Behaviors such as levels of activity and social engagement are associated with common mental health conditions such as depression and anxiety [1-4]. Retrospective evaluations of these behaviors through self-report measures or interviews rely heavily on retrospective recall, which is subject to systematic biases [5,6]. Even more accurate methods, such as ecological momentary assessment (EMA), which acquire self-reported experiences in the course of peoples’ lives [7,8], have not proven practical over extended periods outside of research settings [7]. Accelerometry data from wearable devices that measure activity levels have also been associated with depression [9], but many people stop wearing the devices within the first weeks [10].

Smartphones are becoming ubiquitous. As of 2019, 81% of Americans owned a smartphone [11], as did 76% of people in countries with advanced economies, and 45% in emerging nations [12]. Smartphones are fully integrated into our lives, supporting a growing number of activities. Smartphones contain embedded networked sensors that provide continuous, objective data without user effort, which can be used to produce behavioral markers. A growing body of research suggests that these sensor data can be associated with common mental health problems [13]. Location features derived from GPS actual measurement of patterns of locations visited, time in locations, and travel in phone communications such as text messaging and phone and app use have been associated with depression, anxiety, and social anxiety [14-17].

The potential for personal mobile sensing to improve our understanding of the relationship between behavior and mental health, as well as to advance clinical care, has been widely recognized [18]. However, although there is promise, research to date has had a number of weaknesses. Many of these studies have been conducted in small, relatively homogenous groups, such as students [15,16,19-21]. Although there has been some specificity, with location features tending to identify depression [14,15,19,22,23] and communication features tending to predict social anxiety [17], there have also been a few studies that have found the opposite [16,24]. This may be because of the heterogeneity in symptoms and comorbidities [25], which are common and have not been considered in existing research [13].

To date, studies have focused on using sensed behavioral features to estimate a person’s state, either the presence or absence of a condition or symptom severity. With some exceptions, they have generally not evaluated the capacity for sensed behavioral features to predict whether symptoms will increase or decrease in the future. Among the few studies that have examined the capacity of sensed features to predict symptom change, one small study of 18 patients with bipolar disorder found that greater inconsistencies in rates of typing on a smartphone keyboard were related to future greater depressive symptom severity [26]. Relative to studies that use sensed behavior to estimate a person’s state, the temporal relationship between sensed behaviors and symptom change has received relatively little attention.

In this study, we examine the temporal relationship between changes in sensor features and subsequent changes in mental health symptoms in a large sample of participants. The aim of this exploratory study is to evaluate whether changes in classes of smartphone sensor features were associated with changes in symptom severity for depression, anxiety, and social anxiety, across all participants as well as within groups clustered based on symptoms.

Participants were recruited from July 15 to July 26, 2019, through ResearchMatch, a National Institute of Health-funded volunteer network, and the Center for Behavioral Intervention Technologies research registry. Participants were included if they were US citizens and residents, age ≥18 years, could read English, and had an Android smartphone. Participants were excluded if they endorsed, via self-report, having been diagnosed with a severe mental illness, defined as bipolar disorder, schizophrenia, or other psychotic disorder. Participants were compensated for completing measures at set assessment points as well as for completing EMA check-ins. Compensation for completing assessments increased as the study period progressed, such that participants were compensated relatively less for early assessment points and relatively more for assessments toward the end of the study. No single assessment was compensated at more than US $32.50 per assessment time point. Recruitment was advertised as a study on depression and deliberately oversampled depressed participants such that at least 50% of the sample experienced at least moderate depression symptom severity according to the 8-item Patient Health Questionnaire (PHQ-8).

Participants downloaded the Passive Data Kit [27] mobile app, which unobtrusively collects phone sensor data and administers surveys. Web-based questionnaires were administered every 3 weeks. Participants were enrolled in the study for 16 weeks. All procedures were approved by the Northwestern University Institutional Review Board, and informed consent was obtained from all participants before participation.

Participants completed web-based symptom severity assessments at baseline and every 3 weeks until the end of the study period (ie, weeks 4, 7, 10, 13, and 16). Symptom measures included depression severity (PHQ-8) [28], generalized anxiety disorder (Generalized Anxiety Disorder 7-item scale [GAD-7]) [29], and social anxiety disorder (Social Phobia Inventory [SPIN]) [30]. The PHQ-8 was administered as an EMA survey after the baseline assessment point and, subsequently, had one additional assessment point (week 1) relative to other symptom measures.

Phone sensor data included GPS coordinates sampled once every 5 minutes, communication information (ie, phone logs and duration, text message logs, and length), and open apps. Assessment weeks occurred every 3 weeks, during which participants were asked each evening to label the semantic location (type of location) that they had visited for more than 10 minutes [31]. A series of maps identifying each location were presented, and participants selected the category of each place (eg, home, work, errand, entertainment, place of worship, etc).

The flow of participants in this study is shown in Figure 2. Participant demographic and baseline symptom severity characteristics across the entire sample and participant clusters are detailed in Multimedia Appendix 3.

Ordinary least square regression revealed no significant change in symptom severity as a function of time (PHQ-8: P=.80; GAD-7: P=.83; SPIN: P=.57). However, there was substantial within-participant variability depending on the symptom measure, with mean SDs of 2.66, 3.50, and 5.90, for the PHQ-8, GAD-7, and SPIN, respectively.

There were no significant correlations between changes in any symptom severity measures and subsequent changes in sensor-derived behavioral features.

