Variability in self-reported depression symptomology and associated behavioral markers in digital phenotyping

Abstract

Digital phenotyping studies using smartphone-sensed data have identified several behavioral markers associated with depression. However, the generalizability of these markers is constrained by multiple factors, including variability in depressive symptoms and associated behaviors, both between and within individuals over time. This study explores variability in depression symptoms and associated behaviors using smartphone-sensed data collected from participants diagnosed with depression. It examines behavioral features derived from smartphone sensing as potential markers of depression. We analyzed smartphone-sensed behavioral data from 62 patients with major depressive episodes across three subgroups: major depressive disorder (MDD, n = 41), borderline personality disorder (BPD, n = 12), and bipolar disorder (BD, n = 9). Depression symptoms were assessed with the 9-item Patient Health Questionnaire (PHQ-9). Symptoms differed between subgroups and across severity levels. Association analysis suggested variability in correlations between depression severity and behavioral features, both between participants and over time. Multilevel modeling identified two demographic predictors, employment status (β = − 4.79, 95% CI [− 7.65, − 1.80], p = 0.004) and age (β = − 0.12, 95% CI [− 0.25, 0.00], p = 0.050), along with one behavioral predictor, lower nighttime movement (β = − 0.79, 95% CI [− 1.29, − 0.29], p = 0.024), that were associated with depression severity.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-48625-w.

Introduction

Depressive disorders rank among the most prevalent mental disorders and are among the most burdensome health conditions worldwide¹. In 2021, an estimated 330 million people worldwide were living with depressive disorders, affecting approximately 4.4% of the global population². Clinical assessment of depression relies on structured interviews, observer scales, and questionnaires³, which are prone to biases such as recall bias⁴. Additionally, the heterogeneous nature of depressive symptoms⁵ adds further complexity to diagnosis.

The ubiquity of smartphones and other personal smart devices has inspired research to use data from these devices to detect and predict depression symptoms objectively. Digital phenotyping is a field of research that utilizes personal digital devices to measure human behavior and physiology unobtrusively and continuously in real-life settings⁶. Collecting and analyzing data on individuals’ behavior using these devices can provide new information about mental health disorders, including depression and behavioral patterns associated with them⁷. Ideally, the knowledge produced by digital phenotyping research could be applied in practice as new methods and tools for symptom monitoring and prediction, supplementing current means of mental healthcare. Furthermore, these methods could help design early warning systems, which can be integrated with interventions for symptom worsening⁸.

Research on digital phenotyping of depression has mainly focused on identifying associations between smartphone-sensed behavior and depression symptoms, predicting depression symptom severity and diagnostic classification, and forecasting future mood⁹. The studies typically rely on behavioral data representing participants’ physical activity, mobility, sociability, sleep, and smartphone usage patterns¹⁰. Although these approaches have achieved promising results, the reported associations between behavior and depression vary across studies in both statistical significance and direction (positive or negative) (see, e.g.,^10,11, implying that the results may not generalize well across different samples and contexts. The differences may arise partly from heterogeneity in study objectives, samples, used methodologies, and reporting practices¹⁰, device usage differences between the study participants¹², but also from the differences in depressive symptoms between participants¹³ and temporal variability of symptoms within the participants^14–18. The heterogeneity in depression symptoms and related observable behavior characterizes the depressive disorder; for example, both insomnia and hypersomnia, a decrease or increase in appetite, are regarded as depression symptoms¹⁹. Further, research has identified heterogeneous temporal characteristics of individual depression symptoms²⁰ and variability in long-term trajectories of depression severity and stability²¹.

This study aims to explore variability in depression symptoms and related behavioral features as observed through digital phenotyping. Moreover, it assesses the underexplored contribution of between- and within-participant effects in behavior-depression relationships, aiming to identify features associated with depression. The distinction between these effects is important for evaluating how both long-term, between-participant differences in behavior and short-term fluctuations within participants’ behavior reflect depression severity. We investigated variability from two angles. The first angle looked at differences in self-reported depression symptoms through answers to questionnaires. The second angle focused on the smartphone-sensed behavioral features and their associations with depression severity. The first angle was represented by our first study objective, in which the questionnaire answers were compared at the symptom level, across diagnostic groups, and by severity level. The second angle led to our second and third study objectives. The second objective was to explore correlations between behavioral features and depression across the study population, at the participant level, and over time. The third objective was to assess both within- and between-participant variability in behavioral features using multilevel modeling and to identify potential behavioral features that could be indicative of depression while controlling for participants’ background and contextual factors.

