As the World Health Organization (WHO) notes, mental well-being is “crucial to personal, community, and socioeconomic development” [1]. Mental well-being, or positive mental health, is more than just the absence of mental illness. It also includes having positive psychological resources [2]. Integrating hedonic and eudaimonic perspectives, researchers have highlighted 3 aspects of mental well-being, namely positive affect (eg, optimism, cheerfulness, and relaxation), positive functioning (eg, energy, clear thinking, and competence), and satisfying interpersonal relationships (a sense of belonging and being loved) [3]. Naturally, mental well-being fluctuates over time, and these dynamics typically manifest in daily life [4,5]. They are reflected in and can thus be inferred from everyday behavioral patterns [6]. Previous research in this area relied on self-reports and laboratory experiments [7-9]. However, self-reports are vulnerable to recall and social desirability biases [10], while laboratory paradigms lack ecological validity [11]. Mobile sensing, which passively collects usage records and sensor data from smartphones, helps overcome both limitations by continuously and unobtrusively sampling behavior, providing a more accurate reflection of people’s daily lives [12]. With smartphones now in the pockets of over 3‐quarters of the world’s population [13], capturing these fine‐grained contours of daily life [14,15] has become feasible, allowing us to access app‐use durations, GPS‐derived mobility, and ambient light or sound levels that may reflect well‐being [16,17].

To date, most work in digital mental health has pursued two primary paths: (1) correlational designs that link individual phone-use logs or sensor data to mental well-being or illness assessed by self-report measures or diagnostic interviews [18-27]; or (2) machine learning models aimed at maximizing predictive accuracy of mental health indicators from high‐dimensional sensor inputs [28-32]. While valuable, these strategies are primarily population-level because they often focus on average between-person relationships that can hide the rich heterogeneity of individual experiences [14,33]. In particular, correlational designs and many machine-learning pipelines, especially those trained on temporally aggregated features, can overlook how combinations of behaviors co-occur within the same day, information that is crucial for identifying well-being-related lifestyles and designing personalized interventions [34-36].

Recently, a growing consensus emerged that mental well-being is rarely driven by any single behavior in isolation; what matters are combinations of co-occurring behaviors and the ways they recur as patterns [37,38]. In other words, it is about the synergy, not the dose, of certain behaviors, and mental well-being is likely reflected in the patterned ensemble of behaviors [39]. For example, researchers have found that social engagement [40], mobility patterns [41], physical activity [42], phone use [43], and the environment [44] predict mental health, not in isolation but in combination. Measuring these indicators in real-world settings and using a person-centered modeling approach enables researchers to identify distinct daily behavioral patterns. Furthermore, by examining how individuals assemble and repeat these daily patterns, we can identify broader profiles of lifestyles [16,17] and see which lifestyles demonstrate better mental well-being.

A person-centered approach seeks to identify subpopulations that exhibit distinct patterns across certain variables [45]. Latent profile analysis (LPA) is particularly well suited to this purpose [46]. It is a statistical method that estimates “latent” (ie, not directly observed but inferred) profiles for each unit (eg, one day) and—based on these profiles—groups units that exhibit specific levels on observed measures [47]. LPA aims to identify groups of units that show similar patterns on these measures. For a detailed methodological illustration, see S1 in Multimedia Appendix 1. In digital phenotyping, selecting the unit to be profiled is consequential. A conventional single-level LPA would typically aggregate sensing indicators across the observation period and classify individuals according to their average or total levels of these indicators. Although this approach can identify broad between-person differences, aggregation removes information about how behaviors are organized within individual days and how different types of days recur within the same person. It may also introduce ecological fallacy if associations observed for person-level averages are interpreted as reflecting behavioral configurations occurring within individual days [48]. Two persons may have identical 2-week averages but arrive at those averages through different daily configurations. For example, one person may consistently display moderate mobility and phone use, whereas another alternates between highly active, socially engaged days and sedentary, digitally intensive days. A single-level analysis of aggregated indicators may place these individuals in the same profile, despite potentially meaningful differences in the regularity, variability, and composition of their lifestyles. Profiling days before profiling persons therefore preserves information that is lost when observations are immediately collapsed to person-level values. Usually, repeated digital traces (eg, app-use sessions) are nested within individuals, yielding a hierarchical structure: daily behavioral patterns (Level 1) [12] are nested within persons, and across days, these patterns aggregate into individual lifestyles (Level 2) [49]. A 2-level LPA can capture this structure well. At Level 1, multivariate sensing features from each day can be clustered into daily behavioral profiles; at Level 2, individuals can be clustered by their distribution across these daily profiles to yield higher-order personal profiles [50]. This hierarchical approach has been widely used across organizational, educational, affective, and psychopathological research and has proven effective for identifying subpopulations that differ in distal outcomes (eg, mental health and academic performance) [51-55]. However, to date, no study has applied this approach to identifying behavioral patterns captured via smartphones. By moving beyond variable-centered correlations and single-stage pipelines that operate at a single timescale (eg, collapsing days into person-level summary features before modeling or modeling daily events without considering person clusters), 2-level LPA yields interpretable lifestyles that may offer useful insights for future personalization efforts, positioning it to substantially advance digital mental health research.

