The protocol was prospectively registered in the PROSPERO database (CRD42024518136) and prepared following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) protocol checklist (Checklist 1).

This systematic review was exempted by the Institutional Review Board (IRB) of Gachon University (IRB number: 1044396‐202403 HR-052-01).

The inclusion criteria were as follows: (1) inclusion of participants with depression, (2) use of digital biomarkers to assess depression severity, and (3) reporting diagnostic concordance with validated assessment tools. The exclusion criteria were as follows: (1) non-English or non-Korean articles, (2) duplication, (3) inaccessible full text, and (4) reviews or qualitative studies.

This systematic review involved a search of major academic databases, including the Cochrane Library (Wiley), MEDLINE (Ovid), PsycINFO (Ovid), CINAHL (EBSCOhost), IEEE Xplore (IEEE), Web of Science (Clarivate), and Korean academic databases, such as KISS, RISS, KMbase, and KoreaMed. The final search of all sources was conducted on December 28, 2025. Trial registries, gray literature, and author contacts were omitted as the primary search provided sufficient data. Additionally, we manually screened reference lists and removed duplicates using EndNote 20.

The search strategy combined Medical Subject Headings (MeSH) and free-text keywords related to depression and digital biomarkers. These terms were adapted for each database to maximize search sensitivity. The key terms included “depressi*,” “MDD,” “wearable,” “application,” “smartwatch,” “biomarker*,” “sleep*,” “speech,” “behavioral parameter*,” “electroencephalogram,” and “electrocardiogram.” The search process followed the PRISMA Search Strategy (PRISMA-S) extension [15]. The full search strategy, including specific search strings, limits applied, and the number of records retrieved per database, is provided in Multimedia Appendix 1.

Two independent reviewers screened the titles and abstracts after removing duplicates. Potentially relevant studies and manually identified records underwent full-text assessment. Any disagreements were resolved through consensus with a third reviewer.

Data were extracted by 2 independent reviewers using standardized forms. Extracted characteristics included author, year, country, study design, number of participants, age, sex, biomarker measurement device, measurement period, depression indicators, and analytical methods.

Quality was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool to evaluate diagnostic accuracy and the Scottish Intercollegiate Guidelines Network (SIGN) tool for case-control studies. All eligible studies were included irrespective of their quality scores.

A meta-analysis was performed on quantitatively synthesizable digital biomarkers, and a systematic review was conducted on other biomarkers. All statistical analyses were performed in R software (version 4.5.0; R Foundation for Statistical Computing) using the “meta” and “pimeta” packages. Pooled effects, expressed as mean differences (MDs) or standardized mean differences (SMDs), were estimated using a random-effects model with the Hartung-Knapp-Sidik-Jonkman (HKSJ) method [16,17]. Restricted maximum likelihood was used for τ² estimation. Results have been presented with 95% CIs. To account for the distribution of true effects across different settings, 95% prediction intervals (PIs) were additionally calculated using the parametric bootstrap approach proposed by Nagashima et al [18], which is robust for small study numbers.

Parameters that could not be meta-analyzed due to reporting inconsistencies or insufficient data were narratively synthesized. These included specific sleep measures (eg, sleep fragmentation, rapid eye movement [REM] sleep, REM latency, and slow-wave sleep [SWS]); physical activity (eg, light physical activity [LPA] and energy expenditure); cardiac HRV indices (eg, the root mean square of successive differences [RMSSD], low frequency [LF], high frequency [HF], and LF/HF ratio); and parameters related to speech, GPS, and circadian rhythms.

Reporting bias was considered qualitatively, as a formal statistical assessment (eg, funnel plot) was infeasible due to fewer than 10 studies per outcome. This was supported by comprehensive database searches and manual screening of reference lists. The certainty of evidence was assessed using the GRADE (Grading of Recommendations, Assessment, Development, and Evaluation) approach, considering risk of bias, inconsistency, imprecision, and indirectness [19].

The initial search yielded 39,617 studies. After removing duplicates and excluding records automatically identified as ineligible through journal-type filtering, 21,915 studies remained. Following the screening of titles and abstracts, 21,649 studies were excluded. Before conducting a full-text review, a manual search through reference lists and citation tracking of relevant systematic reviews identified 17 additional studies. Consequently, full-text reviews were conducted on 283 articles, of which 132 studies involving 57,852 participants met the inclusion criteria (Figure 1).

