Digital Phenotyping to Quantify Psychosocial Wellbeing Trajectories after Spinal Cord Injury

Abstract

Objective

Explore feasibility of smartphone-based digital phenotyping methods to examine depression and its relation to psychosocial wellbeing indicators after spinal cord injury (SCI).

Design

Smartphone research platform obtained smartphone sensor and survey data among community-living adult wheelchair users with SCI. Weekly measurements over four months included Patient Health Questionnaire-8, SCI-Quality of Life Satisfaction with and Ability to Participate in Social Roles and Activities (SRA), GPS-derived community mobility metrics, health conditions, and physical activity.

Results

Forty-three individuals were enrolled. Study retention was higher among individuals offered financial incentives (78%) compared with participants enrolled prior to incentives (50%). Participants who dropped out more commonly had nontraumatic or acute SCI, were older, and had less satisfaction and lower participation in SRA. Among 15 individuals with complete data, half had ≥one week of mild depression. Those with depression had frequent health issues, and less satisfaction and lower participation in SRA. Those without depression experienced increased social engagement over time. Average community mobility was similar across depression groups. Relationships were typically in-phase, but also varied by individual.

Conclusion

Smartphone-based digital phenotyping of psychosocial wellbeing after SCI is feasible but not without attrition challenges. Individual differences in depression patterns highlight clinical utility of scaling these methods.

Introduction

Spinal cord injury (SCI) profoundly impacts all aspects of life, and results in seemingly insurmountable barriers to daily life. Depression is common after SCI,¹ with rates ranging from 19 to 37% ^{2, 3} regardless of demographic or injury-related characteristics.³ Of these, 78% have increased risk for suicide,¹ yet only 8–29% receive treatment. ^{3, 4} Moreover, due to underutilization of mental health services, those with SCI are among the highest at risk for social isolation, restricted mobility, and poor health.^{1, 5, 6}

Depression is not a fixed health condition, but symptom severity varies with time. In fact, those with SCI can demonstrate a range of patterns in depression: high unremitting, high with resolution, latent onset, and low stable.^{7, 8} These patterns of depression can be understood as distinct courses of depression symptom severity over a period of time. Higher depression severity is correlated with less social engagement and community mobility,^{5, 9} and lesser depression is related to better health and more exercise participation.^{10, 11} Collectively these constructs are indicators or contributors to psychosocial wellbeing, operationally defined as happiness, satisfaction, and functioning effectively¹² in both psychological and social domains. However, current understanding of depression and its relationships to social engagement, community mobility, health, and physical activity is limited by a lack of robust, time-dependent measures. Numerous factors influence the trajectory of psychosocial wellbeing. Psychosocial distress is most often characterized via questionnaire-based data at discrete time points three to six to twelve months apart.^{7, 8, 13} Hence, trajectories cannot be predicted to allow preventive interventions when most needed. Additionally, what is reported in the clinic or laboratory may not reflect daily life, due to recall errors, response bias, or inherent variability.¹⁴

Smartphone-based research platforms allow self-reported symptoms in the naturalistic setting, minimizing response bias and more accurately tracking depression severity across time.¹⁴ Greater measurement frequency also facilitates phasic correlations of wellbeing to leading or lagging indicators, such as social engagement, community mobility, health status, and exercise participation. Another strength of smartphone platforms is collection of behavioral information with minimal user input. In fact, phone use logs and global positioning system (GPS)-derived mobility metrics have been suggested to be among the most indicative objective features to determine depression.¹⁵ Moreover, individual-specific models provide more accurate predictions than models averaged from an entire sample.¹⁶ Hence, these methods are highly relevant to a personalized medicine approach to patient care; predictions from even a biweekly frequency may forecast an escalation in symptom severity and inform timely interventions.¹⁷ The resulting rich dataset produces a digital phenotype of an individual’s health, defined by the quantification of social and behavioral phenotypes in situ from personal smartphone data.¹⁸