Additionally, theory suggests that person-behavior fit plays a crucial role in well‐being. Individuals thrive when their activities align with their personalities [56,57]. For example, extraverts may flourish on highly social, connected days, whereas conscientious individuals might derive greater satisfaction from structured, goal‐driven routines [58,59]. Thus, personality traits may be integrated into such latent profile frameworks to test whether different personality traits are associated with mental well-being in the same way across different lifestyles and whether individuals with lower or higher levels of these traits exhibit different levels of mental well-being across profiles.

Implementing this approach, the present study used a 2-level LPA on 2 weeks of phone-use logs and sensor data capturing social engagement, mobility, physical activity, phone usage, and environment in a German adult sample. At Level 1, we uncovered recurring daily behavioral phenotypes; at Level 2, we identified personal profiles based on the distribution of daily profile proportions. The study had three objectives: (1) to determine whether passively collected smartphone-sensing data could be used to derive interpretable day-level and person-level lifestyle profiles; (2) to examine whether the resulting person-level profiles differed in overall mental well-being and its dimensions of positive affect, positive functioning, and satisfying interpersonal relationships; and (3) to test whether Big Five personality traits—extraversion, agreeableness, conscientiousness, openness, and negative emotionality—moderated these associations.

The data used in the current study were obtained from the Smartphone Sensing Panel Study (SSPS), a prospective longitudinal cohort study conducted by Ludwig-Maximilians-Universität München (LMU Munich) in cooperation with the Leibniz Institute for Psychology [60]. The SSPS collected data from May 2020 to November 2020. Participants were asked to install the self-developed Android-based mobile sensing app PhoneStudy (LMU Munich) on their private smartphones, which continuously collected various data in the background throughout the respective study duration. Online surveys, experience sampling, and smartphone-sensing data were collected.

The original SSPS aimed to recruit 800 participants. This target was determined pragmatically by balancing the planned study duration, expected attrition, and participant compensation costs. Based on an anticipated monthly dropout rate of 7%‐18.5%, approximately 430 participants were expected to remain after 6 months. Ultimately, 850 participants installed the application and participated in the study, of whom 750 completed the initial survey containing demographic information in May 2020. All procedures adhered to the General Data Protection Regulation. Reporting follows the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology [61]) guidelines for observational studies.

In the present study, we used questionnaire data from the July 2020 survey, which measured mental well-being and Big Five personality traits. A total of 636 participants completed this wave of questionnaires, from whom we extracted sensing data. We excluded 8 participants with no sensing records and 16 who changed phones during data collection, leaving 612 participants. Sensing data were extracted up to the day before each user filled out the questionnaire (July 12, 2020‐July 25, 2020). Because mental well-being was assessed from each participant’s subjective evaluation over the preceding 2 weeks, we first extracted the most recent 14 days of data for everyone. We then applied the 4/7 validity rule [62], which required at least 4 days of valid data within any 7-day window and corresponded to a minimum of 8 valid days over the 14-day period. A day was classified as invalid when all sensing variables were either zero or missing. Participants with fewer than 8 valid days were excluded, leaving a final sample of 553 individuals (mean age 42.12, SD 12.89 years; 44.65% female) with 7635 days of sensing data.

