A multilevel meta-analysis of passive smartphone sensing of adolescent mental health Authors: Merel M. L. Leijse, Aaliyah Hansen, Imara K. Semeijn, Eleonore Breuning, Janneke P. C. Staaks, Thimo M. van der Pol, Geert-Jan J. M. Stams, Arne Popma, Levi van Dam Published: 2026 ABSTRACT Adolescence is a critical developmental stage in which the risk of mental health problems peaks. Passive smartphone sensing offers valuable opportunities for moment-to-moment assessment of such risk. The amount of research is growing steadily, but a quantitative synthesis of the literature is lacking. Therefore, this meta-analysis evaluated the association between passive smartphone data (e.g., GPS, call logs, and notifications) and adolescents’ (12–24 years) mental health outcomes across 45 independent samples (N = 2939). Findings revealed a small but significant overall effect (r = 0.12). Most passive data correlated with mental health outcomes, except for the number of contacts, while linguistic markers showed a significant negative correlation. Associations were not moderated by study design features or type of outcome, but were significantly stronger in non-student than in student samples. These results highlight passive smartphone sensing as a promising tool for assessing precursors of adolescent mental health and provide guidance for future research. Introduction Adolescence is both a dynamic and challenging phase of life, characterized by significant physiological, psychological, and emotional changes, which increase vulnerability to mental health problems1,2. Notably, the peak age to develop a first mental health disorder typically emerges at the age of 143,4, with 75% developing before the age of 253,5. According to the World Health Organization, one in seven adolescents is affected by a mental health disorder, representing approximately 15% of the global disease burden in this age group6. The assessment and subsequent intervention of early signs and symptoms in adolescence are critical for reducing the risk of more persistent mental health problems in adulthood7,8. Despite high prevalence rates of adolescent mental health problems, symptoms often remain undetected, overlooked, and untreated9,10. Explanations can be found in the limited accessibility of mental health services11,12, long waiting lists, and the persistent influence of limited awareness and stigma of mental health problems13,14. Within services, detection and monitoring are further constrained by reliance on retrospective self-reports and brief clinical observations, which obscure the day-to-day mechanisms that sustain mental health problems15. These limitations require tools that are easily accessible and provide continuous information for the timely detection of mental health outcomes in adolescence. In recent years, there has been growing attention to the integration of new technologies for mental health assessment16–18. Digital phenotyping, defined as the continuous collection of data from smartphones, wearables, and other sensor-enabled devices to monitor human behavior in daily life, has become a widely studied approach15,19–21. This encompasses both the collection of active data, requiring a direct user’s input, like self-report questionnaires or daily mood ratings22,23. In contrast, passive data that are generated without direct effort from the individual and are typically derived from sensors or device log files, such as GPS, accelerometers, call logs, or app usage21,24. The combination of active and passive data has shown more accurate, personalized, and ecologically valid predictions. At the same time, it still entails methodological challenges. Research has consistently shown that active self-report measures are affected by declining adherence over time and substantial, non-random missingness, often related to momentary burden, motivation, or symptom severity25–27. Despite the methodological strength of combining active and passive data, these approaches remain strongly dependent on active input and are therefore constrained by engagement-related limitations27–29. Passive data sensing entails technical and ethical challenges, including privacy concerns, high battery consumption, and data fragmentation28,30. However, these challenges are increasingly being addressed through methodological and technological developments31,32. For example, improvements in energy-efficient sensing and adaptive sampling have contributed to reducing battery-related burden, while advances in data processing and feature extraction have increased robustness in the analysis of fragmented or incomplete data streams32–34. Together with developments in longitudinal modeling and analytical methods, these advances have improved the usability and accuracy of passive sensing approaches, enabling the identification of behavioral patterns and risk-related changes using passive data alone32,35 As such, passive sensing has shown promising ways to capture mental health outcomes in a methodologically consistent and low-burden manner36,37. Smartphones in particular offer low-threshold opportunities for the continuous monitoring of behavior23,38. In contrast to wearables and other sensor-enabled devices, smartphones are nearly ubiquitous among adolescents, reflecting their widespread and routine use among adolescents. Moreover, smartphones enable a more comprehensive picture of daily functioning by capturing not only behavioral features commonly assessed by other wearables (e.g., activity and sleep) but also communicative and social data, such as call logs, messaging, and social media use. Together, this makes smartphones a low-threshold and comprehensive tool for assessing behavioral factors in relation to mental health. The type of smartphone data can be generated through embedded raw sensors and device log files that can be transformed into low-level features, such as step-based activity, sleep duration, or call duration, which in turn capture higher-level behavioral markers including physical activity, sleep quality, and social interaction. Such markers are increasingly linked to mental health outcomes such as depression, anxiety, and stress16,29,39. For example, accelerometer data can capture patterns of movement that indicate physical activity, while call logs and phone usage reflect levels of social engagement39. Changes in these behavioral markers may signal lifestyle disruptions or early symptom deterioration, which can represent potential risk for mental health problems16,22. Beyond early detection, such objective real-time measures also offer valuable opportunities for self-monitoring and clinical assessment, supporting the delivery of more personalized interventions targeting specific mental health outcomes23,40. Research on passive smartphone sensing has increased substantially in recent years. Reviews have described research on monitoring by sensor-derived trace data for youth, highlighting both their promise, yet also emphasizing the heterogeneity of study designs and outcomes41,42. Some reviews have focused on the predictive value of passive sensing for particular mental health problems, such as depression and anxiety43 or suicidal thoughts and behavior44. These studies indicate that effect sizes vary considerably across data streams and outcomes, leading to inconsistent results. In addition, work in clinical populations has so far mainly relied on small samples and focused on feasibility and acceptability, with relatively few studies evaluating clinical outcomes45,46. Despite the growing number of studies, overall syntheses of the association between passive smartphone sensing and mental health outcomes remain scarce. Existing meta-analyses on digital interventions