The individual phenotype could capture indicators of time-dependent relationships and thereby inform personalized and responsive mobile health interventions to prevent or disrupt downward trajectories after SCI. This is particularly important for individuals with SCI who are susceptible to frequent health complications and mobility restrictions that affect psychosocial wellbeing. Therefore, this methodological study set out to explore the feasibility of smartphone-based digital phenotyping methods to examine patterns of depression and how they relate to psychosocial wellbeing and physical activity after SCI. Participants were selected from either of two dynamic periods: the first was during a whole-body exercise intervention and the second followed newly-injured adults upon inpatient rehabilitation discharge. Exercise is a preferred treatment for depression among those with SCI ¹⁹ and often positively impacts mood and health.²⁰ In contrast, the transition to community living requires major life adjustments and psychosocial wellbeing tends to diminish, and issues return or worsen.^{1, 7} The powerful combination of self-report surveys with passively collected smartphone data may be uniquely suited to analyze psychosocial wellbeing during these two dynamic periods after SCI. This initial foray into digital phenotyping methods sought to determine if feasibility of acquiring these data is different in these two populations, considering their different lived experiences, and if psychosocial wellbeing trajectories can be characterized over a significant period.

Methods

This methodological study used the Beiwe research platform^{18, 21} for smartphone-based digital phenotyping to assess adults with SCI at either of two time points and clinical locations: 1) initiation of a novel whole-body exercise intervention (hybrid functional electrical stimulation row training; FESRT) at an outpatient hospital-based exercise program or 2) after initial SCI, at discharge from inpatient rehabilitation to a community dwelling. Recruitment began in March 2017, and the participants in this study were enrolled between April 2017 and July 2019. Eligible participants were adults (at least 18 years old) with SCI who used a wheelchair as their primary mobility device and communicated in English. Further, individuals had to demonstrate independent use of their own smartphone (Android or iOS) using hands or adapted devices. Exclusion criteria were SCI due to suicide attempt; all other etiologies were included. After providing informed written consent, the app was downloaded, and subjects demonstrated independent navigation of the app. This was approved by the Partners Healthcare Network Institutional Review Board (IRB). Use of the Beiwe app and platform was approved by the Partners Healthcare Research Information Security Officer. All data was encrypted and indirect identifiers were hashed using an industry standard hashing algorithm, rendering all data unidentifiable. Baseline demographic and clinical data was reported by participants and confirmed by medical records. Initially this study did not offer compensation to participants, however this was revised with IRB approval to promote study retention. After this revision, participants who completed at least 70% of surveys received a $30 check payment after each two-month period of enrollment. The IRB determined that, although this may be a factor in the decision to participate, the incentive would not unduly influence participation. To encourage data completeness, study staff contacted participants after two missing weeks of survey or smartphone sensor data.

Active Survey Data Measurements

Assessments occurred weekly in Beiwe, presented via a persistent notification. The Patient Health Questionnaire-8 (PHQ-8)²² was measured depression severity, modified to reference the previous 7 days. Frequency of experiencing each symptom of depression was noted (scored 0–3) and higher total scores indicated greater depression. Participants were also asked the extent that depressive symptoms interfered with activities or getting along with others. For safety, participants could access an emergency mental health clinician at any time through clicking a “Call Crisis Clinician” button on the app’s home screen. In the voluntary informed consent process, all participants were made aware that responses were not reviewed in real time and this was not an intervention study or substitute for any mental health services.

The SCI-Quality of Life (SCI-QOL) subscales were used, including Ability to Participate in Social Roles and Activities (SRA) and Satisfaction with SRA.²³ The former measures frequency of being able to engage in various social activities (e.g., “I can keep up with my family responsibilities”), and the latter asks for respondents’ level of agreement with statements about satisfaction or disappointment with those activities (e.g., “I am satisfied with my current level of activities with my friends”, “I am disappointed by my limitations in regular family activities”). SCI-QOL subscale scores are normalized to a T-score for comparison to the general and neurologic norm populations.

Participants were asked about health issues and mobility changes that could significantly impact social engagement or community mobility, indicating whether any following events occurred: hospitalization, new infection (i.e. respiratory or urinary tract), new musculoskeletal injury, new illness, and mobility equipment change or malfunction. Finally, participants reported frequency and intensity (mild, moderate, high) of physical activity performed in ≥10 minute bouts, modified from the Leisure Time Physical Activity Questionnaire for People with Spinal Cord Injury.²⁴ This excluded time in necessary physical activity such as therapy or propelling a wheelchair for transportation. Whenever possible, the number of required screen taps was restricted to facilitate usability among those with limited hand function. The surveys totaled 10–15 minutes per week and notifications were spaced throughout the week and allowed participants to answer at any time.