The study analyzed smartphone-collected behavioral data from outpatients diagnosed with depression. Self-reported scores from the 9-item Patient Health Questionnaire (PHQ-9)²² were used to assess the severity of depression. We extracted features from smartphone data that reflect daily behavioral aspects: physical activity, mobility, sociability, phone usage, and sleep. These features served as proxies for behavioral patterns, possibly associated with depression.

Exploratory analysis indicated heterogeneity in depressive symptom representations between the patient groups and in behavior-depression associations across the study population. The associations showed differences in strength and direction (positive and negative) over time within participants. This observed variability aligns with the prior literature describing the heterogeneous nature of depressive symptoms and suggests that such heterogeneity may also be reflected in behavioral data collected through smartphones. Furthermore, multilevel modeling suggested that variability in behavior over time within individual participants was more predictive of depression severity than the behavioral differences between participants. These findings suggest that smartphone-sensed behavioral features may have limited predictive utility in population-level analyses, particularly when simple, unimodal linear models are applied and when homogeneous associations between behavioral features and depression are assumed across participants.

Instead, the results indicate that modeling may benefit from focusing on behavioral changes over time within participants and on individual symptoms, which could improve the accuracy of depression monitoring and prediction.

Results

PHQ-9 questionnaire analysis

Our first objective was to assess the variability in depression symptoms through PHQ-9 answers at the item (symptom), diagnostic group, and severity levels. The summary statistics showed that the MDD patients had, on average, the lowest average depression level (mean = 11.6, SD = 6.0), while BPD had the highest average (mean = 14.9, SD = 4.8). At the study population level, moderate depression severity (PHQ-9 scores from 10 to 14) was the most frequent (27.5%). For a group-level summary of depression severity, refer to Table S1 in Multimedia Appendix 1. Considering PHQ-9 score variability, the BD group showed a wider distribution than others, reflecting higher within-group differences in depression variability. Nonetheless, the Kruskal-Wallis test²³ indicated that these differences in mean score or standard deviation were not statistically significant. Figure 1 shows the distribution of PHQ-9 mean scores and standard deviations for each group.

Fig. 1.

PHQ-9 score group-level summary statistics. The scores were first averaged within participants and then visualized. (A) Boxplots of group-level mean scores: MDD (µ = 11.65, σ = 5.96), BPD (µ = 14.89, σ = 5.87), and BD (µ = 12.12, σ = 4.83), (B) Boxplots of standard deviations of scores: MDD (µ = 2.65, σ = 1.28), BPD (µ = 3.44, σ = 2.49), and BD (µ = 3.78, σ = 2.61). A wider interquartile range for BD suggests greater fluctuations in depression severity among BD patients compared to others.

An exploratory comparison of the PHQ-9 item distributions at different depression severity levels exhibited differences across the patient groups. However, conducting the Kruskal-Wallis test indicated that these differences were not statistically significant. At the lowest depression level, items 3 (trouble with sleep) and 4 (tiredness or low energy) are more pronounced in the BPD group, while item 7 (trouble concentrating) has higher scores for the BD group than others. At the mild and moderate depression levels, the items are more evenly distributed across all groups. At the moderately severe level, item 9 (suicidal ideation) scores higher in the BD group, while item 1 (anhedonia) is pronounced in the MDD group, and the BPD group also exhibits higher scores for item 6 (low self-worth). Finally, at the highest severity level, the BD group differs from the BPD group, particularly in items 8 (psychomotor retardation) and 9 (suicidal ideation). Figure 2 illustrates these differences in distributions across groups and depression severity levels. For detailed results, refer to Appendix 1 Table S2 and Table S3 for a summary of PHQ-9 questions.

Fig. 2.

Average PHQ-9 item scores by diagnosis group and depression severity. For visualization purposes, we treated PHQ-9 items (ordinal) as continuous variables. The radar charts show the average of PHQ-9 item scores across different depression severity levels for BD, BPD, and MDD groups. Letters from I₁ to I₉ denote the items, and different colors represent severity levels. The scale, ranging from 0 (center) to 3 (outer circle), denotes item values.

Further exploration of the item score distributions across depression severity levels revealed that, on average, PHQ-9 items do not increase at the same rate or linearly with the total score. Items 3 (trouble with sleep) and 4 (tiredness or low energy) constantly contributed to the total score across all the severity levels. Also, item 6 (low self-worth) had high scores across the severity levels. On the other hand, item 8 (psychomotor retardation) stayed relatively low across all depression severity levels. Items 3, 4, 5 (poor appetite or overeating), and 6 showed the greatest relative changes, transiting between mild depression and moderate depression. These changes suggest that these items (and corresponding symptoms) might be particularly informative for detecting depression severity above the clinical threshold (a PHQ-9 score of 10). Further, item 9 (suicidal ideation) only emerged at higher depression levels. Figure 3 illustrates the PHQ-9 item distributions across the severity levels.

