This study is a systematic review and meta-analysis. All procedures have been preregistered in the open science framework under trial registration link. Reporting is conducted in accordance with the updated PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines for the reporting of systematic reviews [40] (Multimedia Appendix 1).

Given the interdisciplinary nature of the field, we searched MEDLINE, PsycINFO, Embase, CENTRAL, ACM, IEEE Xplore, PubMed, and Web of Science. The database-specific search strings can be found in Multimedia Appendix 2. After the removal of duplicates by automatic tools of the databases, all identified articles were screened in a two-stage process: (1) title and abstract screening and (2) full-text screening, both performed by 2 independent reviewers (YT and JK). If duplicates were encountered in this screening process, they were removed by hand. In addition to the database search, we searched all reference lists of the included studies for further eligible studies. This search and inclusion process was started on December 6, 2022 (database searches from December 06 until June 13), and completed by March 24, 2023 (last reference search). Any disagreements between the reviewers were resolved in discussions.

To be eligible, the following inclusion criteria had to be met: (1) any kind of GPS sensor data collection, (2) the data were collected by a wearable device such as smartphone or smartwatch, (3) an assessment of depressive symptoms was conducted either by self-report or by clinical diagnostic scales, and (4) reported outcomes included correlations between the collected GPS sensor data and depressive symptoms. Studies that (1) comprised participants younger than 18 years of age or (2) included participants with disorders other than depression (eg, bipolar disorder) were excluded.

All data were extracted by 2 independent reviewers (JK and YT). The following study characteristics and empirical data points were obtained.

We identified a total of k=9499 unique records in the systematic literature search. One additional study was screened and included after being forwarded by researchers contacted due to insufficient data reported in an identified study. In the following screening and inclusion process conducted by 2 independent reviewers (YT and JK), k=19 studies were finally included [24,34-37,49-62]. In total, 2 of the studies were eligible for between- and within-person correlation meta-analysis [34,53]. For further details, please refer to the PRIMSA flowchart in Figure 1.

Combined, the included studies comprised a total of N=2930 participants. The sample sizes ranged from a minimum of n=18 to a maximum of n=1046 (mean 154.21, SD 235.54; median 72, IQR 83). The mean age across all included studies was 38.42 (SD 18.96) years. The average percentage of female participants was mean 59.64% (SD 22.99%). Of the k=19 included studies, k=8 (42%) studies explicitly targeted students and young adults (eg, recruitment at universities), k=3 (16%) studies recruited from the general public, k=5 (26%) studies used nontailored online and social media recruitment, k=2 (11%) studies targeted older adults, and k=1 (5%) study provided no further information on the recruitment strategies and targeted population. Please refer to Table 1 for further details on the descriptive characteristics of the studies.

A total of k=13 unique sensing frameworks were applied among the included studies. PurpleRobot was the most frequently used framework (k=3/19, 16%). Based on the sensed GPS data, the studies reported n=12 distinct activity and mobility features. The number of unique studies per activity and mobility feature ranged from k=1 to k=14 and a combined sample size of n=69 to n=2287 per feature. A summary of the features can be found in Table 2 for between-person features and Table 3 for within-person features.

Distance was the mobility and activity feature most strongly associated with depressive symptoms, with a meta-analytically pooled between-person correlation of r=–0.25 (95% CI –0.29 to –0.21), followed by normalized entropy (r=–0.17, 95% CI –0.29 to –0.04), location variance (r=–0.17, 95% CI –0.26 to –0.06), entropy (r=–0.13, 95% CI –0.23 to –0.04), number of clusters (r=–0.11, 95% CI –0.18 to 0.03), and home stay (r=0.10, 95% CI 0.00 to 0.19). In contrast, the features circadian movement, transition time, speed moving and time moving indicate no significant correlations with depression. Table 4 summarizes all of the pooled between-person correlations. Feature-specific forest plots can be found in Multimedia Appendix 5.

The 3 identified studies reporting on within-person correlations (Table 3) differed widely in the applied methodology of analysis. Saeb et al [34] correlated GPS mobility and activity features with the change in PHQ-9 scores, while Zhang et al [53] used autoregressive models to estimate the correlations over time, and Canzian and Musolesi [57] applied time series analysis to investigate the correlations within each participant across multiple days. Due to this heterogeneity in these studies to derive the reported correlations, we did not perform a meta-analysis for within-person correlations.