Passive Smartphone Sensor Data Measurement

The Beiwe platform was configured to sample GPS data for two-minute epochs spaced 18 minutes apart based on phone movement and usage. Given that physical structures can interfere with GPS signals, unreliable data (>20 meters) was discarded before smoothing with an accuracy weighted five-minute moving average. Smoothed longitudes and latitudes were projected to an x-y plane and passed through a density-based spatial clustering algorithm (DBSCAN) to separate locations visited from travel (i.e. noise). A location was defined as a minimum cluster of four points (i.e. where the subject spent at least twenty minutes) within 50 meters. For each twenty-four-hour period, locations visited were determined relative to the subject’s primary location: a person not leaving the house would have zero locations visited. In addition, each subject’s home was determined post-hoc as the location they spent the most time between midnight and four AM over the total data collection period. Exercise and medical care locations were also identified using Open Street Maps to allow quantification of time spent at these types of locations. Finally, rough estimates of distance travelled were calculated as the sum of straight lines between locations and the radius of the smallest circle that encompasses all locations visited. Summary statistics were created weekly.

Analyses

The first 16 weeks of data after exercise program enrollment or after hospital discharge were analyzed. Descriptive statistics examined the characteristics of the sample, study adherence, and distributions across all outcomes. For comparisons between groups, continuous variables were assessed with 2-sided t-tests and categorical variables were assessed with chi-square or Fisher’s exact tests. Differences over time within a group were analyzed with repeated measures analysis of variance and follow-up pairwise comparisons. Given the potential clinical utility of these methods for personalized medicine, correlations were also determined among variables within individuals. These were followed by linear regressions to examine the impact of time on depression. Finally, bivariate cross correlations determined the phase of significantly correlated variables. Analyses were performed in Stata 15.1 and MATLAB R2018b. This study used a sample size calculator developed by two of the co-authors (JT and JPO), available at https://onnela-lab.shinyapps.io/digital_phenotyping_sample_size_calculator/. To determine a modest correlation of depression with the seven other weekly measurements, this study assumed 50% missing data and alpha level 0.05, and data from 40 individuals would power this study at 99%. With 10 individuals and the same parameters, this study would be powered at 69%. This study followed the STROBE reporting guidelines for observational studies.

Results

Participants

Across 27 months of the ongoing study, 105 adult wheelchair users with SCI were approached to determine eligibility, and 43 were enrolled (Figure 1). The majority of those who did not enroll either anticipated walking to become their primary means of mobility prior to study end or had no interest in research participation. Additional reasons for not joining included insufficient hand function, concern for privacy, and not owning a personal smartphone. Among the 43 enrolled participants mean age was 40.0 years [16.4], and the majority were male (77%) and white or Caucasian race (77%; Table 1). Median SCI duration was 4.3 months and ranged from 1.2 months to 43.7 years. Neurologic level of injury (motor) ranged from C1 to T12; 30% had motor complete tetraplegia, 30% had incomplete tetraplegia, 17% had complete paraplegia, and 23% had incomplete paraplegia. Of those with high cervical SCI 5 were American Spinal Injury Association Impairment Scale (AIS) C, 2 were AIS A and 2 were AIS D, and all demonstrated independent use of their smartphones to research staff. Falls and non-traumatic causes of SCI were equally common (23% each), followed by sports/recreation (21%) and vehicular (19%; Table 1).

Figure 1.

Flow of participants through the study: Enrollment and study adherence

notes: Study enrollment, attrition, adherence, and completion by pre-post compensation years. Abbreviation: mo: months

Table 1.