Fig. 3.

PHQ-9 item average score distributions across depression severity levels, combining the data from all patient groups. For exploratory analysis purposes, we treated the (ordinal) items as continuous variables and visualized the distribution shapes. The comparison shows that the items neither increase at the same rate nor in a linear manner with total depression severity. Notably, items 8 and 9 have higher values only at higher severity levels. Additionally, comparing the mild (scores from 5 to 9) and moderate (scores from 10 to 14) levels of depression severity, items 3 to 6 exhibit the highest differences.

Association analysis

To address the second objective, we examined the associations between behavioral features and depression severity. We conducted an exploratory correlation analysis, calculating pairwise Kendall rank correlation coefficients between the behavioral features and PHQ-9 scores by pooling data from participants, and at the individual participant level, thus comparing general, study population-level associations and individual participant-level variability. The results revealed that pooling the participants, all behavioral data features had weak to moderate correlations (Kendall rank correlation coefficient between − 0.3 and 0.3) with PHQ-9 depression scores. Nonetheless, these correlations were not significant after controlling for false positives with false discovery rate (FDR) correction at a significance level of = 0.05. On the other hand, inspecting the correlations at the individual participant level resulted in a broader range of coefficients, summarized in Table 1.

Table 1.

Association analysis summary: the proportion of correlation types across the behavioral features.

Correlation type	Range	Pooled participants (%)	Individual participants (%)
Very strong negative	0.5 to 1.0	0.0	5.7
Strong negative	0.3 to 0.5	0.0	10.8
Moderate negative	0.1 to 0.3	29.7	17.5
No/Weak correlation	− 0.1 to 0.1	53.6	23.0
Moderate positive	− 0.3 to − 0.1	16.7	20.6
Strong positive	− 0.5 to − 0.3	0.0	11.8
Very strong positive	− 1 to − 0.5	0.0	10.6

The correlation was measured by Kendall rank correlation coefficients, calculated by pooling the participants and at the participant level. In general, participant-level coefficients exhibited broader distributions than pooled-level coefficients.

Further, we examined how the associations varied over time within the participants with exploratory rolling-window correlation analysis. We counted the proportion of users and features exhibiting both positive and negative correlations during the data collection period across different window sizes. The results indicated that correlation directions (positive or negative) might change over time. In general, the proportions of mixed correlation, reflecting that participants had both positive and negative correlations, are higher with shorter windows. These findings suggest that the associations are not static and should be considered dynamic in modeling tasks. Appendix 1 Figure S1 summarizes the average proportion of features exhibiting both positive and negative correlations across participants.

Multilevel modeling

To address the third objective, we used multilevel modeling to examine both within- and between-participant variability in behavior and to identify features indicative of depression. We employed a stepwise modeling strategy, starting with a baseline model that included only the random intercepts for each participant and subsequently adding background information, between- and within-effects as predictors. The baseline model, including only random intercepts, explained 72.0% of the total variance. The first model, which included demographic and contextual information, increased the conditional R²-score (accounting for both fixed and random effects) to 0.737 and the marginal R²-score (reflecting only fixed effects) to 0.162. The second model, adding time-varying within-participant features, further increased the conditional R² to 0.743 and the marginal R²-score to 0.172. Finally, the complete model combines background, contextual, and within- and between-participant behavioral features, resulting in a conditional R²-score of 0.797 and a marginal R²-score of 0.207.

The results from multilevel models showed that the intraclass correlation coefficient (ICC) was relatively high across models (ranging from 0.69 to 0.73), indicating that a substantial proportion of the variability in self-reported depression severity was attributed to the between-participant differences. Model comparisons with ANOVA revealed that adding within-participant behavioral features (Model 2) improved model fit over the baseline model (Model 1) (χ² = 26.22, df = 16, P = 0.05; AIC = 3613.76, BIC = 3721.02). However, incorporating between-participant features (Model 3) led to a significant chi-square change over Model 2 (χ² = 54.78, df = 16, P < 0.001; AIC = 3590.9, BIC = 3769.74). Therefore, the increase in BIC suggests that Model 2 represents the most parsimonious model.