Across 553 participants, we sampled a total of 7635 participant-days of sensing data. Although there is no consensus on a minimum sample size for multilevel latent profile models, sample size planning rules of thumb formulated by Park and Yu [63] recommend at least 20 Level 2 groups and 10 Level 1 units. From this sample, 495 participants provided their demographic information. Age ranged from 18 to 65 years, with an average of 42.12 (SD 12.89) years, and 44.65% (n=224) of participants indicated they were women, and 55.56% were men (n=220). Regarding education, approximately 35% had an intermediate secondary school qualification, 30% had a higher education entrance qualification, 20% had a university or higher education degree, and 14% had a lower secondary school qualification; only a small proportion reported having no school-leaving qualification or a doctorate or habilitation. Regarding employment status, 69% were employed, 10% were in education or training, 9% were not in the labor force, 7% had family or household responsibilities, and 5% were unemployed or seeking work, while only a very small proportion were engaged in voluntary or civic service. Regarding income, 20.6% reported €1501-€2000 (€1=US $1.1770 as of July 27, 2020), 18.6% reported €501-€1000, 17.6% reported €2001-€2500, 14.3% reported €1001-€1500, 6.7% reported €2501-€3000, 6.5% reported €3500 or more, 5.5% reported €3001-€3500, 5.3% reported less than €500, and 5.1% reported no income. See Figures 1A and 1B for an overview of the demographic characteristics of the sample.

Before recruitment began, quotas were established to approximate the German population in terms of age, gender, education, income, religious status, and relationship status [60]. These quotas guided the initial recruitment process. However, as recruitment progressed, some quota targets proved difficult to achieve, and the restrictions were subsequently relaxed to facilitate enrollment. Consequently, although recruitment was initially guided by the prespecified quota framework, the final sample did not strictly reflect the intended quota distribution.

No a priori power analysis was performed because there is no established method for determining statistical power for multilevel LPA. Instead, we considered published recommendations for sample size in multilevel latent profile models. Although no consensus exists regarding the minimum required sample size, Park and Yu [63] recommended a minimum of 20 Level 2 units and 10 Level 1 units per Level 2 unit as a general guideline. In this study, Level 2 units corresponded to participants, and Level 1 units corresponded to daily observations. The final analytic sample included 553 participants and 7635 daily observations (mean 13.81, SD 0.83 days per participant), substantially exceeding these recommended minimum sample size guidelines.

This study used a 2-week intensive longitudinal observational design based on data from the SSPS. Participants installed the PhoneStudy app on their personal Android smartphones, which passively and continuously recorded smartphone use and sensor data during the study period. For the present analysis, sensing data from the 14 days preceding each participant’s completion of the mental well-being and personality questionnaires were examined. Daily observations were nested within participants, allowing behavioral patterns to be modeled at both the day and the person level. Specifically, multilevel LPA was used to identify recurring day-level behavioral profiles and higher-order person-level lifestyle profiles based on the distribution of these daily patterns. The resulting person-level profiles were then compared in terms of mental well-being, and Big Five personality traits were examined as potential moderators of the associations between lifestyle profiles and mental well-being. Because the study was observational, the analyses were intended to identify associations rather than establish causality.

The SSPS was reviewed and approved by the Ethics Committee of the Department of Psychology at LMU Munich under the study title “A Longitudinal Panel Study Combining Smartphone Sensing and Survey Methods.” At LMU Munich, the Institutional Review Board (IRB) does not assign approval numbers. Instead, ethical approval is granted and documented under the study title. Before participating, all individuals received detailed information about the study procedures, smartphone-sensing activities, data processing, compensation, and data protection measures in accordance with the General Data Protection Regulation. Participants provided informed consent before installing the PhoneStudy app and before beginning data collection. To protect privacy and confidentiality, participants were assigned randomly generated 8-character pseudonymous identifiers that were used to store and link sensing, experience-sampling, and survey data. Data were transferred to the study server through encrypted connections and analyzed using pseudonymized identifiers. Participants received monetary compensation for individual study activities, including monthly surveys, experience-sampling phases, and continuous smartphone sensing. Compensation increased with the duration of participation. Participants assigned to the 3-month study condition could receive up to €44.50 (€1=US $1.1770 as of July 27, 2020), whereas those assigned to the 6-month condition could receive up to €131.50 for completing all study activities. No images in the manuscript or supplementary materials permit identification of individual participants. Geographic information is presented only in an aggregated manner, and no individual GPS trajectories or precise residential locations are displayed.