All included studies (k=19) followed an observational study design. Assessment of the international reporting standards for observational studies revealed that not a single study referenced the international reporting guidelines STROBE. Comparing the k=18 included published studies against the STROBE checklist showed an overall agreement across all items of 73.84% (Stamatis et al [61] excluded from comparison as the provided data originated from additional analyses not covered in that study). Only k=9 (50%) of the studies reported how potential sources of bias were addressed. Regarding missing data, k=6 (33%) reported how missing data was handled and k=3 (17%) reported rates of missingness for the variables of interest. A total of k=7 (39%) of the studies listed reasons for the exclusion of participants at each stage of the study. For all STROBE ratings, please refer to Multimedia Appendix 4.

In addition, neither preregistrations nor study protocols were found for the included studies. A priori power analyses for sample size planning or post hoc for discussion were also not reported in any of the included studies. Power analyses of the included studies indicate that k=16 (84%) studies were sufficiently powered with a power of 80% to detect a correlation of r=.50, k=7 (37%) studies for r=.30, k=4 (21%) studies for r=.20, and k=1 (5%) studies for r=.10.

For all mobility and activity features except for time moving, Egger test showed a significant asymmetry in the funnel plot (Table 5). Funnel plots for features with less than k=10 studies, in addition indicate asymmetry. Please refer to Multimedia Appendix 6 for the funnel plots of all between-person features.

This systematic review with meta-analysis provides pooled correlation estimates for the between-person correlation of GPS-based smart sensing mobility and activity features and depression. We identified robust small to medium between-person correlations for multiple mobility and activity features (ie, distance, normalized entropy, entropy, homestay, number of clusters, and homestay). In contrast to the findings on between-person correlations, this literature for within-person correlations did not allow for meaningful meta-analyses. Furthermore, this study clearly highlights a lack of quality in the literature. While all studies followed an observational study design, not a single study referenced the international reporting guidelines for observational studies (STROBE [38]). Although the reporting in studies addressed STROBE items in most cases (14/19, 74%), deficits were found in addressing sources of bias and missing data. Moreover, we found strong evidence for small study effects potentially indicating publication bias. Most studies were underpowered to detect correlations of the magnitude indicated by the meta-analysis.

Overall, the present analysis highlights the potential of GPS features to infer depression, which is in line with previous published descriptive reviews in the field [26,27,30,32,33]. For instance, with a total sample size of N=1824 and homogenous findings across studies (I²=0%), distance (r=–0.25) might be a promising objective and unobtrusively collectible GPS feature in expert systems for (assisted) diagnoses, just-in-time interventions and personalization of treatment of depression in future [20,65]. Analogously, normalized entropy, entropy, homestay, number of clusters, and homestay were significant markers (r=0.10-0.17). However, it is important to note that all studies followed an exploratory design. Both the correlations and the clinical feasibility of such smart sensing augmented expert systems need to be proven in confirmatory trials before clinical applications. The identified pooled correlations and their CI offer a strong foundation for a-priori power analysis to guide confirmatory studies in their study and sample planning [66,67].

However, while some of the investigated correlations are statistically significant and pooled estimates can serve as a basis for future study planning from a statistical point, this does not release from the question of what a clinically relevant correlation is. The process involved in obtaining GPS features (eg, collection of GPS coordinates) as well as the features itself (eg, how much time was spent at home or at work) are highly sensitive. Hence, a discussion taking the clinical significance as well as ethical and privacy issues into account is of utmost importance [20,21,65,68-70]. Furthermore, it needs to be discussed to which extent GPS features as a standalone metric can be used in the diagnostic process or rather as an add on to existing procedures and measurements (eg, patient-reported outcome, ecological momentary assessment, medical record data) [20,24,71]. This discussion would strongly benefit from future studies illuminating these questions from patients’ and health care providers’ (eg, psychotherapists) perspectives.