In Model 2, employment status (β = − 4.79, 95% CI − 7.65 to − 1.80, P = 0.004) and age (β = − 0.12, 95% CI − 0.25 to − 0.00, P = 0.050) emerged as significant demographic predictors of depression severity, with unemployment and younger age linked to higher symptom levels. Among behavioral features, higher nighttime movement (β = − 0.35, 95% CI − 0.63 to − 0.06, P = 0.024) was linked to reduced symptom severity. Notably, when taking into account the between-participant effect, outgoing SMS count emerges as a significant predictor. The effect is positive at the between-participant level (β = 2.20, 95% CI 1.12 to 3.32, P = < 0.001) but negative at the within-participant level (β = − 0.56, 95% CI − 1.00 to − 0.17, P = 0.004). Table 2 shows the condensed model summaries, displaying only the features with P-values below 0.05. For the detailed results, refer to Table S4 in Multimedia Appendix 1.

Table 2.

Multilevel model summaries.

	Model 1			Model 2			Model 3
	Estimate	CI	p	Estimate	CI	p	Estimate	CI	p
Age	− 0.1	− 0.22–0.03	0.142	− 0.12	− 0.25 – − 0.00	0.05	− 0.09	− 0.24–0.05	0.218
Gender (Male)	1.05	− 2.28–4.34	0.534	1.3	− 2.08–4.48	0.458	1.53	− 2.15–5.22	0.404
Children (Yes)	− 1.9	− 5.05–1.09	0.23	− 2.39	− 5.69–0.95	0.162	− 2.60	− 5.58–1.047	0.166
Pet (Yes)	− 0.67	− 3.53–2.29	0.656	− 1.13	− 4.14–1.83	0.442	− 1.48	− 4.99–1.87	0.400
Employed (Yes)	− 4.18	− 7.01 – − 1.38	0.004	− 4.79	− 7.65 – − 1.80	0.004	− 5.40	− 8.75 – − 1.93	0.002
SMS count out W				0.01	− 0.38–0.38	0.954	− 0.56	− 1.00 – − 0.17	0.004
Acc STD night W				− 0.38	− 0.68 – − 0.07	0.012	− 0.79	− 1.29 – − 0.29	0.002
SMS count out B							2.20	1.12–3.32	< 0.001
Random effects
σ2	11.85			11.15			10.89
τ00	25.95 user			27.36 user			29.49 user
ICC	0.69			0.71			0.73
N	57 user			57 user			57 user
Observations	645			645			645
Marginal R2/conditional R2	0.162/0.737			0.172/0.743			0.201/0.797

For brevity, the table shows only the statistically significant behavioral predictors (P-values ≤ 0.05 are shown in bold). Model 2 is the best-fitting and most parsimonious, showing within-participant behavioral features: nighttime accelerometer variability with lower severity. Additionally, higher age and employed status are associated with lower depression severity.

W denotes the within-participant effect, σ2 residual variance, τ00 intercept variance, and ICC intra-class correlation coefficient.

Model 1 includes background demographics and contextual predictors, with random intercepts for each participant.

Model 2 adds within-participant effects (features) into Model 1.

Model 3 adds between-participant effects (features) into Model 2.

Discussion

This study investigated the variability in self-reported symptoms of depression and associated smartphone-sensed behaviors. We observed variability in PHQ-9 items between the patient groups and across different levels of depression severity. Associations between behavioral features and depression severity varied considerably between the study participants and over time within participants. The multilevel model results indicate that approximately 70% of the variability in self-assessed depression is attributable to between-participant differences in depression severity. Adding demographic background, context, and time-varying within-participant behavioral features improves the model’s explanatory power by up to 76.3%. Multilevel modeling identified that demographic factors, age, and employment status, and within-participant behavioral features of daily outgoing SMS count, as well as physical nighttime activity, are associated with depression severity. Notably, while these findings exemplify variability in this study population, the limited sample size warrants interpreting them as exploratory observations.

Inspection of PHQ-9 item distributions suggested differences in the representation of depression across the diagnostic subgroups. The differences suggest that accommodating individual differences in depression manifestations at the symptom level and modeling each participant individually or pooling together participants with similar symptoms might benefit the modeling accuracy. Additionally, these differences in symptom profiles may hint at possible subgroup distinctions within depressive disorder; a unique symptom profile may characterize each subgroup. However, as the subgroups in this study were small, caution should be exercised when generalizing such findings.

Further, we observed the highest difference in PHQ-9 score distributions between mild and moderate depression on items 3 (trouble with sleep), 4 (tiredness or low energy), 5 (poor appetite or overeating), and 6 (low self-worth). These differences suggest a possible pattern worth further investigation, specifically exploring whether the transition from lower depression into moderate level may be identifiable through monitoring behavior related to these items, such as sleep disturbances, daily activity and mobility, and sociability. Additionally, we noticed that item 8 (psychomotor retardation) exhibits higher scores only at higher depression severity levels, aligning with clinical knowledge that psychomotor retardation typically occurs only in severe depression and is an important characteristic of the DSM-5 melancholic specifier or subtype of major depressive episodes^19,24. Similarly, item 9 (suicidal ideation) gets higher scores only at higher depression levels. This finding aligns with other studies and is a clinically important point—the alleviation of depression results in the alleviation of suicidal thoughts²⁵.