The PhoneStudy app recorded many different types of smartphone-sensing data, providing information about participants’ daily lives. To capture the full richness of participants’ daily living patterns, we derived a set of digital phenotyping variables spanning 5 key domains known to be associated with mental well-being in synergy with one another—social engagement on smartphone, mobility, physical activity logs, physical sensor logs, and screen logs—by applying a suite of computational algorithms to raw sensor streams. Sensing features were computed for each person at the day level (12:00 AM-11:59 PM) according to pipelines described by Schoedel et al [69]. To improve LPA convergence and simplify interpretation, we selected 2 features for each of the 5 domains, prioritizing those with minimal missingness and low intercorrelations to avoid redundancy. A brief summary is shown in Table 1.

We quantified social engagement by total time, in seconds, spent in communication apps and social media apps. App categories were derived from the systematic classification developed by Schoedel et al [71], which assigned 3091 commonly used Android apps to 26 psychologically meaningful categories. The classification system was developed iteratively and applied by 2 independent raters; disagreements were discussed and resolved with the involvement of a third rater. Interrater agreement was κ=.71 for communication apps and κ=.63 for social media apps. Communication apps were defined as apps whose primary purpose was direct interpersonal communication, including calling, text messaging, instant messaging, email, and video calling. Social media apps were defined as apps whose primary purpose was sharing, browsing, or interacting with content within an online community. Although many apps provide both functions, each app was assigned exclusively to one category according to its main functionality or primary selling point.

Data collection began in May 2020 and was conducted simultaneously for all participants. After enrollment, participants installed the self-developed Android-based PhoneStudy app on their personal smartphones, which continuously collected passive smartphone-sensing data in the background throughout the study period. Participants also completed monthly online questionnaires assessing demographic and psychological variables. For this study, only the survey assessing mental well-being and personality traits and the corresponding 14 days of smartphone-sensing data preceding survey completion were analyzed.

Smartphone-sensing variables were computed using previously validated processing pipelines. Daily features were derived automatically from raw sensor streams using standardized algorithms, minimizing observer-related measurement error. Self-report measures were administered online using validated German versions of established instruments.

Data were collected using the self-developed Android-based PhoneStudy app and online questionnaires. The PhoneStudy app continuously recorded passive smartphone-sensing data by accessing built-in Android system APIs and device sensors, including app use, GPS, activity recognition, ambient light, microphone, and screen-state information. Self-report data were collected through secure online questionnaires, including the WEMWBS and the BFI-2-XS, both of which are validated German-language instruments. Smartphone-sensing features were derived from the raw sensor streams using previously validated processing pipelines described by Schoedel et al [69].

Because this was an observational study without experimental conditions, masking or blinding of participants, investigators, or outcome assessors was not applicable.

For the German version of the WEMWBS, because the present study examined both the composite score and 3 subdimensions, we first evaluated the factorial validity of the instrument using confirmatory factor analysis. A correlated 3-factor model representing positive affect, positive functioning, and satisfying interpersonal relationships provided a better fit than a one-factor model (see S3 in Multimedia Appendix 1). Internal consistency in the current sample was excellent for the composite score (ω=.95), positive affect (ω=.95), positive functioning (ω=.84), and satisfying interpersonal relationships (ω=.83).

For the German version of the BFI-2-XS, previous validation studies reported Cronbach α coefficients ranging from 0.45 to 0.67, reflecting the brevity of the 3-item scales, and test-retest reliabilities ranging from 0.72 to 0.88 [68]. In the present sample, McDonald ω coefficients were 0.44 for extraversion, 0.58 for agreeableness, 0.55 for conscientiousness, 0.75 for negative emotionality, and 0.50 for openness. These reliability estimates are comparable to those reported in the original validation study and are consistent with expectations for very brief personality measures.

No experimental manipulation was implemented. The study used a nonexperimental, observational, intensive longitudinal design with repeated daily observations nested within participants.

We conducted an attrition analysis to examine whether participants retained in the analysis differed from those excluded between enrollment and analysis. No significant differences were found in age, gender, education, income, religion, or relationship status; full results are reported in S1 in Multimedia Appendix 1. We further compared the demographic composition of the final sample with the initial quota design using chi-square goodness-of-fit tests. The final sample retained the initial quota composition for age, gender, and relationship status, whereas significant deviations were observed for education, income, and religious status. See S2 in Multimedia Appendix 1 for details.

The flow of participants from recruitment to the final analytic sample is presented in Figure 2.