Another question in the field is, whether GPS features cannot only serve to determine depression from a between-person perspective but also to inform researchers and clinicians about changes within persons (eg, is the increase of time spent at home a reliable indicator for rising depression severity?). This systematic review showed too sparse evidence and too high heterogeneity in applied analyses to run a meta-analysis on the studies conducted so far. Although this study cannot provide answers regarding the robustness across studies and hence evidence for the significance of within-person correlations, in particular, the included study by Zhang et al [53] on the longitudinal relationship of GPS mobility and activity features and depression found higher within-person correlations than between person correlations, underlining the potential benefit of within-person features in clinical applications and the importance of further studies on within-person correlations.

In line with previous reviews on the methods used in smart sensing studies [26,27,31], this study critically highlights gaps in the adherence to international reporting guidelines, potential publication bias, and high heterogeneity in assessment and analysis methodology. Guidelines like STROBE [38] for observational studies and other study design–specific guidelines offer a strong starting point to increase the study and especially reporting quality in the field. However, research on smart sensing and objective sensor data collection comes with its own challenges like (1) high heterogeneity in hardware (eg, used devices) and GPS sampling (eg, many different manufacturers of sensors and variety in the precision of sensing and data quality), (2) a plethora of data preprocessing decisions (eg, sampling rate, outliner detection or feature selection and definition), (3) missing data handling in large fine-grained datasets, and (4) analyses methodology (eg, how to deal with dependencies resulting from multiple measurements from the same individual [24]) [31]. Preregistration and open-access scripts for data preparation and analyses and extensions to existing international guidelines (eg, standards for feature calculation, reporting guidelines, and missing data handling) are highly needed to move towards a standardized and reliable research field.

The current lack of standardization and heterogeneity between studies in assessment frameworks, samples (eg, age and range), sensor sampling rates, devices, feature calculation, methods of handling missing data or accounting for dependencies in the data structure should also be considered when interpreting these findings. Due to the already low number of studies, we were unable to conduct meaningful sensitivity analysis on more homogenous groups of studies (eg, based on data quality, and devices) or meta-regression models to control for other variables (eg, for systematic mobility and activity differences across age groups). As the number of studies and their quality are likely to increase over time, analyses on homogenous groups and meta-regression analyses might become feasible in future. Nevertheless, we want to point out that the heterogeneity for some features (eg, distance) was very low, underlining the robustness of the pooled correlation estimates for these features. Moreover, the weighted pooling in the meta-analysis gave stronger influence on larger studies with more precise estimations of correlations, counteracting the influence of small studies to some extent [41].

Besides analyses on GPS mobility and activity features in more homogenous study groups, other GPS features should be focused on in the future. For instance, GPS sensor data can be used to derive environmental features such as green space, blue space, air pollution, or even regional data like social deprivation scores (eg, zip code based). First studies using smartphone-based GPS collection to infer depression based on such environmental GPS features are promising, and it needs to be investigated to what extent such features show meta-analytically robust correlations [72-74]. Analogously, meta-analytical evidence for other sensors and features is lacking (eg, smartphone screen usage, app usage, and communication and sociability features) and of high importance to inform researchers and practitioners on which objective markers might be suited for depression diagnosis, early-warning, just-in-time interventions, and other clinical applications [21,30-32,65,69,75]. Furthermore, extensions to other prevalent mental disorders (eg, anxiety) and conditions of leading disease burden (eg, pain) are highly needed to investigate the potential of unobtrusive and objective smart sensing data in health care.

This systematic review and meta-analysis provides evidence for small to medium between-person correlations of GPS mobility and activity features collected through smartphones and depression. In the future, GPS mobility and activity features such as distance may become an important augmentation in the assessment of depression and clinical applications (eg, decision support systems). However, replications in confirmatory studies and improvements in the study quality and standards in the field are needed to draw robust conclusions. Besides, more research on the within-person correlations is necessary to determine the potential of GPS features in not only between-person applications but also from a longitudinal perspective. To fully exploit the potential of smart sensing, further research on associations of depression and other mental disorders with GPS mobility and activity features as well as environmental GPS features (eg, green space) and other sensor modalities (eg, smartphone usage features), and how smart sensing features can be integrated in existing information systems and complex prediction models alongside patient-reported outcomes, clinician ratings, ecological momentary assessments, and medical record data, is of high importance.