Correlation analysis results indicated that most behavioral features had weak correlations with PHQ-9 depression scores at the population level. In contrast, at the individual participant level, the correlations exhibited a wider range of both positive and negative values, which may cancel each other out when aggregated at the population level. Notably, the number of data points provided by each participant varies. Thus, some of the higher correlations might stem from the relatively low observation count by change. On the other hand, the low variability in PHQ-9 scores could explain the low correlations for some participants. Likewise, considering application usage-related features, low correlations may occur because some participants use smartphone applications infrequently. Finally, the complementary rolling-window correlation analysis explored the within-participant variability in time, revealing periods with varying correlation strengths and directions. This observation suggests that the relationship between behavioral features and depression severity should be considered time-dependent rather than static, aligning with existing literature that emphasizes the need for models that can accommodate the time-varying dynamics of the depression-behavior relationship^14,16.

The multilevel modeling results regarding the explanatory power of random and fixed effects align with other studies using hierarchical linear modeling (e.g.,^26–28. Adding behavioral features increased the marginal R² -score by 2.1%, indicating modest explanatory power of passively sensed, time-varying participant-level behavioral features. Further, a significant portion of PHQ-9 variance is explained by between-participant variability in depression severity. Together, these findings suggest that for the purpose of depression monitoring, the clinical focus should be on detecting deviations from a patient’s own baseline (within-person) rather than comparing them to others (between-person), while accounting for participants’ depression severity history. Notably, outgoing SMS count showed opposite associations at the between- and within-participant levels. At the between-participant level, higher average outgoing SMS count correlated with higher depression severity, possibly reflecting a tendency for more distressed individuals to rely on text-based communication as a means of seeking support. At the within-participant level, however, the association reverses: temporary increases in outgoing SMS relative to one’s own baseline may reflect periods of better functioning or greater social engagement.

The validity and generalizability of this study’s findings are primarily limited by the amount of available data; the sample size, participants dropping out of the study, and missing intermediate observations (the dropout rate and data missingness are discussed more thoroughly in our previous work^29,30. These factors may reduce the statistical power and contribute to greater uncertainty in estimates for subgroup comparison and heterogeneity analyses. Limited yet high-dimensional data is a common issue across studies involving personal digital devices for data collection^31,32. While the quantity of passively collected raw data is relatively high in this study, more bi-weekly sampled active data for each participant would have benefited the analysis, leading to higher statistical power and making modeling less prone to overfitting. Also, combining data sets across different studies may provide a feasible approach to obtaining larger and more heterogeneous samples.

Additionally, the temporal nature of the data might bias the analysis results. The feature standardization and imputation techniques used in this study assume stationarity of variance over time, which might not hold with behavior in participants with mood disorders. These issues could be mitigated by considering time-aware standardization and thoroughly analyzing the nature of missing observations (whether they are missing completely at random, at random, or not at random) and thoroughly comparing between different imputation techniques, such as multiple imputation using chained equations (MICE)³³.

It is also worth exploring whether the changes in behavior are more common in moderate to severe depression. If sensitive digital phenotyping measures for the milder range of depression are to be explored, they probably must tackle more subjective and internal aspects of depression. Achieving this might require measures that are sensitive to the negative emotional and self-referential cognitive biases in depression, such as changes in speech prosody or the linguistic content of written texts (for example, in emails and social media platforms), consumed media content, or facial expressions.

Due to the limited explanatory power of linear modeling in this dataset, we suggest that future work explore multiple features in modeling, including feature interactions and non-linear relationships in the data with modeling techniques, such as gradient-boosting trees, and personalizing the models. Previous research has addressed the problem of between-participant variability through personalized modeling that adapts to the differences between participants. The models can be personalized, for example, by using a participant-specific (idiographic) approach (e.g.,²⁸), personalizing the depression symptom profiles (e.g.,³⁴), hierarchical modeling (e.g.,^35,36), symptom similarity-based grouping (e.g.,³⁷), combining the training data from participants with similar behavioral feature distributions (e.g.,³⁸), or collaborative filtering modeling (e.g.,³⁹).