Participant recruitment was conducted between May and November 2020 as part of the SSPS. The questionnaire wave analyzed in the present study was completed in July 2020, and smartphone-sensing data from the preceding 14 days were included in the analyses.

For the current study, 10 features spanning five domains were computed per participant and day: (1) social engagement on smartphones: communication app usage duration, social media app usage duration, (2) mobility: travel distance, radius of gyration, (3) physical activity: walking probability, on-bicycle probability, (4) environment: loudness and brightness, (5) phone-use intensity: screen-on duration and screen session count. Of these 10 features, 8 exhibited high skewness and thus were Box-Cox transformed (see Figure S1 in Multimedia Appendix 1 for details). Subsequently, all variables were z score normalized to obtain standardized values across variable types. See Table 2 for correlational results.

To our knowledge, this is the first study to use a latent profile framework with mobile sensing data to characterize daily behavioral patterns for digital phenotyping. By capturing smartphone-based records of social engagement, mobility, physical activity, environment, and phone-use intensity, we identified 8 daily behavioral profiles and aggregated them into 7 person‐level profiles. In this German adult sample, we found that person-level profiles did not differ in composite mental well-being or 2 of its subscales (ie, positive affect and satisfying interpersonal relationships), but did differ in a third subscale, positive functioning. Specifically, a profile characterized by high physical activity and low digital engagement showed higher levels of positive functioning compared to the profile marked by extensive mobile and social media use. Moreover, we found no evidence that personality moderated these associations in the current sample. The findings primarily demonstrated the feasibility of deriving interpretable multilevel behavioral profiles from mobile-sensing data, whereas evidence that these profiles meaningfully distinguish mental well-being was limited and requires replication.

The overall pattern of findings indicates that the identified lifestyle profiles were only weakly and selectively associated with mental well-being. Across the 7 person-level profiles, we observed no significant differences in the composite mental well-being score, positive affect, or satisfying interpersonal relationships. Moreover, most pairwise comparisons of positive functioning were also nonsignificant. The only significant contrast indicated higher positive functioning in the “active and unplugged” profile than in the “mobile and always-on social” profile. Accordingly, the results do not support that the profile solution consistently distinguishes individuals according to their mental well-being. Instead, they suggest that particular behavioral configurations may be associated with a specific functional dimension of well-being.

Several explanations for the predominance of null findings are possible. First, behavioral patterns captured through smartphones may show only modest and inconsistent associations with broad self-reported well-being, particularly in generally healthy or population-based samples [96]. Second, the same observable behavior may have different meanings across individuals [97]. The sensing indicators captured the duration and frequency of app and screen use but not the content, purpose, reciprocity, or subjective quality of that use. Digital trace volume should therefore not be equated with actual exposure or lived engagement [98]. For example, 60 minutes recorded in a social media app could reflect scrolling, distressing content consumption, reciprocal communication, information seeking, content creation, or leaving an app open without sustained attention [99]. These forms of engagement may have different associations with well-being but cannot be distinguished using app duration, screen-on time, or session frequency alone [100]. Consequently, labels such as “high social media use” describe recorded behavioral volume rather than the psychological meaning or quality of the experience. This measurement gap may partly explain why the profiles showed limited differentiation in mental well-being. Future studies should combine sensing data with content-sensitive measures, experience sampling, or participant reports to address this limitation. Third, positive affect and relationship satisfaction may depend more strongly on the quality and subjective meaning of activities and interactions than on the behavioral quantities captured by the sensing variables [100]. Finally, the data were collected in July 2020, when pandemic-related restrictions, altered work and commuting arrangements, reduced opportunities for in-person activities, and increased reliance on digital communication may have constrained both behavioral variability and the associations between behavior and well-being [101]. The predominance of null findings may therefore partly reflect this unusual temporal context rather than associations that would be expected under postpandemic conditions. These possibilities should be examined directly in future studies rather than assumed to account for the null results.