in adolescents have mainly focused on mobile technologies or standalone apps, where passive data may play a role but have not been explicitly examined18,47,48. Meta-analytic reviews have addressed specific passive data streams, such as GPS traces in relation to mental health problems like depression49 or PTSD50. However, these meta-analyses have predominantly targeted adult populations and did not systematically investigate the broader use of passive sensing for mental health outcomes in adolescence. While conceptual frameworks have outlined the potential value of such approaches in adolescents16,39, a comprehensive quantitative synthesis is still lacking. To our knowledge, no prior meta-analysis has systematically examined passive smartphone sensing in relation to adolescent mental health outcomes. Therefore, this study primarily aimed to synthesize existing evidence and quantify associations between passive smartphone markers and adolescent mental health outcomes, focusing on adolescence as a developmental period ranging from early adolescence to young adulthood51,52. A secondary aim was to examine whether these associations varied across different passive sensor markers and whether study characteristics did affect these associations. By synthesizing the emerging evidence from empirical studies, this meta-analysis aims to clarify the potential and limitations of passive smartphone sensing for identifying early signs of mental health problems in adolescence. Results A total of 3860 records were identified through electronic databases. An additional 36 studies were screened from the reference lists of related reviews, of which 2 studies met the inclusion criteria. After the screening process and retrieval, 45 studies with a total of 2939 participants were included. See Fig. 1 for the PRISMA 2020 flow diagram for the full search and selection process53.Fig. 1PRISMA 2020 flow diagram.This diagram represents the systematic search strategy and the process for the selection of studies included in this review. This figure is adapted from the PRISMA 2020 statement, which is distributed under the terms of the Creative Commons Attribution 4.0 International License (CC BY 4.0). PRISMA 2020 flow diagram. This diagram represents the systematic search strategy and the process for the selection of studies included in this review. This figure is adapted from the PRISMA 2020 statement, which is distributed under the terms of the Creative Commons Attribution 4.0 International License (CC BY 4.0). Study characteristics The characteristics of all included study reports are summarized in Table 1. Most studies were conducted in North America (80%), followed by Europe (11%), Asia (9%), and Oceania (0.02%). The mean age of participants across studies was 20.05 years (SD = 2.01), across an age range of 12–25 years. Notably, the majority of samples consisted of students (69%), followed by community (13%), mixed (11%), and clinical (7%). Regarding passive smartphone sensing methods, a combination of three or more markers was most frequently applied (44%). Thereafter, GPS was the most common single marker used (24%). Sleep, keystroke, and linguistic markers followed, each accounting for 4% of studies. With respect to mental health outcomes, depression was the most frequently examined outcome within studies (44%). All papers were published between 2016 and 2025.Table 1Study characteristicsStudySamplePassive smartphone sensingMental health outcomesAuthor (year)CountryDesignNAge (M)% MaleType of sampleType of sensingSensing durationCombined active dataOperating systemType of mental health outcomeMeasureInformantAlamoudi et al. (2024)110U.K.Long1121.5036%StudentsSleep56 daysNoCombinedAnxiety and depressionPHQ-9 and GAD-7AdolescentAuerbach et al. (2025)61U.S.Long18616.4320%MixedGPS~180 daysESMCombinedSuicideSSIAdolescentAyranci et al. (2022)111U.S.Long124N/A69%StudentsGPS~90 daysNoAndroidAnxiety and depressionPHQ-2 and GAD-2AdolescentBosma et al. (2024)112U.S.Long6919.70N/AStudentsGPS7 daysNoCombinedNegative feelings and internalizingSSIAdolescentBoukhechba, Chow et al.(2018)113U.S.Long22819.4338%StudentsGPS14 daysNoCombinedAnxietySIASAdolescentBoukhechba, Daros et al (2018)114U.S.Long7219.8049%StudentsCalls and texting14 daysEMAAndroidAnxiety, depression, positive feelings and negative feelingsSIAS, DASS and PANASAdolescentBraund et al. (2023)115AustraliaLong93413.8436%StudentsKeystrokeN/ANoCombinedAnxiety, depression and stressCAS-8, PHQ-A and DQ-5AdolescentBrogly (2022)116U.S.Long101N.A.24%StudentsGPS, accelerometer, contact, app usage, and screen usage280 daysNoCombinedAnxiety, depression, stress, addiction, positive and negative feelingsDASS-21 and PANASAdolescentByrne et al. (2021)117U.S.Long2520.6452%StudentsKeystroke and linguistic marker2 ×7 daysNoAndroidAnxiety, depression and stressDASS, PSS and STRAINAdolescentCao et al. (2020)118U.S.Long1314.9315%ClinicalCalls, texting, sleep and screen usage56 daysEMACombinedAnxiety, depression and internalizingPHQ-9, HAM-D and HAM-AAdolescentChikersal et al. (2021)119U.S.Long138N/AN/AStudentsCalls, GPS, accelerometer, Bluetooth, and sessions16 ×7 daysNoCombinedDepressionBDI-IIAdolescentChow et al. (2017)120U.S.Cross and long7220.5073%StudentsGPS14–16 daysEMAAndroidAnxiety, depression, positive feelings, negative feelingsSIAS, DASS-21 and SensusAdolescentCunningham (2017)121U.S.Long48N.A.N.A.StudentsGPS70 daysNoCombinedDepressionPHQ-9AdolescentCurrey and Torous (2022)122U.S.Long147N/AN/AStudentsCalls, GPS, Bluetooth, sleep, and screen28 daysNoCombinedAnxiety, depression, loneliness, stress and psychosisGAD-7, PHQ-9, UCLA, PSS, PQ-16AdolescentCurrey and Torous (2023)123U.S.Long672022%StudentsGPS28 daysNoCombinedAnxiety, depression, loneliness and stressGAD-7, PGQ-9, UCLA and PSSAdolescentDissing et al. (2021)124DenmarkLong81621.6077%CommunityCalls, texting, and contact75–163 daysNoCombinedDepression and lonelinessMDI and UCLAAdolescentElhai et al. (2021)125U.S.Cross10319.2833%StudentsNotification, accelerometer, and screen usage~56 daysNoIOSAnxiety, depression and stressDASS-21 and FOMO ScaleAdolescentFarhan et al. (2016)126U.S.Long79N/A26%StudentsGPS~240 daysNoCombinedDepressionPHQ-9AdolescentFunkhouser et al. (2023)127U.S.Long9016.5737%ClinicalLinguistic marker365 daysDaily self-reportCombinedDepressionA-LifeAdolescentGansner et al. (2023)128U.S.Long1815.6050%ClinicalGPS, accelerometer, sessions, and screen usage42 daysDaily self-reportCombinedAnxiety, depression and addictionGAD-7, PGQ-8 and PIU-SF-6AdolescentGong et al. (2019)129U.S.Long5220.4032%StudentsAccelerometerN/ANoAndroidAnxietySIASAdolescentGyorda et al. (2024)130U.S.Long33019.6741%StudentsLock/unlock70 daysNoIOSDepressionPHQ-4AdolescentHuang et al. (2016)131U.S.Long18N/AN/AStudentsGPS10 daysNoStudentsAnxietySIASAdolescentIlyas et al. (2023)132U.S.Long1919.0032%CommunityGPS~180 daysNoAndroidDepression and stressPCL and PHQ-9AdolescentJeyashree et al. (2021)133IndiaCross16320.0064%StudentsScreen usage28–70 daysNoAndroidStressPSS-10AdolescentKim et al. (2021)134South KoreaLong3721.5049%StudentsGPS, lock/unlock, Bluetooth, sleep, app usage, and screen usage22 daysNoAndroidInternalizingCIPSAdolescentKuhney et al. (2025)135U.S.Long19321.7023%MixedCalls, texting, screen usage6 daysEMAAndroidAnxiety, depression, psychosisBAI, BDI, SIPS, and NSI-PRAdolescentLi et al. (2023)136U.S.Long8316.4939%StudentsLinguistic marker90 daysDaily self-reportCombinedDepression and internalizingIDAS-IIAdolescentLim et al. (2023)137South KoreaLong6622.4435%CommunityKeystroke35 daysNoAndroidLonelinessULCAAdolescentMarin-Dragu (2022)138CanadaLong45118.5017%MixedLock/unlock and screen usage14 daysNoCombinedLoneliness and internalizingSDQAdolescentMarin-Dragu