This study illustrates how self-reported depressive symptoms and associated mobile-sensed behaviors vary between and within patients clinically diagnosed as experiencing major depressive episodes. Our multilevel modeling indicated that a substantial portion of depression variability is due to individual differences. In summary, these findings suggest that smartphone-sensed behavioral features may have limited clinical utility in monitoring depression using digital phenotyping within this study population. Hence, we propose that future research could focus on improving depression prediction and forecasting accuracy by utilizing data from multiple sources using advanced techniques capable of modeling non-linear and time-varying relationships, feature interactions, and differences between participants. Through these results, we bring into focus the importance of addressing and reporting heterogeneity within the study population, as it may provide useful information for future research design, identifying useful behavioral markers, and optimizing sample size through power analysis.

Methods

Setting

This study analyzed behavioral data from the Mobile Monitoring of Mood (MoMo-Mood) study²⁹, which includes continuously sensed smartphone behavioral data (passive data), self-reported responses to biweekly depression questionnaires (active data), and background demographic and contextual information. The study involved 164 participants: healthy controls (n = 31) and patients (n = 133). All patients were experiencing an ongoing major depressive episode (MDE) at the time of recruitment and were diagnosed with either major depressive disorder (MDD, n = 85), major depressive episodes with bipolar disorder (MDE|BD, n = 21), or major depressive disorder with comorbid borderline personality disorder (MDD|BPD, n = 27). The patients were voluntarily recruited from primary care and psychiatric outpatient treatment clinics in Finland, while healthy controls were recruited from university students and healthcare personnel. The patients were under treatment and were diagnosed as having ongoing major depressive episodes with structured interviews, the Mini-International Neuropsychiatric Interview⁴⁰, and the Structured Clinical Interview for DSM-IV Axis II Personality Disorders⁴¹.

The study used continuous enrollment, allowing the participants to enter or withdraw from the study at different times, with a recommended participation period of one year. All participants were informed about the study prior to enrolment, including the data collection and their option to leave the study at any time. Informed written consent was obtained from all the study participants. The participants were compensated with four movie tickets. The Helsinki and Uusimaa Hospital District’s Ethics Committee approved the study research protocols (including the data stream and collection platform security), and Helsinki and Uusimaa Hospital District Psychiatry granted a research permit (approval number § 125/2018). A written data security statement was authorized by the local research ethics committee and IT support. All research procedures followed the ethical standards of the Declaration of Helsinki. For more details, refer to earlier work^29,30,42.

Dataset description

The passive data used in our analysis originated from various smartphone sensors and logs, collected using the AWARE app⁴³ and curated using the Niima platform⁴⁴. The data encompasses five aspects of daily behavior: physical activity and mobility, sociability, phone interactions, and sleep. Behavioral features related to these aspects have been identified as being associated with depression severity in previous research ( e.g., for physical activity^45–47, mobility^48,49, sociability^50,51, phone interactions^52,53, and sleep^39,54). We operationalized these behavioral aspects through specific features extracted from raw smartphone data, serving as proxies for real-world behaviors. The active data has participants’ self-reported PHQ-9 questionnaires, which assess the depression severity. The PHQ-9 is a tool used for screening for diagnosis and monitoring of depressive symptoms⁵⁵. The PHQ-9 survey comprises nine items, each addressing the main symptoms of depression, such as having little interest or pleasure in doing things. According to the survey, patients were asked to respond based on their experience over the last two weeks, with each item rated from 0 (not at all) to 3 (nearly every day). The Background demographics include categorical variables, including participants’ age and sex, used as control variables. Contextual information includes details about full-time work, shift work, cohabitation status, the number of children, and whether the participants have pets. We selected these factors since they may affect participants’ daily behaviors. Participation adherence, assessed via PHQ-9 responses, declined over the study period, with 65.7% of participants still active by week 8. Adherence varied across subgroups, ranging from 50% in BPD to 83% in BD. Passive data missingness, proxied by battery data availability, averaged 10.5% across all participants, with notable variation by group (BD: 4.5%, BPD: 20.4%, control: 1.2%, MDD: 14.1%). A survival analysis of dropout and missingness is provided in our previous work²⁹.

Data preparation and preprocessing

Of the 164 participants, we excluded the control participants from the analysis since they exhibited low levels and variability (mean = 1.2, std = 1.8) of self-measured depression³⁰. Further, we excluded the participants who provided less than one month of passive data (equaling at least two biweekly PHQ-9 responses), resulting in a total of 62 (39%) participants from three different groups: (1) patients with BD (n = 9), (2) patients with BPD (n = 12), and (3) patients with MDD (n = 41). Descriptive statistics and demographic information of the participants are presented in Table 3.

Table 3.

Demographic and clinical characteristics.

