarity; however, participants described the first hospitalization as “something shocking, intolerable, and terrible” [3]. For example, consider participant A whose recovery transition showed to shift from a self-awareness focused PSS to a withdrawal focused PSS after their first hospitalization. Participant A posted 15 status updates on Facebook during the two weeks prior to their hospitalization, such as the following paraphrased posts: “Live as if you were to die tomorrow and learn as if were to live forever; The only competition you’ll ever face is with your own ignorance.” However, since their hospitalization A made no posts on Facebook, except accepting new friends requests. In contrast, consider participant B. Around their fifth hospitalization B showed a recovery transition from re-adjustment focused PSS to reintegration focused PSS with paraphrased posts such as “Don’t lend people money because that person might be a *** and not pay you back.” before hospitalization to posts such as “my brother..the love we share is eternal. You live on in your loved ones memories and hearts.” after hospitalization.

5.2. Using Possible Selves to Understand Reintegration Trajectories

As a second deeper dive into recovery trajectories, Fig. 6(b) shows the top 10 transitions that begin after the hospitalization and end long after the hospitalization. The post-discharge period after hospitalization involves a shift in focus from institutionalized treatment, to coping and management of illness. This transition is accompanied by mechanisms to break inhibitions and stigma, open up about their experiences with intimate others, reach out for social support and improve overall well-being [78]. The two most common trajectories we observe during the post-hospitalization period end in the self-awareness or incorporation focused PSS. While both of these PSS involve self-referential thinking and self-reflection, only the incorporation focused PSS shows signals of social functioning (such as messaging friends, one-on-one interactions, co-tagging, etc.) that are considered positive signs of reintegration. Among people who transition into the incorporation focused PSS, 53.8% transition from the self-awareness focused PSS, and 24% are those who have maintained the incorporation focused PSS. Studies from the psychiatry literature report that the stigma related to hospitalization could be transitory in cases where people resume occupancy of normal societal roles [68]. This suggests that the transition from self-awareness focused PSS to incorporation PSS is a positive reintegration trajectory where people move past the uncertainty post-hospitalization and re-establish their online social connections as part of the incorporation-focused PSS.

Next, we find that every PSS exhibited immediately after the hospitalization (Af_hosp) has a transition ending in the “self-awareness” focused PSS long after the hospitalization (Af_long). While self-focus demonstrated by the awareness focused PSS is beneficial, excessive self-reflection is considered a central feature in mood and anxiety disorders [118], and greater focus on self-concept is also related to internalized stigma in mental health [99, 147]. In contrast to the social stigma about mental illness that entails discrimination, negative stereotypes, loss of opportunities, etc., self-stigma relates to awareness, agreement, and application [37] – a person with mental illness must first be aware of corresponding stereotypes before agreeing with them and then apply self-stigma to one’s self. A behavioral consequence of self-stigma is social avoidance [37, 100], that is the person may avoid situations where they might feel publicly disrespected because of self-stigma and low self-esteem. Awareness, reflection and social avoidance that are related to self-stigma are representative features of the self-awareness focused PSS, that most individuals are found to transition into long after hospitalization (Af_long). Based on this literature, the majority of reintegration trajectories we see ending in the self-awareness focused PSS might be reflective of the long term self-stigma, that happens as a consequence of psychiatric hospitalizations. For instance, participant C who transitioned into a self-awareness focused PSS long after their hospitalization posted on Facebook about the stigma they have been experiencing, saying (paraphrased): “I think the stigma about mental health really needs to be broken. I’m tired of every single person treating me like the plague. For the last year, I have been dealing with bi polar disorder. It was the scariest thing I ever dealt with but I learned that I’m not alone. People automatically treat you like your a different person, but don’t show me your sympathy. Treat me the same. I’m stronger than I’ve ever been.”

7.1. Theoretical Implications

Anchoring on psychiatric hospitalizations as a liminality, in this paper, we combined clinical perspectives – around symptomatic expression and recovery, with social perspectives – around stigma and reintegration. Guided by the Possible Selves framework [103], we provided an empirically derived taxonomy to represent and then understand individuals’ mental health status transitions. Our approach thus enabled incorporating people’s heterogeneous experiences around recovery and reintegration in the computational modeling of mental health. By combining the two perspectives, we presented an exploration of the intersection and the significance of considering social and clinical perspectives on mental health. This illuminates important theoretical implications.

As noted in Section 2.2, computational approaches to study mental health within HCI, digital psychiatry, and machine learning frame the individual in two broad, distinct ways: as a patient who is receiving clinical care and institutionalized treatment for a validated diagnosis [15, 43], or as a vulnerable individual who is seeking social care and support for management of a mental health condition, such as from their networks of loved ones and peers, offline and online [7, 8, 119]. The word patient comes from the Latin “patiens,” from “patior,” meaning to suffer or bear. Scholars have expressed criticism of this conceptualization. It tends to objectify the person with mental illness [116], disregards their sufferings and identity outside of the clinical definition of the illness [85], and fails to consider the challenges people need to overcome beyond symptom management to get back to life following hospitalizations [129]. In contrast, when the clinical facet of one’s mental illness trajectory is ignored only to consider their vulnerability and self- and peer-supported management efforts, this latter conceptualization fails to account for the obstacles an individual faces in navigating a formalized treatment plan in concert with their own attempts. Research has shown that in some cases this lack of consideration of the relationship of social and clinical care can interfere with evidence-based treatment, create tensions in the patient-provider alliance, and even lead to detrimental mental health outcomes [117].