et al. (2023)65CanadaLong45120.9717%MixedLock/unlock and screen usageContinueNoCombinedLoneliness and internalizingSDQAdolescentPrasad et al. (2018)139IndiaLong14022.8950%StudentsLock/unlock, app usage, and screen usage7 daysNoAndroidStress and addictionSASAdolescentQirtas et al. (2023)140U.S.Long48N.A.N.A.StudentsCalls, GPS, lock/unlock, Bluetooth, sleep, and sessionsN/ANoCombinedDepression and lonelinessPHQ-9 and UCLAAdolescentRozgonjuk et al. (2018)141U.S.Long10119.5324%CommunityLock/unlock and screen usage7 daysDaily self-reportIOSAnxiety and depressionDASS-21 and PHQ-9AdolescentSaeb et al. (2016)142U.S.Long48N.A.79%StudentsGPS and accelerometer10 ×7 daysNoAndroidDepressionPHQ-9AdolescentŠutić and Novak (2024)143CroatiaLong10215.1332%CommunitySleep8 daysESMCombinedNegative feelingsESIRAdolescentThakur and Roy (2020)144U.S.Long45N/A84%StudentsCalls, GPS, accelerometerN/ANoCombinedDepression, loneliness, stress and positive feelingsPHQ-9, UCLA, PSS and FSAdolescentTruong et al. (2017)145U.S.Cross1314.50N.A.CommunityCalls, accelerometer, and texting56 daysDaily self-reportAndroidInternalizingPHQ-9 and HAM–AAdolescentVan de Casteele et al. (2024)146BelgiumLong10715.2847%StudentsApp usage and screen usage7 daysDaily self-reportAndroidPositive feelings and negative feelingsN/AAdolescentWang et al. (2018)147U.S.Long8320.1348%StudentsCalls, GPS, lock/unlock and sleep63 daysDaily self-reportCombinedDepressionPHQ-8AdolescentWare et al. (2020)148U.S.Long18221.5049%StudentsGPSN/ANoCombinedDepressionPHQ-9 and QIDSAdolescentWare et al. (2020)149U.S.Long5921.50N/AMixedKeystroke, calls, texting, and contact14 days (sliding window)NoAndroidDepressionQIDSAdolescentYang et al. (2017) 150U.K.Long48N/AN/AStudentsGPS and accelerometer28 daysNoCombinedDepression and stressPHQ-9 and PSSAdolescentYuan and Cao (2024)151U.S.Cross and long725N/A70%StudentsLock/unlock, accelerometer, sleep and screen usage28 daysDaily self-reportAndroidDepressionPHQ-9 and SASAdolescentYue (2020)152U.S.Long7921.50N.A.StudentsGPS, accelerometer, sessions and screen usageN/ANoCombinedDepressionPHQ-9ProfessionalAge is reported in years. N/A = not-applicable; Cross = cross-sectional; Long = longitudinal; Mixed = sample included both community and clinical. Table 1Study characteristicsStudySamplePassive smartphone sensingMental health outcomesAuthor (year)CountryDesignNAge (M)% MaleType of sampleType of sensingSensing durationCombined active dataOperating systemType of mental health outcomeMeasureInformantAlamoudi et al. (2024)110U.K.Long1121.5036%StudentsSleep56 daysNoCombinedAnxiety and depressionPHQ-9 and GAD-7AdolescentAuerbach et al. (2025)61U.S.Long18616.4320%MixedGPS~180 daysESMCombinedSuicideSSIAdolescentAyranci et al. (2022)111U.S.Long124N/A69%StudentsGPS~90 daysNoAndroidAnxiety and depressionPHQ-2 and GAD-2AdolescentBosma et al. (2024)112U.S.Long6919.70N/AStudentsGPS7 daysNoCombinedNegative feelings and internalizingSSIAdolescentBoukhechba, Chow et al.(2018)113U.S.Long22819.4338%StudentsGPS14 daysNoCombinedAnxietySIASAdolescentBoukhechba, Daros et al (2018)114U.S.Long7219.8049%StudentsCalls and texting14 daysEMAAndroidAnxiety, depression, positive feelings and negative feelingsSIAS, DASS and PANASAdolescentBraund et al. (2023)115AustraliaLong93413.8436%StudentsKeystrokeN/ANoCombinedAnxiety, depression and stressCAS-8, PHQ-A and DQ-5AdolescentBrogly (2022)116U.S.Long101N.A.24%StudentsGPS, accelerometer, contact, app usage, and screen usage280 daysNoCombinedAnxiety, depression, stress, addiction, positive and negative feelingsDASS-21 and PANASAdolescentByrne et al. (2021)117U.S.Long2520.6452%StudentsKeystroke and linguistic marker2 ×7 daysNoAndroidAnxiety, depression and stressDASS, PSS and STRAINAdolescentCao et al. (2020)118U.S.Long1314.9315%ClinicalCalls, texting, sleep and screen usage56 daysEMACombinedAnxiety, depression and internalizingPHQ-9, HAM-D and HAM-AAdolescentChikersal et al. (2021)119U.S.Long138N/AN/AStudentsCalls, GPS, accelerometer, Bluetooth, and sessions16 ×7 daysNoCombinedDepressionBDI-IIAdolescentChow et al. (2017)120U.S.Cross and long7220.5073%StudentsGPS14–16 daysEMAAndroidAnxiety, depression, positive feelings, negative feelingsSIAS, DASS-21 and SensusAdolescentCunningham (2017)121U.S.Long48N.A.N.A.StudentsGPS70 daysNoCombinedDepressionPHQ-9AdolescentCurrey and Torous (2022)122U.S.Long147N/AN/AStudentsCalls, GPS, Bluetooth, sleep, and screen28 daysNoCombinedAnxiety, depression, loneliness, stress and psychosisGAD-7, PHQ-9, UCLA, PSS, PQ-16AdolescentCurrey and Torous (2023)123U.S.Long672022%StudentsGPS28 daysNoCombinedAnxiety, depression, loneliness and stressGAD-7, PGQ-9, UCLA and PSSAdolescentDissing et al. (2021)124DenmarkLong81621.6077%CommunityCalls, texting, and contact75–163 daysNoCombinedDepression and lonelinessMDI and UCLAAdolescentElhai et al. (2021)125U.S.Cross10319.2833%StudentsNotification, accelerometer, and screen usage~56 daysNoIOSAnxiety, depression and stressDASS-21 and FOMO ScaleAdolescentFarhan et al. (2016)126U.S.Long79N/A26%StudentsGPS~240 daysNoCombinedDepressionPHQ-9AdolescentFunkhouser et al. (2023)127U.S.Long9016.5737%ClinicalLinguistic marker365 daysDaily self-reportCombinedDepressionA-LifeAdolescentGansner et al. (2023)128U.S.Long1815.6050%ClinicalGPS, accelerometer, sessions, and screen usage42 daysDaily self-reportCombinedAnxiety, depression and addictionGAD-7, PGQ-8 and PIU-SF-6AdolescentGong et al. (2019)129U.S.Long5220.4032%StudentsAccelerometerN/ANoAndroidAnxietySIASAdolescentGyorda et al. (2024)130U.S.Long33019.6741%StudentsLock/unlock70 daysNoIOSDepressionPHQ-4AdolescentHuang et al. (2016)131U.S.Long18N/AN/AStudentsGPS10 daysNoStudentsAnxietySIASAdolescentIlyas et al. (2023)132U.S.Long1919.0032%CommunityGPS~180 daysNoAndroidDepression and stressPCL and PHQ-9AdolescentJeyashree et al. (2021)133IndiaCross16320.0064%StudentsScreen usage28–70 daysNoAndroidStressPSS-10AdolescentKim et al. (2021)134South KoreaLong3721.5049%StudentsGPS, lock/unlock, Bluetooth, sleep, app usage, and screen usage22 daysNoAndroidInternalizingCIPSAdolescentKuhney et al. (2025)135U.S.Long19321.7023%MixedCalls, texting, screen usage6 daysEMAAndroidAnxiety, depression, psychosisBAI, BDI, SIPS, and NSI-PRAdolescentLi et al. (2023)136U.S.Long8316.4939%StudentsLinguistic marker90 daysDaily self-reportCombinedDepression and internalizingIDAS-IIAdolescentLim et al. (2023)137South KoreaLong6622.4435%CommunityKeystroke35 daysNoAndroidLonelinessULCAAdolescentMarin-Dragu (2022)138CanadaLong45118.5017%MixedLock/unlock and screen usage14 daysNoCombinedLoneliness and internalizingSDQAdolescentMarin-Dragu et al. (2023)65CanadaLong45120.9717%MixedLock/unlock and screen usageContinueNoCombinedLoneliness and internalizingSDQAdolescentPrasad et al. (2018)139IndiaLong14022.8950%StudentsLock/unlock, app usage, and screen usage7 daysNoAndroidStress and addictionSASAdolescentQirtas et al. (2023)140U.S.Long48N.A.N.A.StudentsCalls, GPS, lock/unlock, Bluetooth, sleep, and sessionsN/ANoCombinedDepression and lonelinessPHQ-9 and UCLAAdolescentRozgonjuk et al. (2018)141U.S.Long10119.5324%CommunityLock/unlock and screen usage7 daysDaily self-reportIOSAnxiety and depressionDASS-21 and PHQ-9AdolescentSaeb et al. (2016)142U.S.Long48N.A.79%StudentsGPS and accelerometer10 ×7 daysNoAndroidDepressionPHQ-9AdolescentŠutić and Novak (2024)143CroatiaLong10215.1332%CommunitySleep8 daysESMCombinedNegative feelingsESIRAdolescentThakur and Roy (2020)144U.S.Long45N/A84%StudentsCalls, GPS, accelerometerN/ANoCombinedDepression, loneliness, stress and positive feelingsPHQ-9, UCLA, PSS and FSAdolescentTruong et al. (2017)145U.S.Cross1314.50N.A.CommunityCalls, accelerometer, and texting56 daysDaily self-reportAndroidInternalizingPHQ-9 and