Across all treatment weeks, missingness (ie, the number of missing observations/total number of possible observations for all 282 participants) was higher for the PHQ-8 (277/1692, 16.37%) assessments than for the GAD-7 (104/1410, 7.38%) and SPIN (104/1410, 7.38%). PHQ-8 assessments were administered through our smartphone app, whereas the GAD-7 and SPIN were administered through REDCap (Research Electronic Data Capture) [39]. In addition, across symptom outcome measures and relative to baseline symptom levels (Multimedia Appendix 3), those with missing assessments tended to have slightly higher baseline symptom severity (PHQ-8 range: 10.94-12.82; GAD-7 range: 9.95-11.5; and SPIN range: 24.52-27.48).

Changes in numerous phone sensor–derived behavioral features were associated with subsequent changes in mental health symptoms among people with elevated symptoms of depression. However, changes in symptoms were not associated with subsequent changes in behavioral features. GPS location features were fairly consistently and negatively, albeit modestly, related to subsequent changes in depression severity across symptom groups. This is consistent with a number of previous relatively small studies showing correlations between GPS features and depression [14,19,20,22,40,41]. This larger study confirms these earlier findings, and importantly, indicates a directional relationship in which GPS features are associated with subsequent increases or decreases in depressive symptoms, but not with anxiety or social anxiety.

The types of locations (work, shopping, etc) people visited were less consistently related to changes in depression. This does not necessarily mean that specific locations are unimportant at the individual level: one person may like shopping, whereas another may detest it. However, this suggests that patterns of movement through geographic space, irrespective of the destinations or locations to which one travels, are indicators of symptom change among people with depression. We speculate that this may reflect a loss of motivation expressed through geographic movement. Perhaps more speculatively, it may also be that changes in neurocognitive mechanisms, such as executive control, affect, and motivation, impact both depression and basic mechanisms involved in movement through geographic space [42,43].

The different constellations of symptoms that participants experience impact the salience of some sensed behaviors in predicting outcomes. For example, only within participants in the Depression and Social Anxiety symptom cluster was the amount of time spent in locations related to social activities associated with (at the trend level) subsequent change in depressive symptoms, suggesting that although locations are generally useful for depression prediction, social activities may be particularly useful when social anxiety symptoms are present. Active App Use (texting, email, and mapping) was associated with depression change among those with multiple and more severe comorbidities. Although GPS features were generally useful, features integrating time and location were not useful among those with comorbid generalized anxiety, and features measuring transitioning between locations were not useful for those with comorbid social anxiety. Thus, there was support for the notion that symptom constellations are important considerations for some features.

Increases in telephone calls were associated with increases in social anxiety symptoms among clusters characterized by depression and social anxiety. This finding notwithstanding, the capacity for sensor-derived features to be associated with changes in social anxiety symptoms was not as consistent as that for depressive symptoms. Furthermore, we did not find any associations between sensor-derived features and generalized anxiety symptom changes.

These findings indicate that sensor-derived behavioral features, which are objective and can be acquired with reduced participant effort, can be a useful tool for investigating the role of some behaviors in changing depressive symptom severity. There has been much speculation about the clinical potential of mobile sensing [18,44]. The effect sizes are modest, albeit consistent with many other studies that have examined the use of sensed behavioral features to estimate the presence or severity of symptoms [16,19]. This study does not support the use of phone sensor data alone in monitoring symptom changes in mental health populations; however, phone sensor data may be useful in conjunction with other networked sensing tools such as wearables. Phone sensor data may be useful for digital mental health interventions [45]. Just-in-time adaptive interventions [46,47] use individualized data to predict risk and deliver context-aware intervention material that is adaptive. These approaches are increasingly applied in digital health interventions, such as identifying when to send messages to increase physical activity such as step counts [48]. The promise of delivering motivational messages at opportune moments that reinforce behavioral activation strategies, such as visiting someplace new, spending more time outside of the home or work, or engaging socially, has the potential to improve engagement with these tools and reduce depression.

This study had several limitations. First, the exploratory nature of this study requires interpreting results with caution and necessitates that future work must explicitly test the a priori hypotheses arising from these results. Next, although we establish significant temporal relationships between sensed behavior changes and subsequent changes in symptom severity, our study is not experimental and does not establish causal relationships. Furthermore, our sensor feature aggregations were limited to single sensor sources and were constructed to maximize interpretability; however, future studies that use data-driven aggregations are necessary to help inform feature aggregations across sensor modalities. Although aggregating across sensor modalities presents a challenge for interpretability, future work that examines cross-sensor aggregations could yield improved estimation of sensed behaviors and, subsequently, more robust associations with changes in symptoms. In addition, although these findings provide some support for the hypothesis that sensed behavior change is associated with subsequent changes in depression and not vice versa, this study examined the associations between changes in sensor features and subsequent changes in symptom severity measures lagged by 2 weeks, and therefore should not be generalized to periods beyond the 2-week window. Another limitation is that our sample only included those who used Android devices and agreed to participate in this research. App use and communication data are not readily available for iOS devices. Regarding data missingness, across all the surveys, individuals who had missing data had higher baseline symptom severity than the overall sample, though not dramatically so; thus, data were missing not at random. These missingness rates are in line with established criteria that are often used as the standard for good trial data [49]. Finally, although we controlled for multiple analyses, we nonetheless caution against overinterpretation of more isolated findings that need to be replicated in future studies.

The ubiquity of smartphones with networked sensors has opened up new opportunities to identify behavioral markers related to mental health that can be acquired continuously and effortlessly. Changes in movement through geographic space were consistently associated with subsequent changes in depressive symptoms; however, there was no evidence that changes in depression were associated with subsequent changes in sensed behaviors. This supports a directional relationship in which changes in movement patterns precede symptom change, but symptom change does not precede changes in movement behaviors.