The characteristics of the included studies are summarized in Tables S1 and S2 in Multimedia Appendix 2. The included studies were categorized based on the types of digital biomarkers investigated. A total of 87 studies used single parameters, including sleep parameters (n=23) [20-42], cardiac parameters (n=19) [43-61], physical activity parameters (n=16) [62-77], smartphone-based parameters (n=9) [78-86], speech parameters (n=10) [87-96], circadian rhythm parameters (n=7) [97-103], electroencephalogram parameters (n=2) [104,105], and video-based parameters (n=1) [106]. Furthermore, 45 studies used multiple parameters to assess digital biomarkers [4,107-150]. Of the 132 studies, 63 (47.7%) used wearable devices to continuously measure biomarkers in daily life, typically for more than a week. Study designs encompassed large-scale cohorts and group comparisons (eg, depression vs control). Analytical approaches varied, including regression analyses for associations, evaluations of diagnostic accuracy, and assessments of group differences (Multimedia Appendix 3).

Validated depression assessment tools were used, with the most frequent being the Patient Health Questionnaire (PHQ), followed by the Hamilton Depression Rating Scale (HDRS), Beck Depression Inventory (BDI), Geriatric Depression Scale (GDS), Korean version of the GDS (SGDS-K), and Center for Epidemiologic Studies Depression Scale (CES-D) (Multimedia Appendix 4). Biomarker categories and their detailed features are illustrated in Multimedia Appendix 5 and Table S3 in Multimedia Appendix 2.

Based on the SIGN assessment, all 73 case-control studies demonstrated acceptable internal validity with clearly defined groups. However, participation rates for each group showed substantial variability, ranging from 20.1% to 100%, and many studies (70/73, 96%) did not report comparisons between participants and nonparticipants. Additionally, potential confounding factors were insufficiently addressed in several studies (Table S4 in Multimedia Appendix 2).

Among diagnostic accuracy studies assessed using QUADAS-2, most studies (58/59, 98%) showed a low risk of bias in the index test, reference standard, and flow and timing domains. In contrast, great concern regarding applicability was identified because study populations often did not align with intended clinical targets, limiting the generalizability of the findings to real-world settings (Table 1).

Meta-analysis results are presented in Figure 2. A detailed overview of the parameters that could not be quantitatively synthesized and their reported associations with depression is provided in Multimedia Appendix 6.

Speech parameters were categorized into speech flow and voice acoustic parameters.

The qualitative assessment of reporting bias suggested a low likelihood of missing relevant studies, supported by comprehensive multidatabase searches and manual reference screening. According to the GRADE approach, the certainty of evidence ranged from low to very low across the key digital biomarkers (Table 2). The certainty was low for SOL, TIB, and physical activity counts, whereas it was very low for TST, SE, WASO, and mHR. The overall certainty was mainly downgraded due to inconsistency across studies and imprecision associated with wide CIs or PIs.

This systematic review synthesized digital biomarkers for depression across diverse domains—sleep, physical activity, cardiac parameters, smartphone usage, speech, GPS data, and circadian rhythms—to identify more consistent indicators across multiple digital signals. Our meta-analysis identified prolonged SOL, increased TIB, and reduced activity counts as the most consistent behavioral features associated with depression. These findings support the hypothesis that digital phenotyping can capture objective manifestations of depression, particularly sleep initiation difficulties and reduced energy expenditure, which are often difficult to quantify through traditional self-reports.

A key contribution of this systematic review is the application of quantitative meta-analyses in a field where such synthesis was previously considered methodologically challenging [10,11]. Unlike prior meta-analyses that focused on single domains or summarized findings narratively because of methodological heterogeneity [13,14], this review systematically evaluates digital biomarkers across diverse domains. Consistent with earlier reports, the results indicate that depression is not characterized by a single physiological signature but rather by a constellation of behavioral and biological changes [6,149].

Furthermore, unlike recent scoping reviews on general mental illness [151], our analysis is specifically tailored to MDD. Importantly, the interpretation of these findings must consider the substantial heterogeneity observed across studies. The CIs reflect the average effect across the included studies, whereas the PIs indicate the range of effects that may occur in future settings. For several outcomes, wide PIs crossing the null value suggest that effect sizes may vary considerably depending on the study population, device type, measurement protocol, and analytical approach. These findings indicate that group-level average effects may not generalize consistently across contexts, supporting the use of personalized, longitudinal multibiomarker models for effective monitoring and intervention.