HAM–AAdolescentVan de Casteele et al. (2024)146BelgiumLong10715.2847%StudentsApp usage and screen usage7 daysDaily self-reportAndroidPositive feelings and negative feelingsN/AAdolescentWang et al. (2018)147U.S.Long8320.1348%StudentsCalls, GPS, lock/unlock and sleep63 daysDaily self-reportCombinedDepressionPHQ-8AdolescentWare et al. (2020)148U.S.Long18221.5049%StudentsGPSN/ANoCombinedDepressionPHQ-9 and QIDSAdolescentWare et al. (2020)149U.S.Long5921.50N/AMixedKeystroke, calls, texting, and contact14 days (sliding window)NoAndroidDepressionQIDSAdolescentYang et al. (2017) 150U.K.Long48N/AN/AStudentsGPS and accelerometer28 daysNoCombinedDepression and stressPHQ-9 and PSSAdolescentYuan and Cao (2024)151U.S.Cross and long725N/A70%StudentsLock/unlock, accelerometer, sleep and screen usage28 daysDaily self-reportAndroidDepressionPHQ-9 and SASAdolescentYue (2020)152U.S.Long7921.50N.A.StudentsGPS, accelerometer, sessions and screen usageN/ANoCombinedDepressionPHQ-9ProfessionalAge is reported in years. N/A = not-applicable; Cross = cross-sectional; Long = longitudinal; Mixed = sample included both community and clinical. Study characteristics Age is reported in years. N/A = not-applicable; Cross = cross-sectional; Long = longitudinal; Mixed = sample included both community and clinical. Main analysis Table 2 presents meta-analytic findings on passive smartphone data in relation to mental health outcomes. In total, 45 independent studies (k), reporting on 663 effect sizes (#ES) across N = 2939 participants. All studies were published between 2016 and 2025. When combining all effect sizes into one overall analysis, passive smartphone sensing showed a small (r = 0.12) but significant (t = 3.274, p < 0.001) association with mental health, after having recoded 6 outlying effect sizes into the highest scores in the normal range. The variability in effect sizes was partitioned across three levels: sampling error (Level 1), within-study variance (Level 2), and between-study variance (Level 3). Level 1 accounted for 12.32% of the variance, Level 2 for 25.37%, and Level 3 for 62.31%, pointing to substantial heterogeneity of effect sizes both within and between studies. Figure 2 presents the forest plot of individual effect sizes per study. The number of effect sizes contributed by each study ranged from 1 to 66. A total of 6 studies showed a significant positive association with mental health outcomes based on the confidence intervals of the study mean effect sizes, whereas 2 studies showed a significant negative association: 37 studies did not show a significant association.Fig. 2Forest plot.This plot displays the effect sizes and 95% confidence intervals for each included study, providing a visual summary of the variation across the literature. The central black markers indicate the mean effect for each study, while the gray horizontal lines represent the sampling variance, with thickness reflecting the relative weight of the data. The overall synthesized effect is presented to illustrate the general trend and statistical significance of the findings.Table 2Overall relation between passive smartphone data and mental health outcomesOutcomek#ESr95% CIpσ2level 2σ2level 3% Var Level 1% Var Level 2% Var Level 3Mental health outcomes456630.120.05; 0.190.0010.021***0.052***12.3225.3762.31k number of studies, #ES number of effect sizes, mean r mean effect size (r), CI confidence interval, σ2level 2 % within study variance, σ2level 3 % between study variance. *** p < 0.001. Forest plot. This plot displays the effect sizes and 95% confidence intervals for each included study, providing a visual summary of the variation across the literature. The central black markers indicate the mean effect for each study, while the gray horizontal lines represent the sampling variance, with thickness reflecting the relative weight of the data. The overall synthesized effect is presented to illustrate the general trend and statistical significance of the findings. Table 2Overall relation between passive smartphone data and mental health outcomesOutcomek#ESr95% CIpσ2level 2σ2level 3% Var Level 1% Var Level 2% Var Level 3Mental health outcomes456630.120.05; 0.190.0010.021***0.052***12.3225.3762.31k number of studies, #ES number of effect sizes, mean r mean effect size (r), CI confidence interval, σ2level 2 % within study variance, σ2level 3 % between study variance. *** p < 0.001. Overall relation between passive smartphone data and mental health outcomes k number of studies, #ES number of effect sizes, mean r mean effect size (r), CI confidence interval, σ2level 2 % within study variance, σ2level 3 % between study variance. *** p < 0.001. Moderator analyses The results of the moderator analyses are presented in Table 3. Study characteristics did not significantly moderate the association between passive smartphone data and mental health outcomes, suggesting that study-level features did not account for meaningful variation in effect sizes. Regarding study sample characteristics, the associations were consistent across age groups, gender, and clinical versus non-clinical populations. In contrast, student status was a significant moderator (F(1, 630) = 5.871, p < 0.05), with a stronger association observed in non-student samples (r = 0.29, p < 0.001). Data characteristics also moderated the overall association (F(10, 322) = 1.918, p < 0.05). Linguistic markers and number of social contacts differed significantly from the reference category GPS. Linguistic markers showed a significant and negative association with mental health outcomes (r = −0.31, p < 0.01), whereas the association between the number of social contacts proved to be non-significant (r = 0.06, ns). For low-level behavioral features, keystroke-derived measures of text variability (r = −0.29, p < 0.01) and text intensity (r = 0.49, p < 0.001) differed significantly from the reference category location context (r = 0.11, p < 0.05). Outcome characteristics showed no moderating effects of either the type of mental health outcome or the instrument used to assess it.Table 3Results of moderator analyses for the association between passive smartphone data and mental health outcomesModeratork#ESΒ0/Fisch. zrt0β1t1F(df1, df2)Study design characteristicsImpact factor404680.120.122.796**0.000.598F(1, 466) = 0.357Publication year456630.120.123.236**0.000.072F(1, 661) = 0.005PublishedF(1, 661) = 0.040 Published404680.120.122.986** Not published51950.140.141.2660.020.199RegionF(2, 660) = 2.318+ Northern America344820.100.102.554* Europe61230.030.030.342−0.07−0.651 Asia5580.320.313.058**0.221.946+DesignF(1, 661) = 0.000 Cross-sectional5240.120.121.631 Longitudinal426390.120.123.236**0.000.004Time between assessments426390.110.112.690**−0.00−0.424F(1, 637) = 0.180Risk bias assessment (index)456630.110.112.906**0.051.433F(1, 661) = 2.0532Sample characteristicsMean age344410.040.040.1070.010.351F(1, 439) = 0.123Adolescence periodF(2, 438) = 1.175 Early3180.330.322.088* Mid5580.040.040.320−0.29−1.512 Late272650.160.163.247**−0.17−1.013Percentage male323930.150.153.273**−0.00−0.507F(1, 391) = 0.257Type of sampleF(2, 660) = 0.020 General375380.120.122.951** Clinical3420.120.120.8540.000.009 Mixed5830.100.100.919−0.02−0.195StudentsF(1, 630) = 5.871* Not students9630.300.293.903*** Students325690.090.092.368*−0.21−2.423*Data characteristicsType of passive data(Raw sensor data)F(13, 645) = 2.464** GPS221790.200.204.270*** Accelerometer11570.130.132.532*−0.07−1.869+ Sleep data8220.130.132.052*−0.07−1.351 Call logs11830.170.173.305**−0.03−0.915 Texting5240.120.121.637−0.08−1.161 Contacts (number)3240.060.060.988−0.14−2.784** Bluetooth4100.130.131.832+−0.06−1.007 Lock/unlock9250.120.121.866+−0.08−1.597 Phone