The difference between the 2 extreme profiles may nevertheless provide a hypothesis for further investigation. Prior work has found that lifestyle was linked to cognitive performance [102], cognitive function [103], and cognitive impairment [104]. Using objectively collected smartphone data, the current study reconfirmed the relevance of lifestyle differences to personal functioning. Specifically, the way people organize their days—including how much they move, how often they check their phones, and how much time they spend on communication and social media apps—relates selectively to positive functioning. The “active and unplugged” profile (person-level profile 5) exhibited higher positive functioning than the “mobile and always-on social” profile (person-level profile 7). One possible explanation is that heavy social app use may fragment attention and deplete cognitive resources, undermining perceived competence and clarity [105,106], whereas physical activity may acutely enhance cognitive performance [107]. However, the present correlational data cannot determine which feature, if any, accounts for the observed difference. The contrast may also reflect unmeasured factors such as occupational demands, commuting, socioeconomic circumstances, physical health, or preexisting functioning. Therefore, interpretations should be regarded as possible explanations for future testing rather than mechanisms demonstrated by this study.

The physically active and digitally unplugged pattern (person-level profile 5) was dominated by days marked by walking/cycling and minimal phone usage (day-level profile 4). Prior work consistently found that reduced smartphone usage was associated with higher levels of well-being, possibly via less digital toxicity and greater engagement with real-world life [108-110]. In addition, the screen logs of people in this profile showed more frequent activity but shorter durations. This pattern indicates how often and for how long the phone was active, but it does not reveal whether these sessions involved purposeful task completion, reciprocal communication, passive browsing, or another form of engagement. Possibly, this combination could suggest purpose-driven interactions (brief checks/replies) rather than passive social media scrolling, which indicates a healthier phone usage style that has been shown to relate to better mental well-being [100]. Additionally, high probabilities of walking and cycling in this profile are strong indicators of better cognitive performance [107]. The relatively low levels of loudness and brightness indicate that the movement of persons in this profile occurs in calmer environments rather than noisy, attention-demanding spaces, an environment previously associated with better positive functioning [111]. Moreover, during the pandemic, greater physical activity and lower phone use may have reflected differential access to outdoor spaces, employment arrangements, caregiving responsibilities, or local restrictions rather than a stable post-pandemic lifestyle. Thus, the observed association with positive functioning should not be generalized beyond the study context without replication.

Person-level profile 7 was dominated by day-level profile 1, days with high social media app duration, above-average mobility, relatively loud and bright contexts, and frequent phone sessions, with a smaller share of day-level profile 3 (sedentary, social-heavy days with longer sessions, but fewer pickups). Social media duration was consistently high, while communication-app duration is moderate. On most days, screen frequency was high alongside moderate-to-high screen time. Variability arose on days from profile 3; lower frequency, but longer durations. Mobility and activity are above average on the dominant days from day-level profile 1, interleaved with lower-movement and social media–heavy sedentary days from day-level profile 3. In summary, person-level profile 7 comprised individuals with high recorded social media app duration, frequent phone sessions, moderate-to-high screen time, and above-average mobility. These measures indicate the volume and frequency of device activity, but do not establish sustained attention, exposure to particular content, passive use, or active social interaction. One plausible explanation is that these were long-distance commuting days on which people were heavily engaged with their smartphones, which can be exhausting [112]. In addition, mobility and digital engagement in July 2020 may have been shaped by pandemic-related arrangements. Profile 7 should therefore not be interpreted as a stable “always-on” lifestyle without replication using postpandemic data.

Taken together, only positive functioning reliably differed across 2 of 7 person-level profiles; positive affect and personal relationships may require longer observation windows or additional features (eg, social-interaction content/quality) to detect profile differences.

The examination of lifestyle profile×personality interactions provided no evidence that personality moderated the associations between behavioral profiles and mental well-being. A likely explanation is limited statistical power to detect moderation effects. Interaction effects are typically harder to detect than main effects, and statistical power is limited unless designs are optimized and samples are very large [113]. A priori power calculations using Cohen f²= .02 (small effect) indicated that approximately n=1261 participants would be required to achieve 80% power for the omnibus interaction test (computed with the pwr R package [114]), which exceeds the available sample size in the present study. Accordingly, the current findings should not be taken as strong evidence that moderation is absent; rather, they suggest that any profile×personality interactions—if present—are likely small and will require larger, confirmatory samples to detect reliably.

Findings from this study have important implications. First, we advanced a digital phenotyping approach on mental well-being, showing that well-being may be reflected in specific combinations of daily behaviors. This approach underscores the value of person-centered analytic approaches for capturing heterogeneity in lifestyles.