Several biomarkers emerged as important indicators of depression across multiple domains. In the sleep domain, individuals with depression showed significantly longer SOL and increased TIB compared to controls without depression. These findings suggest that the sleep-wake cycle in MDD is characterized more by structural fragmentation than by simple reductions in sleep duration. Specifically, increased TIB likely reflects hallmark symptoms such as psychomotor retardation and lethargy. While polysomnography remains the gold standard [152], these results demonstrate that wearable devices can capture such patterns in ecologically valid, naturalistic settings.

The absence of significant differences in TST, SE, and WASO, which are parameters frequently associated with depression in previous reviews [4], highlights the clinical heterogeneity of MDD [153,154]. Depression may manifest as either insomnia or hypersomnia, depending on the subtype, potentially neutralizing average effects in pooled analyses [155]. The wide PIs further indicate substantial heterogeneity, likely driven by differences in study populations, medication use, and device-specific scoring. These findings underscore that absolute sleep quantity alone is an insufficient marker, and the relationship between sleep and depression is shaped by complex biopsychosocial factors [153,154]. Consequently, temporal and qualitative features, such as SOL and TIB, may serve as more clinically informative indicators than total sleep volume [9,156].

With respect to sleep architecture, evidence regarding REM-related parameters remains inconclusive. While some studies suggest that shortened REM latency may precede depression [157], results for REM duration and frequency are inconsistent [26,31,41,115,116]. Age-related changes [33] and reliance on single-night laboratory measurements further limit generalizability. These findings highlight the need for longitudinal, real-world assessments to clarify the relationship between sleep architecture and depression.

Physical activity counts emerged as a sensitive state-dependent marker, showing significant reductions in depression groups. Compared with simple step counts, activity counts better capture movement intensity and frequency [65,142,158], reflecting the energy deficits associated with MDD. Although daily step counts were not significantly different in pooled analyses, longitudinal evidence suggests that they may be sensitive to individual recovery trajectories [65,142]. Several studies also reported more pronounced reductions during daytime hours and on weekends, emphasizing the importance of temporal patterns rather than simple daily averages [158,159]. In contrast, MVPA, sedentary time, and LPA showed inconsistent results. This lack of robust significance likely reflects high heterogeneity and the limited specificity of these markers when used in isolation.

GPS-derived parameters, including total distance traveled, location entropy, time spent at home, and number of locations visited, capture behavioral and environmental changes associated with depression. Mobility patterns, such as reduced diversity in visited locations (entropy) and increased homestay durations, have emerged as promising indicators of depressive symptoms [160]. Recent longitudinal evidence suggests that early changes in homestay and mobility patterns may predict symptom improvement [136], highlighting the value of GPS-derived features in multimodal monitoring frameworks [149].

Cardiac parameters showed less consistent results. Individuals with depression tended to exhibit higher nocturnal mHR and lower HRV, although these differences were not statistically significant in pooled analyses. These trends may reflect physiological hyperarousal and autonomic dysregulation [161,162], but their effects appear to vary across populations and study conditions. Accordingly, these markers may be more informative when interpreted within multimodal models that account for age, medication use, and comorbid conditions. Despite this overall inconsistency, short-term HRV indices and frequency-domain features (eg, HF and VLF) remain meaningful components in diagnostic and multimodal frameworks, particularly as sensitive indicators of autonomic imbalance in first-episode depression or high-risk individuals [54,125,149].

Speech parameters, including slower speech rates and longer pauses, consistently distinguished individuals with MDD from controls in several studies [87,144]. These acoustic features likely reflect psychomotor retardation.

Digital biomarkers offer several advantages over traditional self-reported measures by providing objective, continuous data that reduce recall bias and capture early physiological and behavioral changes [163,164]. Passive monitoring enables large-scale, low-burden data collection, supporting early detection and personalized, data-driven interventions [5,8,165].

However, the substantial variability observed across individuals highlights the need to interpret these signals within a personalized framework, as traditional group-based models may be inadequate in many contexts [166]. The predictive power of biomarkers varies with age. TIB, daily step counts, and MVPA are more predictive in older adults [11,109,167], whereas smartphone usage patterns show stronger associations in younger individuals [11,122]. High-resolution temporal data, such as nighttime HRV or weekend reductions in MVPA, may provide additional insights into symptom dynamics, underscoring the importance of temporal patterns in digital measurements [69]. Monitoring sleep-related biomarkers in relation to changes in individual symptoms may further improve the prediction and tracking of depressive episodes [6]. Together, these findings suggest that digital monitoring strategies should focus on detecting deviations from an individual’s baseline rather than relying on universal thresholds.