sessions4320.180.182.933**−0.02−0.472 Screentime15860.150.153.099**−0.05−1.424 Notifications140.120.121.036−0.08−0.714 App usage4260.120.121.913+−08−1.698+ Keystrokes4280.040.040.460−0.16−1.927+ Linguistic marker359−0.32−0.31−2.954**−0.51−4.693*** Program settingF(2, 656) = 0.174 Combined233210.100.101.901+ Android182620.140.142.635**0.040.577 IOS6760.150.152.401*0.050.570Type of passive data(Low-level behavioral features)F(19,639) = 3.282** Location context16410.110.112.124* Mobility7380.140.142.625*0.030.563 Location variability12800.130.132.994**0.020.426 Connect context2190.400.382.504*0.291.777 Activity intensity6350.060.061.130−0.5−0.888 Step activity6230.100.102.339*−0.1−0.219 Sleep duration512−0.00−0.00−0.000−0.11−1.559 Sleep distribution3100.230.232.321*0.121.181 Call interaction12970.050.051.236−0.06−1.100 Text interaction5240.020.020.245−0.09−1.201 Proximity4100.060.060.787−0.05−0.725 Contact count2100.030.030.305−0.08−0.905 App use4260.070.071.343−0.03−0.538 Screen time15860.090.092.197*−0.02−0.304 Phone session4320.110.111.969*00.005 Lock/unlock9240.050.050.947−0.05−0.842 Notifications150.100.100.919−0.01−0.065 Text intensity3140.530.494.796***0.423.486** Text variability114−0.30−0.29−2.921**−0.41−3.848*** (Emotional) word use3590.020.020.155−0.09−0.788Type of passive data(High-level behavioral marker)F(4, 654) = 1.999+ Physical262360.160.163.772*** Sleep8220.110.111.782+−0.05−1.025 Social interaction131410.110.112.518*−0.05−1.599 Phone engagement191730.140.143.123**−0.02−0.899 Cognitive-emotional687−0.04−0.04−0.513−0.20−2.578*Outcome characteristicsMental health outcomeF(10, 652) = 0.596 Stress11740.160.163.410*** Anxiety17710.120.122.577*−0.04−1.135 Depression313080.100.102.409*−0.06−1.828+ Loneliness6450.100.102.038*−0.06−1.310 Positive Feelings5260.130.132.245*−0.03−0.635 Negative Feelings6180.110.111.746+−0.05−0.889 Internalizing7640.140.142.086*−0.02−0.267 Externalizing250.190.191.768+0.030.304 Addiction3300.110.111.824+−0.05−1.075 Psychosis2160.070.071.145−0.09−1.615 Suicide160.310.301.2930.150.610InstrumentF(1, 661) = 0.685 Questionnaire (RC)436460.120.123.383*** EMA/ESM4170.050.050.529−0.08−0.827Predictor assessmentF(4, 658) = 0.639 Daily6990.050.050.520 Weekly317−0.01−0.01−0.076−0.06−0.362 Interval61790.060.060.6960.010.052 Study period283610.150.163.350***0.100.933 Single measurement270.130.130.7600.080.405Outcome assessmentF(4, 658) = 1.307 Daily5930.000.000.000 Weekly3150.120.120.8320.120.656 Interval163130.110.111.748+0.110.860 Pre-post measurement575−0.03−0.03−0.237−0.03−0.166 Single measurement161670.210.213.398***0.211.679+k number of independent studies. #ES number of effect sizes, B0/Fisch. z intercept/mean effect size (z), t0 difference in mean r with zero, β1 estimated regression coefficient, t1 difference in mean r with reference category, F(df1, df2) omnibus test, RC reference category.+p < 0.10; *p < 0.05; **p < 0.01; ***p < 0.001. Table 3Results of moderator analyses for the association between passive smartphone data and mental health outcomesModeratork#ESΒ0/Fisch. zrt0β1t1F(df1, df2)Study design characteristicsImpact factor404680.120.122.796**0.000.598F(1, 466) = 0.357Publication year456630.120.123.236**0.000.072F(1, 661) = 0.005PublishedF(1, 661) = 0.040 Published404680.120.122.986** Not published51950.140.141.2660.020.199RegionF(2, 660) = 2.318+ Northern America344820.100.102.554* Europe61230.030.030.342−0.07−0.651 Asia5580.320.313.058**0.221.946+DesignF(1, 661) = 0.000 Cross-sectional5240.120.121.631 Longitudinal426390.120.123.236**0.000.004Time between assessments426390.110.112.690**−0.00−0.424F(1, 637) = 0.180Risk bias assessment (index)456630.110.112.906**0.051.433F(1, 661) = 2.0532Sample characteristicsMean age344410.040.040.1070.010.351F(1, 439) = 0.123Adolescence periodF(2, 438) = 1.175 Early3180.330.322.088* Mid5580.040.040.320−0.29−1.512 Late272650.160.163.247**−0.17−1.013Percentage male323930.150.153.273**−0.00−0.507F(1, 391) = 0.257Type of sampleF(2, 660) = 0.020 General375380.120.122.951** Clinical3420.120.120.8540.000.009 Mixed5830.100.100.919−0.02−0.195StudentsF(1, 630) = 5.871* Not students9630.300.293.903*** Students325690.090.092.368*−0.21−2.423*Data characteristicsType of passive data(Raw sensor data)F(13, 645) = 2.464** GPS221790.200.204.270*** Accelerometer11570.130.132.532*−0.07−1.869+ Sleep data8220.130.132.052*−0.07−1.351 Call logs11830.170.173.305**−0.03−0.915 Texting5240.120.121.637−0.08−1.161 Contacts (number)3240.060.060.988−0.14−2.784** Bluetooth4100.130.131.832+−0.06−1.007 Lock/unlock9250.120.121.866+−0.08−1.597 Phone sessions4320.180.182.933**−0.02−0.472 Screentime15860.150.153.099**−0.05−1.424 Notifications140.120.121.036−0.08−0.714 App usage4260.120.121.913+−08−1.698+ Keystrokes4280.040.040.460−0.16−1.927+ Linguistic marker359−0.32−0.31−2.954**−0.51−4.693*** Program settingF(2, 656) = 0.174 Combined233210.100.101.901+ Android182620.140.142.635**0.040.577 IOS6760.150.152.401*0.050.570Type of passive data(Low-level behavioral features)F(19,639) = 3.282** Location context16410.110.112.124* Mobility7380.140.142.625*0.030.563 Location variability12800.130.132.994**0.020.426 Connect context2190.400.382.504*0.291.777 Activity intensity6350.060.061.130−0.5−0.888 Step activity6230.100.102.339*−0.1−0.219 Sleep duration512−0.00−0.00−0.000−0.11−1.559 Sleep distribution3100.230.232.321*0.121.181 Call interaction12970.050.051.236−0.06−1.100 Text interaction5240.020.020.245−0.09−1.201 Proximity4100.060.060.787−0.05−0.725 Contact count2100.030.030.305−0.08−0.905 App use4260.070.071.343−0.03−0.538 Screen time15860.090.092.197*−0.02−0.304 Phone session4320.110.111.969*00.005 Lock/unlock9240.050.050.947−0.05−0.842 Notifications150.100.100.919−0.01−0.065 Text intensity3140.530.494.796***0.423.486** Text variability114−0.30−0.29−2.921**−0.41−3.848*** (Emotional) word use3590.020.020.155−0.09−0.788Type of passive data(High-level behavioral marker)F(4, 654) = 1.999+ Physical262360.160.163.772*** Sleep8220.110.111.782+−0.05−1.025 Social interaction131410.110.112.518*−0.05−1.599 Phone engagement191730.140.143.123**−0.02−0.899 Cognitive-emotional687−0.04−0.04−0.513−0.20−2.578*Outcome characteristicsMental health outcomeF(10, 652) = 0.596 Stress11740.160.163.410*** Anxiety17710.120.122.577*−0.04−1.135 Depression313080.100.102.409*−0.06−1.828+ Loneliness6450.100.102.038*−0.06−1.310 Positive Feelings5260.130.132.245*−0.03−0.635 Negative Feelings6180.110.111.746+−0.05−0.889 Internalizing7640.140.142.086*−0.02−0.267 Externalizing250.190.191.768+0.030.304 Addiction3300.110.111.824+−0.05−1.075 Psychosis2160.070.071.145−0.09−1.615 Suicide160.310.301.2930.150.610InstrumentF(1, 661) = 0.685 Questionnaire (RC)436460.120.123.383*** EMA/ESM4170.050.050.529−0.08−0.827Predictor assessmentF(4, 658) = 0.639 Daily6990.050.050.520 Weekly317−0.01−0.01−0.076−0.06−0.362 Interval61790.060.060.6960.010.052 Study period283610.150.163.350***0.100.933 Single measurement270.130.130.7600.080.405Outcome assessmentF(4, 658) = 1.307 Daily5930.000.000.000 Weekly3150.120.120.8320.120.656 Interval163130.110.111.748+0.110.860 Pre-post measurement575−0.03−0.03−0.237−0.03−0.166 Single measurement161670.210.213.398***0.211.679+k number of independent studies. #ES number of effect sizes, B0/Fisch. z intercept/mean effect size (z), t0 difference in mean r with zero, β1 estimated regression coefficient, t1 difference in mean r with reference category, F(df1, df2) omnibus test, RC reference category.