Second, should predictive validity be established, this approach may have the potential to inform mental health monitoring by offering interpretable information about lifestyle configurations that may be higher risk. Unlike many previously proposed monitoring systems that rely on black-box machine learning models [115], a profile-based monitor would surface human-readable patterns (which daily profiles are trending and how often they recur) that both clinicians and users can understand. Such transparency can support shared decision-making [116], guide clinician-patient communications [117], and help users reflect on the routines contributing to their risk flags.

Third, with future experimental validation, the monitoring platforms mentioned above could support the development of just-in-time adaptive interventions (JITAIs) that aim to deliver tailored support when individuals appear to exhibit risk-like patterns, thereby potentially improving the timing of interventions [118,119]. Building on this foundation, JITAIs might use profile-informed triggers to flag candidate “windows of opportunity” and provide context-appropriate options [118]. However, these possibilities require prospective testing for efficacy, feasibility, and safety.

Fourth, rather than targeting a single behavior, prior work has proposed synergistic approaches of intervention (ie, Multiple Health Behavior Change) [120,121]. Our findings reinforce this perspective by showing that well-being could be reflected in configurations of co-occurring routines [122,123]. In practice, profile-guided interventions could coordinate a few low-friction actions (eg, a short movement cue paired with a soft screen-time nudge) during candidate high-risk days, with the goal of shifting the overall configuration. While causal efficacy remains to be established, our findings provide granularity—which behaviors tend to co-occur—that could inform the design of such multicomponent strategies proposed previously.

Despite these contributions, some limitations warrant consideration. First, our sample was drawn exclusively from an online panel of German participants during July 2020. The context of the COVID-19 pandemic may have influenced the behavioral patterns captured in the study. Public policies altered work and commuting arrangements, and reduced access to social activities may shape the ways in which daily activities were organized [124]. Consequently, some of the identified profiles may reflect adaptations to the pandemic context rather than stable and generalizable lifestyle phenotypes. The present findings should therefore not be assumed to represent general behavioral patterns. Given that lifestyle norms and their links to well-being may differ across historical periods and cultures [125], direct replications using post-pandemic data, diverse cultural contexts, and repeated measurement across different time periods are necessary to determine whether the profiles are reproducible and temporally stable. Second, the current study recruited only Android users, as our app operates exclusively on this system. Given documented differences between Android and iOS users in education, income, and extraversion [126], future research should consider platform-related selection effects when replicating the results. Third, for interpretability, we selected only a limited set of sensing features; digital-phenotyping research suggests that more comprehensive multimodal data (eg, physiological data) can yield richer lifestyle profiles and stronger links to mental health outcomes [14]. Finally, our study design was correlational and cannot establish causality. It remains unclear whether certain routines impact well-being, whether well-being levels lead to certain routines, or whether both are explained by third variables, such as age, which also differed across profiles.

In light of these limitations, the present study should be viewed as an initial step toward interpretable, digital trace–based lifestyle profiling. Accordingly, several avenues for future research follow. First, the present findings show that positive functioning is reflected in lifestyles inferred from digital traces in generally healthy populations. An important next step is to examine clinical samples and test whether they show similar lifestyle patterns to those in the general population and whether certain lifestyles indicate risk for particular mental disorders. This would provide a more complete picture of digital phenotyping and inform actionable intervention practices. Second, the current interpretation of the profiles is post hoc and speculative. To validate these profiles, future studies should use methods such as qualitative interviews or daily diary approaches, asking individuals belonging to a specific profile about their daily activities. Combined with asking them how they feel throughout the day, further work could examine dynamic fluctuations in their daily lifestyles and mood. This would allow modeling of dynamic, within-person relationships between daily patterns and well-being. Third, potential weekday–weekend differences may emerge in both behavior and well-being. Future studies could consider this by examining longer timeframes that include sufficient weekend data, which would yield insights into work–life balance and how people spend their workdays and free days in ways that reflect their well-being.

This study demonstrates the feasibility of using multilevel LPA to identify interpretable day- and person-level configurations from mobile-sensing data. Evidence linking these profiles to mental well-being was limited; profiles did not differ in overall mental well-being, positive affect, or satisfying interpersonal relationships, and only 2 of the 7 profiles differed in positive functioning. No evidence of personality moderation was observed. These findings indicate that distinguishable behavioral profiles do not necessarily correspond to distinguishable levels of mental well-being. Future studies should replicate the positive-functioning contrast, examine within-person temporal associations, and establish predictive and incremental validity before considering profile-based monitoring or intervention applications.