Observation counts	671
Age	35.1 (sd 12.2)
Gender
Male	16
Female	46
Employment
Fulltime	20
Other	42
PHQ9 scores	11.3 (sd 5.8)

The raw passive data comprised accelerometer readings, application usage logs, battery level, GPS locations, phone call logs, SMS logs, and screen usage. We used the Niimpy behavioral analysis toolbox⁵⁶ to preprocess the raw sensor data and extract behavioral features. We extracted three types of features: volume-based, temporal, and contextual features. The volume-based features represent counts and durations of specific activities, such as smartphone screen usage time. Temporal features were derived by segmenting the volume-based features into 6-hour bins based on the time of day (12:00 a.m.–6:00 a.m., 6:00 a.m.–12:00 p.m., 12:00 p.m.–6:00 p.m., and 6:00 p.m.–12:00 a.m.) for more precise information in daily rhythms, similar to our previous work³⁰. Considering the behavioral aspects, physical activity was captured through phone accelerometer data, providing temporal information about movement intensity and variability. Mobility was assessed using GPS location data, which provided spatial (e.g., total distance traveled) and contextual information (e.g., daily time spent at home) on behavior. Sociability was measured by phone call counts, durations, and the number of messages sent and received. Phone interactions were inferred from application usage logs (e.g., application category, usage count, and duration), screen usage counts (on and off), usage duration, and battery charge level. Sleep duration and timing were approximated using the longest period of inactivity measured through screen activity data, following a process similar to that demonstrated in²⁹. This method gives comparable results with other methodologies for measuring sleep in this population⁵⁷. For further details, refer to Table S1 in Multimedia Appendix 1. The average number of completed PHQ-9 assessments and aligned passive data observations per participant, reported overall and by diagnostic group, are provided in Table S4 in Multimedia Appendix S4.

For active data, we used scores from individual items and assessed the depression severity with the total sum of item scores, with the result ranging from 0 (no depression) to 27 (severe depression). Further, categorical background demographics and contextual variables were used as control variables. We used demographics, including age and gender, as well as contextual variables that might influence daily behavior, such as having children, owning a pet, and being employed. The variables were used in multilevel modeling, encoded using one-hot encoding.

Extracted temporal and volume-based behavioral features were aligned with biweekly depression questionnaires and aggregated by calculating the arithmetic mean over the 14 days preceding the days the PHQ-9 measurement took place. Location data features (including distance-based and significant places-based features) were calculated directly over 14-day windows, thus requiring no averaging. Further, we pre-filtered redundant features based on prior knowledge (for example, screen off count, since it reflects the same information as screen on count). The resulting data comprised 671 biweekly observations covering 224 behavioral features, two background demographics, five contextual variables, and one external variable. Figure 4 shows the details of data preprocessing and aggregation schematics.

Fig. 4.

Participant selection, data collection, and processing schema. For this study, we selected participants from three patient groups (BD, BPD, and MDD) who answered at least two bi-weekly (14-day) PHQ-9 questionnaires and provided passive data for 14 days prior to each PHQ-9 answer. Smartphone-collected data encompassed: (1) passive data, consisting of raw smartphone-sensed data, including accelerometer readings, phone call and SMS logs, app and screen usage logs, battery level logs, and GPS location data; (2) active data, consisting of PHQ-9 questionnaires; and (3) demographic information, such as participants’ age and work status. The merged datapoints are denoted by t₁ to t₂₆, and day d₀ represents the day the PHQ-9 was reported, and d_− 14 to d_− 1 the preceding 14-day period, during which passive data was collected. Data collection lasted up to one year.

For the first study objective, to assess how depressive symptoms differ between patient groups, we used the raw values from the PHQ-9 questionnaire items, removing missing or incomplete records. For the second objective, behavior–depression association analysis, we used aligned passive (behavioral features) and active (self-assessed PHQ-9 scores) data, removing observations where either the feature or questionnaire was missing. For the third objective, we imputed the missing intermittent passive data observations. For sociability features (e.g., call and SMS counts), we used zero imputation to distinguish true missingness from observed zeros using battery information: we treat a person-day as missing only if there is no battery record (used as a proxy for the phone not being active or observed). When battery data indicate the phone was active on a given day but no call or text events are detected, we interpret this as a valid zero (that is, no communication recorded that day), and coding it as 0 is consistent with the measurement process rather than an imputation of unknown behavior. For the remaining variables, participant-median imputation was used only for intermittent missingness where battery data indicate the device was active, but a specific feature value is unavailable. Further, to examine within- and between-participant effects of behavior using multilevel modeling, we disaggregated passive data features using participant-mean and grand-mean centering⁵⁸. The within-participant components reflect fluctuations relative to an individual’s average and their relationship to the severity of depression. Contrarily, between-participant effects reflect how differences in average behavior relate to depression severity (e.g., whether participants sleeping longer on average have higher or lower depression levels). For within-participant centering, we centered each variable around the participant’s mean to account for variability over time. Second, for between-participant centering, the variables were averaged within participants and then centered across the study population’s mean. This centering approach enables the detection of both individual-level behavioral fluctuations and broader population-level patterns.