+p < 0.10; *p < 0.05; **p < 0.01; ***p < 0.001. Results of moderator analyses for the association between passive smartphone data and mental health outcomes Type of passive data (Raw sensor data) Type of passive data (Low-level behavioral features) Type of passive data (High-level behavioral marker) k number of independent studies. #ES number of effect sizes, B0/Fisch. z intercept/mean effect size (z), t0 difference in mean r with zero, β1 estimated regression coefficient, t1 difference in mean r with reference category, F(df1, df2) omnibus test, RC reference category. +p < 0.10; *p < 0.05; **p < 0.01; ***p < 0.001. The Egger’s test and two funnel plots were conducted to further examine the possibility of publication bias. Both the Egger’s test (z = –0.779, p = 0.44) and the funnel plot symmetry test (z = –0.558, p = 0.58) were not significant, indicating no evidence of publication bias. The trim-and-fill test showed that no supplementation of studies was required at either the left or right side of the funnel plot. Furthermore, two funnel plots were generated: one depicting within-study effect sizes against their standard error, and the other depicting between-study mean effects against their standard error. Both plots are provided in Supplementary Figs. 1 and 2. Discussion This meta-analysis examined the association between passive smartphone data and mental health outcomes among adolescents, accounting for study design, sample, passive data and outcome characteristics (i.e., type of outcome). Results showed a small but significant association between passive smartphone data and mental health outcomes (r = 0.12), indicating that passive data constitute a modest yet meaningful indicator of mental health difficulties. Moderator analyses showed that the association was weaker in students than in non-students, while the number of contacts was non-significant. Linguistic markers and keystroke-derived text variability were unreliable (inaccurate) predictors of mental health, whereas text intensity proved to be a strong predictor. Region, age, clinical status, and type of outcome did not moderate the association between passive smartphone data and mental health outcomes, indicating the broad applicability of such data as an indicator of adolescents’ mental health. The overall small but significant effect observed in this study aligns with a broader pattern reported in related work. Meta-analytic studies in adult populations have consistently shown small to modest associations between passively collected smartphone data and diverse psychological outcomes, including personality traits54, post-traumatic stress disorder50 and depressive symptoms49. The recurrence of these small effects across domains suggests that passive sensing holds promise for detecting and advancing the understanding of mental health outcomes. While the magnitude of the main effect is comparable to that reported in adult studies, our moderator analyses indicate that the contribution of specific passive markers differs across data streams. These differences occur against a background of generally modest predictive power, with linguistic markers and text variability in particular showing inconsistent performances. Passive measures, therefore, require further refinement, for example, by using context-sensitive language models55, larger and more diverse datasets56, and by combining linguistic with other smartphone data to improve their predictive power, validity, and reliability in detecting mental health outcomes57. More specific results regarding the type of passive data could be detected from the moderator analyses. GPS-based measures provided the most robust signals, confirming earlier work in adults where reduced location variability and spending more time at home were linked to higher depressive severity58–60. Associations with anxiety, stress, and suicidal risk have also been reported, although these findings are weaker and less consistent58,60,61. Notably, GPS is among the few passive indicators frequently studied in isolation, which may help explain its comparatively robust predictive value. At the same time, prediction improves when GPS is integrated with other data streams such as sleep or communication features62. Social interaction (e.g. call logs and texting) and phone-engagement features (e.g. notification, screen time, phone sessions), also showed significant but more modest associations. Unlike GPS, which directly reflects mobility patterns, these indicators capture indirect behaviors such as frequency of communication or device engagement21,39. Their predictive value may be weaker and less stable, which could explain why they are often studied in combination with other features rather than in isolation. For example, call and text logs have often been examined together, both in relation to overall depression severity and to more specific outcomes such as suicidal ideation63,64. Phone-engagement features are often studied in combination (e.g., screen time with number of unlocks), where frequent unlocks predicted fewer externalizing symptoms but only at low levels of screen use65. Finally, studies combining social interaction and phone-engagement features (e.g., call logs, battery charging, screen checking) suggest potential for capturing impulsivity traits66. These findings highlight that communication and phone-use features are weaker than GPS as stand-alone predictors, yet they gain explanatory value when studied in combination with other data streams or when examined at a more granular level. We found that number of contacts, indicative of social (dis)connectedness, was not significant, suggesting that the amount of online contact may not be a reliable indicator of mental health. This aligns with broader evidence that the quality of relationships, such as support and trust, is more relevant than the sheer quantity of contacts67,68. Passive measures that capture richer aspects of social interaction, such as conversational content or in-person meetings, may therefore provide more meaningful indicators16,69. Linguistic markers showed similarly inconsistent and even negative associations. These raw language-based features, such as emotional tone or speech patterns derived from text analysis or natural language processing, are difficult to interpret because the meaning of language is highly context- and aged-dependent70. Prior work in adult populations has shown that linguistic markers are strongly influenced by context and task71. Research in adolescent populations further suggests that language use is more closely tied to momentary mood and internalizing symptoms, reflecting ongoing socio-emotional development72. This contextual and developmental sensitivity of language use may explain the observed inaccuracy of linguistic markers. A similar limitation appears to apply to keystroke-derived text variability at the level of low-level features, which also proved to be ad inaccurate, that is, unreliable predictor of mental health. In contrast to text intensity, which demonstrated consistent associations, text variability may be shaped by contextual and state-dependent factors73,74, probably capturing a range of underlying psychological processes. This complexity suggests that variability-based measures may be less informative when interpreted in isolation73,74. Consequently, the application of linguistic and text variability markers as indicators of psychopathology likely requires approaches that preserve contextual meaning, such as discourse analysis75–77. Such analyses could be facilitated by AI-based algorithms that are regularly updated to account for the evolving meanings of words across different social contexts. Results showed that the association between passive data and mental health outcomes was significantly stronger in non-student than in student samples. This may reflect greater heterogeneity in non-student populations compared to the relative homogeneity of student samples78, creating a potential artifact due to restricted