Sample demographic and SCI characteristics

Characteristic	N=43
Age in years, mean (sd)	40 (16.4)
Gender, n (%)
• Male	33 (77)
• Female	10 (23)
Race, n (%)
• Caucasian, White	33 (77)
• African American	4 (9)
• Multiracial, Other	6 (14)
Duration of SCI in months, n (%)
• <3	9 (21)
• 3–6	15 (35)
• 6–12	6 (14)
• 12–24	4 (9)
• >24	9 (21)
Categorized Injury Level and Motor Completeness, n(%)
• Complete Tetraplegia	13 (30)
• Incomplete Tetraplegia	13 (30)
• Complete Paraplegia	7 (17)
• Incomplete Paraplegia	10 (23)
AIS, n (%)
• A	20 (46)
• B	6 (14)
• C	15 (35)
• D	2 (5)
Neurological Level of Injury, n (%)
• High Cervical (C1-C4)	9 (21)
• Low Cervical (C5-C8)	17 (39)
• High Thoracic (T1-T6)	8 (19)
• Low Thoracic (T7-T12)	9 (21)
SCI Etiology, n (%)
• Vehicular	8 (19)
• Violence	3 (7)
• Sports/Recreation	9 (21)
• Falls	10 (23)
• Pedestrian	3 (7)
• Non-Traumatic	10 (23)

abbreviation: AIS, American Spinal Injury Association Impairment Scale

Of the enrolled sample, 19 individuals were from FESRT and 24 individuals were from the community reintegration group (CRG). The only demographic or clinical differences between the FESRT and CRG participants were longer duration of SCI (6.5 years [2.5] vs. 2 months [0.2], p<0.05) and lower frequency of non-traumatic SCI etiology (5% vs. 42%, p<0.05) in the FESRT group.

Study attrition

Fourteen participants dropped out by two months and five dropped out between two and four months (Figure 1). In general, individuals who dropped out tended to be CRG rather than FESRT participants (63% vs 21%, p<0.05) and to be older (44.4 [17.1] vs 36.5 [15.2] years, p=0.12). Those who dropped out reported lower baseline social participation (39.9 [5.6] vs 45.9 [7.4], p<0.05) and less satisfaction with social participation (42.3 [4.7] vs. 48.5 [6.9], p=0.002). Those who dropped out also exercised less frequently at mild (2.6 [2.7] vs 4.3 [2.3] sessions, p<0.05), moderate (0.7 [1.3] vs 3.2 [2.4], p<0.05), and high intensities (0.3 [0.8] vs 1.9 [2.3], p<0.05). Neurological level of injury and baseline depression severity did not differ between those who dropped out and those who completed 4 months. Prior to providing compensation, reasons for study attrition included demands for time (reported six times), medical complications and limited hand function (each cited four times); six were lost to follow up despite multiple contact attempts. After providing compensation, study retention increased from 50% to 78% (Figure 1). During that time, one enrolled participant was lost to follow up and a second participant desired more privacy and withdrew from the study.

Data completion

Fifty-three percent of FESRT participants (n=10) and 21% of the CRG (n=5) completed 4-months of data. FESRT participants completed more surveys than CRG participants (57% vs. 25%, p<0.05). Participants with complete data (submitted at least 70% of surveys) had longer SCI duration (89.4 [149.8] vs. 7.1 [9.6] months, p=0.005), greater ability and satisfaction with social participation (46.6 [6.4] vs 41.5 [7.2], p=0.02; 43.9 [5.6] vs 49.2 [7.5], p=0.01), and more frequent physical activity (mild p=0.08, moderate p<0.05, and high intensity p<0.05). The majority (n=13) had depression severity within normal range (PHQ-8 <5), though 2 individuals had a more severe average depression (mild n=1, moderate n=1).

Preliminary findings by depression severity

Roughly half of participants with complete data had at least one week of mild depression during the four months, and this delineation was applied to examine the relationship between depression and psychosocial wellbeing across individuals (Depression vs. No Depression). The groups did not differ in demographic or clinical characteristics and had similar data completeness. Average depression severity in the depressed group appeared to decrease over time (Figure 2; Box’s conservative p=0.23) but may have been driven by substantial declines in three individuals. The pairwise difference in mean depression over time was largest between baseline and 8-weeks (p<0.05).

Figure 2.

Average depression severity over time

notes: Average weekly group PHQ-8 depression score with standard error bars. Though average depression declined from mild to minimal severity, the apparent decrease was likely driven by significant declines in three individuals.