Despite identifying several robust biomarkers, parameters, such as LPA, sedentary time, and specific cardiac measures, showed inconsistent results. This inconsistency, together with the wide PIs observed across studies, likely reflects not only diverse device technologies and measurement protocols, but also the symptomatic variability across depression subtypes [155,168]. While wearable devices offer objective physiological data, they do not directly capture emotional states [169]. Furthermore, the predominance of cross-sectional designs limits our ability to determine the temporal or causal relationships between these digital signals and depressive symptoms [170].

Risk-of-bias assessments also indicated several methodological concerns. Although participation rates were somewhat low in certain studies [24,34,59,62,128], they were generally acceptable across the majority of the included research [4,25,27,31,32,35,38,42,45,49,52-54,56,60,67,68,76,89,91,101,111,112,114,115,117,131,145]. However, comparisons between participants and nonparticipants were often not reported, raising the possibility of selection bias. For example, individuals with more severe depressive symptoms, lower motivation, or limited familiarity with wearable technologies may have been less likely to participate or remain in the studies. In addition, adjustment for confounding variables was limited in several studies. For diagnostic accuracy studies, most domains showed low risk of bias, but concerns regarding applicability were common because study populations did not always reflect real-world clinical settings.

Consistent with these methodological limitations, the GRADE assessment indicated low to very low certainty of evidence for the key digital biomarkers. This was primarily due to inconsistency across studies and imprecision associated with wide CIs and PIs. These findings suggest that the pooled estimates should be interpreted cautiously and that further well-designed, standardized studies are needed to strengthen the evidence base.

This systematic review has several limitations. First, substantial heterogeneity across studies limited the precision and generalizability of the pooled estimates, reflecting differences in devices, study populations, monitoring periods, and analytic methods. Second, the overall certainty of evidence was low to very low according to the GRADE framework, which reduces confidence in the pooled estimates. Third, a formal statistical assessment of reporting bias was not feasible because fewer than 10 studies were available for each meta-analysis. Finally, many studies included nonclinical or convenience samples, which may limit the generalizability of the findings to real-world clinical populations.

Future research should move beyond the identification of individual markers toward the development of integrated, clinically actionable digital phenotyping systems. The consistent associations observed for indicators, such as SOL, TIB, and activity counts, suggest that certain behavioral signals may serve as foundational components of continuous mental health monitoring. The next phase of research should therefore focus on embedding these markers within longitudinal, multimodal frameworks that support personalized clinical decision-making and improve the precision of depression monitoring [149,150,155], with the potential to inform more proactive intervention strategies in real-world settings.

Achieving this transition will require greater methodological standardization across devices, measurement protocols, and analytic pipelines. Reducing reliance on proprietary, nontransparent algorithms and promoting device-agnostic, reproducible approaches will be essential for ensuring clinical validity and interoperability across health care systems. Integration of digital biomarker data into electronic health records may further enable real-time, context-aware decision support while minimizing additional cognitive burden on clinicians.

More broadly, digital biomarkers may support earlier detection and more adaptive treatment strategies by providing high-frequency, objective data on symptom trajectories [136,164]. At the population level, scalable signals such as physical activity and mobility patterns may also facilitate screening and risk stratification in settings with limited access to mental health professionals. Future work should therefore prioritize longitudinal, multimodal designs; representative clinical populations; standardized measurement protocols; and transparent reporting of analytic methods.

This systematic review provides a comprehensive synthesis of multimodal digital biomarkers for depression. Unlike previous reviews that focused on single signals or technical feasibility, this systematic review advances the field by establishing a standardized framework for objective clinical decision-making through a rigorous meta-analysis. The findings indicate that while certain markers, particularly SOL and physical activity counts, show consistent average differences, their effects vary substantially across settings, as reflected by wide PIs. These results suggest that depression cannot be reliably characterized by a single digital biomarker. Instead, a multimodal, personalized approach that integrates physiological, behavioral, and contextual signals is likely to be more effective for real-world applications. More broadly, this systematic review demonstrates that quantitative synthesis in digital phenotyping is feasible despite substantial heterogeneity and that meaningful signals can be identified when methodological rigor and transparent reporting are applied. Establishing standardized, clinically interpretable digital biomarker frameworks will be essential for advancing objective, continuous, and personalized assessments of depression in routine care.