variance. Nevertheless, the association in the student sample, although weaker, remained significant. Differences may also relate to the timing of onset of mental health difficulties, as one-third of individuals tend to experience a mental disorder before age 14, and nearly half by age 1879. This suggests that mental health problems may occur less frequently in student populations, which negatively affects the predictive power. Student samples also to differ from non-student populations in broader sociodemographic characteristics, including higher socioeconomic status and higher parental education, both of which are protective factors for mental health outcomes80. Furthermore, student populations are often more female-dominated than non-student populations81,82, which may represent another relevant population characteristic warranting further investigation. Exploring these population differences more systematically could help clarify these findings, and their implications for the generalizability of passive data across populations. Current findings have implications for the development and use of passive smartphone data in adolescent mental health care. Identifying individual behavior patterns through passive data may support the early detection of mental health problems37,83. The small but significant associations observed across different types of mental health suggest that passive data capture a broad and generalizable signal, making them a useful complementary tool in clinical care. Moreover, the affordability of smartphone-based monitoring offers opportunities for scalable early detection and linkage to preventive or low-intensity interventions84,85. In turn, these approaches may help reduce the treatment gap, which refers to the discrepancy between those who need mental health care and those who receive it, by reaching underserved adolescents at risk for, or already experiencing, mental health problems23,86,87. The current study has several strengths. First, to our knowledge, this is the first meta-analysis to investigate multiple types of passive smartphone data in relation to different types of mental health, addressing the absence of an overall synthesis, a gap repeatedly highlighted in systematic reviews in this field29,46,83. Another important strength is the use of a three-level meta-analytic model, which allowed us to examine variation both within and between studies, across study designs, sample characteristics, parameter types and mental health outcomes. The large number of studies and effect sizes improved precision, enabling a more reliable estimate of the overall association between passive data and mental health outcomes in adolescents and its moderators. Another strength is the exclusive focus on adolescents aged 12–24 years, a group that remains mostly underrepresented in reviews of passive sensing research88. This age range spans a critical developmental period characterized by substantial cognitive, emotional, and social change, during which the meaning and interpretation of passive digital markers may differ from those observed in adult populations. By focusing specifically on this group, the present study helps to fill a critical gap in evidence on passive mobile data and adolescent mental health. While the findings of this study offer valuable insights, several methodological limitations should be considered. First, the number of included effect sizes varied widely between categories. This discrepancy affects generalizability. Conclusions based on few studies of a low number of effect sizes must be interpreted with caution, and result in low statistical power to detect effects in particular subgroup analyses. Second, complex passive data streams (e.g., GPS) were aggregated into broader analytic categories to enable quantitative synthesis across studies, which may have further reduced sensitivity to feature-specific effects within these subgroups. Third, the studies included in this meta-analysis predominantly focused on late adolescence and young adulthood, reflecting the relative scarcity of research in early and mid-adolescence. As a result, caution is needed in the interpretation and generalizability of the findings. Fourth, ethical aspects of study design, such as parental involvement and consent procedures for passive data collection, could not be examined as moderators due to limited and inconsistent reporting across studies, but remain important considerations for future research. Fifth, although no indication of publication or selection bias was found, such bias can never be fully ruled out in meta-analysis. Pre-registration of primary studies remains the best safeguard against this risk. Last, as in many meta-analyses, heterogeneity in study design, measures of passive data, and definitions of mental health outcomes may have influenced the pooled estimates. Differences in how passive data were collected or how outcomes were defined could have contributed to unexplained variance. While some of these differences were captured in moderator analyses, they underscore the need for consistent reporting and more standardized approaches in future meta-analyses. In conclusion, the present meta-analysis examined the association between passive smartphone data and mental health outcomes in adolescents. Using a three-level model, we identified moderators at the level of study design, sample characteristics, and parameter type. Results showed that the number of contacts was not a significant predictor, and linguistic markers proved to be unreliable, while associations were weaker in student compared to non-student samples. Overall, the effect was small but significant. These findings suggest that passive smartphone data may offer value as a complementary tool for early detection and monitoring of mental health outcomes in adolescent mental health care. To advance this field, future research should focus on refining which data types are most informative, ensuring standardized measurement, and testing integration of passive monitoring in clinical or curative and preventive settings. Methods The protocol for this study has been registered at PROSPERO (CRD42025596956). We followed the PRISMA 2020 statement for reporting this review53. The full PRISMA checklist is provided in Supplementary Table 1. Eligibility criteria Studies had to meet the following criteria to be included in the review: (1) participants are aged between 12 and 24 years, reflecting a developmental definition of adolescence that encompasses early, middle, and late adolescence51,52. This age range was selected to capture multiple developmental phases relevant to mental health while aligning with the age at which personal smartphone ownership becomes common89,90, enabling the use of passive smartphone sensing; (2) longitudinal or observational design, to capture variation in mental health outcomes over time; (3) objective measure of passive smartphone sensing as the predictor of interest (4) quantitative mental health assessment, either through questionnaire or ecological momentary assessment/experience sampling methods (EMA/ESM); (5) peer-reviewed publication between 1992 and 2025, with 1992 with 1992 chosen as a conservative lower bound corresponding to the release of the first smartphone91; and (6) written in English or Dutch. Both clinical and community samples, including student populations, were eligible. Studies were excluded when passive smartphone sensing data could not be coded separately, for example, when passive and active data were combined within a single predictive model without separate reporting of passive effects. Additional exclusion criteria included studies showing major limitations in the