Those with depression consistently had lower satisfaction with social engagement (42 [1.4] vs 47.7 [1.1], p<0.05) and a restricted ability to participate in SRA (51.3 [7.9] vs 39.1 [3.1], p<0.05). Social engagement ability seemed to improve over time for those without depression (Week 1, 48.0 [0.9] vs Week 16, 56.6 [2.9], p<0.05; Figure 3), but not in those with depression (Week one, 39.1 [1.2] vs Week 16, 39.1 [1.3], p=0.99). According to SCI-specific bookmarking-based scoring standards from Pamela Kisala, the depressed group experienced “Moderate Problems” (36.25 to 41.25), while their non-depressed peers had “No Problems” (>46.25) with social engagement. However, across participants there was variability in time-dependent correlations of depression with satisfaction or ability to participate in SRA: the in-phase Pearson correlation coefficients ranged from −0.57 to 0.56. The data also showed a variety of phasic differences by psychosocial indicator. For example, a clear decline in depression severity was coupled with improved satisfaction within the same week in one individual (r=−0.51, p=0.07; Figure 4), while improvements in social engagement ability were seen following a preceding two-week decline in depression (r=−0.75, p=0.002).

Figure 3.

Changes in social engagement by depression severity

notes: Average group depression scores with standard error bars. Though there were differences between depression in both of these comparisons, the only within-group change during the study was an increase in social engagement participation among those without depression. Abbreviations: SRA: social roles and activities.

Figure 4.

Course of depression and social engagement

notes: Participant example of within-individual variability in depression and corresponding changes in social engagement satisfaction and ability. A decline in depression severity corresponded with improved social engagement satisfaction in the same week and with improved social engagement ability after two weeks.

Average GPS-derived community mobility measurements were similar across depression groups; these metrics included time spent at exercise and medical locations, days out, time spent at home, total locations visited, and distance travelled. Relationships were typically in-phase, but also varied by individual. Two examples are shown in Figure 5. Subject #18 showed a steady decline in depression from mild to normal symptoms without any relationship to community mobility. In contrast, Subject #22 experienced a similar decline accompanied by an in-phase increase in weekly locations (r=−0.47, p=0.09).

Figure 5.

Depression and community mobility correlations over time

notes: Declines in depression severity, shown with trend lines, and corresponding community mobility metrics. Subject 18’s decline in depression did not relate to the weekly days out of the house. Subject 22’s decline in depression was inversely related to an increase in weekly locations visited.

Those with at least mild depression tended to report more health issues (4.4 [4.4] vs. 1.7 [1.9], p=0.14), co-occurring with reported infections (n=5), other illnesses (n=4), hospitalizations (n=2) and single muscle or bone injury (n=1). Surprisingly, the depressed group tended to report more frequent light physical activity (4.4 [1.6] vs. 2.6 [1.9] weekly sessions, p=0.06). However, this was not uniform across all individuals or exercise intensities. Figure 6, for example, illustrates the phase of a moderate association of increased depression in the past week with less high intensity exercise in the current week.

Figure 6.

Phasic relationship of depression with high intensity physical activity

notes: Participant example of phasic relations calculated with cross-correlogram for bivariate time series: Moderate correlation between higher depression in the previous week with decreased frequency of exercise sessions in the current week.

Discussion

This methodological study determined that it is feasible, but not without challenges, to acquire a unique combination of self-report survey data and passively collected smartphone data during a whole-body exercise intervention or during community reintegration after initial SCI. This study expanded digital phenotyping methods to a sample of wheelchair-users with SCI and provided preliminary analyses of time-dependent relationships between depression and indicators of psychosocial wellbeing. With an unprecedented level of detail, digital phenotyping methods identified large individual variability in these relationships. Preliminary findings suggested that depression related to poor social engagement but had no consistent relationship over time with GPS measures of community mobility.

Study attrition did not differ by SCI neurological level, indicating manageable burden for touchscreen interactions even among those with partial hand function. Still, accessibility and utility of mobile monitoring might also be increased with speech-to-text options for survey responses.²⁵ Although participants remained in the study longer, the conditional remuneration did not apparently incentivize survey response behaviors because percentage of complete surveys did not increase. The financial incentive did not appear to alter recruitment appeal to eligible participants: the sample’s demographic characteristics were similar pre and post compensation.