PHQ-9 questionnaire analysis

To assess the heterogeneity in self-reported depression, we conducted a descriptive analysis of PHQ-9 scores, applied the Kruskal–Wallis test²³ to evaluate group-wise differences in score means and standard deviations, and performed an exploratory analysis for each item across the groups and different depression severity levels. Additionally, we utilized exploratory data analysis to examine the scores across the groups and severity levels.

Association analysis

To investigate the associations between behavioral features and self-reported depression among participants, we conducted exploratory correlation analysis, calculating pair-wise Kendall’s rank correlation coefficients between the behavioral features and PHQ-9 scores. We selected Kendall rank correlation since it does not require the compared variables to be normally distributed or linearly related, and it can be used with ordinal data. For analysis, we used a function implemented in the SciPy library⁵⁹. We employed two approaches: pooling data from all participants and calculating the coefficients for each participant individually. Due to the multiple comparisons, we applied the FDR correction with the Benjamini-Hochberg procedure⁶⁰ to control Type I errors. Additionally, we examined whether the associations between behavioral features and depression symptoms varied within participants over time by conducting an exploratory rolling-window correlation analysis. We tested various window sizes, ranging from 28 days to one year. Within each window, we computed the Kendall rank correlation coefficient, shifting the window forward by 14 days to calculate successive correlations. We excluded participants who did not have enough data to fill at least one complete window for a given window size (and subsequent longer window sizes) from the analysis. For each window size, we used 1000 bootstrapping samples to generate more robust distributions of correlation coefficients.

Multilevel modeling

To assess the within- and between-participant effects of behavioral features on PHQ-9 scores, we used multilevel modeling with random intercepts to account for the data’s hierarchical structure. Independent variables consisted of features representing behavioral aspects, including sleep, activity, mobility, sociability, and device usage. PHQ-9 scores were used as the dependent variable. Participants were used as grouping variables. Generalized variance inflation factor (GVIF) was used to assess multicollinearity, keeping only variables with GVIF < 10. We employed a stepwise modeling strategy. First, we started with a baseline model that included only random intercepts for each participant. Second, we included demographic and contextual variables. Next, within-participant variables were added to the model. Finally, between-participant variables were included. The lme4 package⁶¹ in R was used for fitting the models, and 95% confidence intervals were estimated using bootstrapping (1000 iterations) with the sjPlot package⁶². For model comparison, we used the likelihood ratio tests to assess explanatory power and ensure model parsimony. The model formulas and the final list of predictors are shown in Appendix 3.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AIC

Akaike information criterion

ANOVA

Analysis of variance

BD

Bipolar disorder

BH

Benjamini–Hochberg

BIC

Bayesian information criterion

BPD

Borderline personality disorder

CI

Confidence interval

FDR

False discovery rate

GPS

Global positioning system

GVIF

Generalized variance inflation factor

HUS

Helsinki and Uusimaa Hospital District

ICC

Intraclass correlation coefficient

MDD

Major depressive disorder

MICE

Multivariate Imputation by Chained Equations

PHQ-9

9-Item Patient Health Questionnaire

SMS

Short message service

Author contributions

Arsi Ikäheimonen: conceptualization, data analysis, data preprocessing, methodology, study design, visualization, writing—original draft, writing—review & editing, Nguyen Luong: data preprocessing, methodology, study design, multilevel-modeling, writing—review & editing, Ilya Baryshnikov: methodology, data collection, Ti John: methodology, study design, writing—review & editing, Annasofia Martikkala: data collection, writing—review & editing, Erkki Isometsä: conceptualization, data collection, project administration, supervision. Talayeh Aledavood: conceptualization, data collection, methodology, project administration, supervision, writing—review & editing. All authors reviewed the manuscript and approved the final version.

Funding

Wihuri Foundation personal one-year fulltime working grant (00230125).

Data availability

The dataset analyzed in the current study is not publicly available, as the conditions of the granted research permit restrict its use in order to protect the participants’ privacy. However, the data is available from the corresponding author on reasonable request.

Code availability

The code used for preprocessing the data is available on GitHub⁶³.

Declarations

Competing interests

The authors declare no competing interests.