reliability or validity of assessments, such as very small sample sizes, high risk of bias, insufficient data reporting, or lack of quality control. We also excluded studies without an explicit quantitative measure of the association of interest, such as case studies, qualitative work, study protocols, reviews, conference proceedings, and books. Finally, studies using passive data from devices other than smartphones, including wearables, laptops, or tablets, were also excluded. Search strategy Relevant articles were retrieved from PubMed, Medline, Scopus, Embase, and Web of Science (April 16th, 2025). Results for each database are provided in Supplementary Table 2. The search combined terms across four categories: (1) passive smartphone data, (2) psychopathology or mental health outcomes, (3) mobile phones, and (4) adolescents. The full search strategy is provided in the Supplementary Tables 3–7. Additional relevant studies were identified by screening the reference lists of excluded but related articles and by reviewing the publication records of leading researchers in digital phenotyping and passive mobile sensing (e.g., J. Torous, J.P. Onnela, K. Huckvale). Study selection All studies were imported into the electronic database. Duplicates were first removed automatically by the system, followed by manual screening to identify additional duplicates. The first 60 studies were collaboratively screened by the first and second authors in collaboration with research interns on title and abstract, using the predefined inclusion and exclusion criteria. The remaining studies were screened independently and in a blinded manner by two reviewers to maximize interrater reliability. The initial interrater agreement was 81%. The final decision was reached through discussion among the researchers. Data extraction A dataset was created in IBM SPSS Statistics version 29.0 to extract relevant information from the included studies. To enhance transparency, the complete dataset has been made publicly available via OSF (https://osf.io/ns9rd/overview). We collected within- and between-study characteristics that could potentially moderate the association between passive sensing and mental health outcomes. Moderator categories included study design, sample characteristics, raw types of passive data, behavioral features derived from passive data, and outcome characteristics. Following the frameworks of Mohr et al.39 and Vlisides-Henry et al.16, we distinguished between raw sensor signals and higher-order behavioral features, treating these as two separate subcategories. Figure 3 shows how the raw sensors connect to high-level behavioral markers through intermediate low-level features. Some sensors, such as Bluetooth and texting, can contribute to multiple categories. To ensure consistency in this meta-analysis, however, each raw sensor was assigned to only one category, based on its most common application in the included studies. The full operationalization from study-level raw sensor data to behavioral markers is provided in the Supplementary Tables 8–23. Mental health outcomes were used as an umbrella term encompassing all identified categories of mental health problems, which were grouped into related domains92. In line with Keyes’ dual-continuum model93, we defined mental health outcomes broadly to include both problems, such as depression and anxiety, and well-being indicators such as stress and loneliness.Fig. 3Overview of sensor data and behavioral markers.This figure illustrates the connection between raw sensors and high-level behavioral markers through intermediate low-level features. It provides a visual representation of the data processing layers defined in this research. Overview of sensor data and behavioral markers. This figure illustrates the connection between raw sensors and high-level behavioral markers through intermediate low-level features. It provides a visual representation of the data processing layers defined in this research. The first five studies were coded collaboratively by ML, AH and IS under the supervision of GJS. The remaining studies were evenly distributed coded and double-checked by an independent researcher. Any disagreements or uncertainties were resolved through discussion, and when necessary, authors of the included studies were contacted to clarify unclear information or to provide missing details. Effect measures Each study’s effect size was converted to Pearson’s r, as measure for the association between passive smartphone sensing data and mental health outcomes. Correlations were coded according to the expected direction of the association, with values that did not align recoded to the opposite sign. For example, if increased GPS use was reported to relate to lower levels of depression, the correlation was scored as positive. All correlations were transformed to Fisher’s Z prior to calculating pooled effects and conducting moderator analyses94. For the reporting and interpretation of the results, the Fischer’s Z values were once again converted into Pearson’s r values. Effect sizes were interpreted following established guidelines: r < 0.05, tiny; 0.05–0.10, very small; 0.10–0.20, small; 0.20–0.30, medium; 0.30–0.40, large; and ≥0.40, very large95. Synthesis method Analyses were conducted in R (version 4.0.0) and SPSS (version 28). A multilevel meta-regression was used to estimate the overall association between passive smartphone sensing and mental health outcomes, and to test whether study or sample characteristics moderated this relation. The model accounted for three sources of variance: sampling variance of the observed effect sizes (level 1), variance between effect sizes within studies (level 2), and variance between studies (level 3)96–98. Multilevel models were fitted using the metafor package in R98. Regression coefficients were evaluated using a t-distribution, and omnibus F-tests were applied to categorical moderators with three or more groups. Statistical significance was defined as p < 0.05. Outlying effect sizes were winsorized, meaning they were recoded to fall within the normal range, defined as values within ±3.29 standard deviations of the mean99. To visualize the average associations, forest plots were generated following published guidelines tailored to multilevel meta-analysis, providing a graphical summary of effect sizes and confidence intervals100. Reporting for biases Publication bias, or the file-drawer problem, refers to the overestimation of true effects because studies with significant results are more likely to be published than those with null findings101. To address this, we first inspected funnel plots, which plot effect sizes against their standard errors100. In the absence of bias, studies should form a symmetrical funnel around the overall estimate, whereas asymmetry may indicate selective reporting. Missing values on the right side of the funnel suggest selection bias, whereas missing values on the left suggest publication bias. Three methods were applied: Egger’s regression test to statistically evaluate asymmetry102, an adapted funnel plot test for visual inspection in multilevel meta-analysis103, and trim-and-fill analyses to estimate potentially missing studies103,104. In addition, a risk of bias assessment was conducted at the study level. Based on established frameworks105–109, five key criteria were evaluated: (1) measurement validity, (2) measurement reliability, (3) same-measurement variance, (4) shared informant variance, and (5) drop-out bias. The risk of bias assessment for all included studies is provided in Supplementary Tables 24–26. Supplementary information 20260317Supplementary Information 20260317Supplementary Information