The repercussions of this disjoint consideration have been noted in recent social computing literature [60]. Chancellor et al. highlighted that the representation of people’s experiences and data in machine learning work on mental health – as a patient, disorder, data point or a person – may inadvertently risk dehumanization, poor and incomplete characterization of the mental health experience, and present serious consequences to scientific rigor [26]. Relatedly, as also mentioned in Section 2.2, Ernala et al.’s triangulation study on methods for predicting mental health states from social media data surfaced the differentiating uses of social media for mental health, showing that representations of people’s experiences who make disclosures on social media and seek social support for mental health, poorly generalize to those of patients receiving clinical care [54]. At the crux of these observations is the fact that peoples’ mental health statuses and needs are not isolated, discrete attributes, but are temporally situated experiences of the same person who is in transition, given the dynamic nature of their illness.

Framed in terms of Van Gennep’s liminality [154], using our derived taxonomy of PSS, we could examine the transitions of individuals as they enter and leave the role of a patient (Section 5). For example, an individual could move from the self-regulation focused PSS they expressed as a non-patient before hospitalization, to the withdrawal-focused PSS post-hospitalization, that is reminiscent of them being a patient at the hospital facility. Our taxonomy could further reveal the transitions that are indicative of reintegration after psychiatric hospitalizations (Section 6), such as those shifting from the self-awareness PSS immediately after they were a patient, to the incorporatation-focused PSS in the longer-term, as a way to embrace the non-patient identity again. This way, we showcased how a holistic view of an individual’s mental health state, as indicated in their clinical recovery and social reintegration can be derived by understanding their various possible selves surrounding hospitalizations. We did so by demonstrating a theoretically- and clinically-grounded computational approach that amalgamates insights from people’s medical histories as well as their social media data.

Drawing upon our work, future research leveraging digital traces for prediction of mental health states can benefit from noting that people’s mental health statuses and their exhibition on social media are heterogeneous, and contextualized within the individuals’ specific needs for clinical treatment and social support. Multimodal approaches, like Gaussian Mixture Models [108], that we adopt in RQ1, when formulated with theory (such as the Possible Selves framework [103]) and punctuated with clinical insights (Section 4), can reveal differentiating patterns of behaviors and use of social media within the same population. This way our work contrasts the theoretical models used to understand health transitions in prior CSCW research [101], which consider that the experiences of all individuals across the illness journey are the same, and that these transitions are largely uni-directional and permanent. Adopting a personalized approach, our taxonomy gives three distinct sets of behaviors that are dominant before psychiatric hospitalization – self-regulation, self-awareness, and sociality focused PSS. These characterizations of mental health statuses can enrich personalized predictive models for early interventions that consider an individual’s clinical as well as social pathways of care. Similarly, research primarily focusing on social support mechanisms for mental health could leverage how clinical aspects to care such as hospitalizations, medication adherence etc. impact support outcomes. For instance, in this study our proposed empirical framework of transitions revealed that first hospitalization experiences lead most people to transition into the withdrawal focused PSS, whereas those re-hospitalized are able to maintain the incorporation focused PSS. The support seeking goals and outcomes for the former group might be significantly distinct from the latter – insights that could contextualize how different individuals use of social media to find help and advice around their mental distress.

7.2. Clinical Implications

The taxonomy we discovered for RQ1 (Section 4) on distinct PSS, the empirical framework in Section 5 to capture mental health status transitions from RQ2, and the predictive models to assess recovery and reintegration in Section 6 (RQ3) can be useful to improve patient-provider interactions in treatment and management of mental illnesses. We discuss two such opportunities below.

7.3. Design implications

7.4. Limitations and Future Directions

Our work has some limitations. First, we note caveats in our data. The data on medical records comes from consented participants seeking treatment within a single healthcare system in a specific geography comprising English-speaking, Western populations. The demographics therefore, are skewed and we caution against sweeping generalizations. We also acknowledge that our examination of social media is limited to a single platform, Facebook. Moreover, we combined different mental health conditions into a single study population, to understand recovery and reintegration transitions surrounding hospitalization. Future work can extend the current findings to explore commonalities and differences in the transitions experienced by people with different illnesses and comorbid conditions. Second, we were able to capture aspects of clinical care based on hospitalization dates and social care based on digital traces from Facebook data. However, not all aspects of social care might be visible on social media. For instance, caregivers and offline connections, community integration programs and support groups both offline and online also contribute to social aspects of care in mental health. Future work can explore these offline facets. Similarly, the use of the PSS-based trajectories in clinical contexts (as discussed in section 7.2) requires additional rigorous validation in future work. Finally, for easier interpretation, we examined linear transitions that people experience around recovery and reintegration. However, literature suggests that transitions can be non-linear and cyclical [92]. Future work can explore modeling strategies and in-depth interviews to uncover these complex transitions, and pre- and post-hospitalization experiences.