CRG participants were less engaged in the study and were also more severely depressed than the FESRT participants. Newly-injured individuals with SCI may have less clinical and peer support, resulting in many demonstrating overt depression during community life despite no depression during inpatient stay.⁷ Furthermore, depression severity and coping strategies after acute SCI can impact psychosocial wellbeing.²⁶ The three individuals who had significant declines in depression severity were in the FESRT group. Decreases in depression scores have been noted in other monitoring studies with a potential mechanism of action including increased emotional self-awareness.²⁷ These group-level differences emphasize the difficulty in assessing individuals over longer periods of time after acute SCI.

Data from this 4-month methodological study suggested that participants with even one week of depression had significantly restricted ability and satisfaction with social engagement compared to those without depression, to individuals with neurological conditions in general, and to able-bodied adults. To further quantify patterns in psychosocial wellbeing, study authors have begun to analyze additional smartphone sensor data including phone use (call and text logs) as a behavioral representation of social engagement.²¹ Additionally, this ongoing study has incorporated the Personal Network Survey (developed by Dhand and colleagues in 2018) to measure changes in the structure and health of the network and type of personal ties.

Smartphone sensors and in-situ surveys have been used separately to examine indicators of psychosocial wellbeing ^{28, 29} and even to prompt self-monitoring interventions among those with SCI.¹³ To establish more comprehensive, personalized models of psychosocial wellbeing, the most powerful solution would be a combination, as introduced in ambulatory adults to enable prediction of depression severity and guide intervention delivery.^{30, 31} The individual variability in depression and indicators of psychosocial wellbeing identified in this study suggest digital phenotyping methods could more completely explain psychosocial wellbeing trajectories than current standard measurements. This would allow personalized healthcare decisions that reference the individual-level phenotype rather than relying group normative data for more effective and timely interventions. Moreover, computer modeling brings scalability to enhance clinical utility of digital phenotyping methods.

Study Limitations

This study did not collect clinical, demographic, or socioeconomic data on individuals who were approached by study staff and were found to be either ineligible or not interested in research. It is possible that individuals from marginalized communities experience greater surveillance in public situations and may be more likely to avoid a research study that collects GPS data. The SCI sample was less racially diverse than a study in able-bodied individuals that also combined psychosocial survey data with GPS¹⁶ but slightly more diverse than the composition of the Boston metro area, and this study had successful data collection and retention across races. Though age was not different by data completion, study participants who dropped out were slightly older. In spite of potential generational differences in smartphone technology familiarity, individuals from young to older adulthood enrolled in this study and overall this sample was older than those in a majority of similar studies.¹⁵ Attrition in this study suggests that the unique demands of a new SCI contributed to drop out. Greater study attrition and fewer survey responses among those with acute SCI may indicate needed restructuring of the app interface. One of the most common reasons for drop out was time constraints and so monitoring psychosocial wellbeing via passive smartphone sensor data that does not require participant input may be warranted. Additionally, study engagement could be increased by selecting items to answer via branching logic based on current mood. Finally, imputation methods for within-individual GPS processing could decrease error variance due to missing data.

Conclusion

Smartphone-based sensors combined with in-situ survey methods are feasible monitors of psychosocial wellbeing among wheelchair users with SCI. Higher attrition among acute SCI suggests a need for even less obtrusive smartphone platform. This new approach to traditional psychosocial health measurement techniques after SCI revealed great variability in depression and its relationship to social engagement, community mobility, health, and physical activity. Models of psychosocial wellbeing specific to the individual’s behavior and in-situ self-report would be highly relevant to a personalized medicine approach to patient care.

What is Known

Previous approaches characterized changes in psychosocial distress after SCI from questionnaire-based data at discrete time points three to six to twelve months apart. These data cannot predict trajectories that would allow preventive interventions to be delivered when most needed.

What is New

Smartphone research platforms are uniquely equipped to collect in-situ self-report and objective behavioral indicators of psychosocial wellbeing. A preliminary exploration of this fine-grained assessment found that even one week of depression over 4 months is related to worse health and social engagement. These methods may hold clinical utility to predict psychosocial distress development or resolution and inform timely interventions after SCI.