Throughout the follow-up period, every 3 months, participants were asked to complete a battery of questionnaires; these included, among others, the Inventory of Depressive Symptomatology–Self Report (IDS-SR; [28]), the questionnaire assessing the 7-item Generalized Anxiety Disorder (GAD-7; [29]), and the Work and Social Adjustment Scale (WSAS; [30]).

Digital sleep and activity data were collected through Fitbit Charge 2 and later the Charge 3 wearable devices, which recorded information on sleep and physical activity throughout the follow-up period. Table 1 summarizes the sleep and activity features included in the analysis. Of the features available for extraction, these were deemed to provide the most comprehensive and clinically relevant picture of aggregated sleep and activity. Previous literature reviews have identified them as some of the most frequently included digital features in studies linking depression and RMTs [11,13,14,31,32].

Assessed at baseline, sociodemographic and clinical characteristics were used to describe differences in the resulting solutions (see Table 2 for the variables included). These variables were chosen based on data availability and clinical relevance.

We decided to analyze sleep and activity data separately because (1) despite their inherent connection, they represent distinct behavioral and physiological constructs, (2) the exploratory nature of this analysis benefits from focusing on each domain independently, avoiding interference or dilution of signals, and (3) it simplifies model complexity and enhances interpretability.

Following best practice guidelines [24], sleep and activity data were explored using 2 probabilistic unsupervised machine learning–based clustering approaches: Gaussian mixture models (GMMs; [37]), a static model, and hidden Markov models (HMMs) with Gaussian emissions [38], which consider repeated measures. Using 2 complementary approaches, we could check whether the solution was highly dependent on the algorithm used or method-agnostic, suggesting better generalizability.

Clustering was conducted at the observational level, with baseline and follow-up timepoints included simultaneously in each model. Thus, participants contributing multiple timepoints could be assigned to different clusters across assessments. In the GMM approaches, observations were treated as independent, consistent with identifying cross-sectional latent structure across all observations. In contrast, the HMM explicitly incorporated the temporal ordering of repeated measures, enabling the estimation of transition probabilities that describe movement between states over time.

We used a diagonal covariance matrix for the GMM and the HMM, as it requires less computational power than a full matrix. Additionally, it reduces the risk of overfitting in smaller datasets (compared to a full matrix) [39]. For the GMM, we used “k-means++” as the setting for the initialization parameter to ensure that the initial cluster centers are spread out in the data space, thereby avoiding convergence on a suboptimal solution. This also reduces computational cost compared to running a full k-means [39]. All other model specifications were set to default according to the Gaussian HMM model in the hmmlearn package [40] and the GaussianMixture in the Scikit-learn toolkit [39,41]. All analyses were conducted using Python (version 3.11.0) [42]. All figures were created in R (version 4.3.1) [43] using the ggplot2 package [44].

We undertook several steps to determine the optimal number of clusters or states for the GMM and the HMM separately. Our approach aimed to ensure that the chosen model complexity was based on robust, generalizable performance rather than overfitting or sensitivity to random initializations. For each algorithm, we tested 2‐10 cluster or state solutions, with 5 folds stratified by participant identifier, to ensure that all observations for a particular participant were always in the same set. The process is outlined in Figure 1.

Once the optimal number of clusters was identified using grouped 5-fold cross-validation, the model was refitted on the full dataset (ie, all 5 folds combined) using the selected number of clusters and the corresponding “median-fit” seed. No separate test set was retained at this stage because generalization performance had already been estimated via cross-validation. This model was then refitted, and cluster assignments (state labels) were predicted for each observation (ie, each observation was assigned to its most likely cluster). Once each observation was assigned to a given cluster, descriptive statistics for each profile were calculated. For each cluster, the mean values of the input RMT metrics were computed and interpreted to ascertain digital feature differences between the identified clusters. To facilitate comparisons across different features, we used the standardized and centered feature values, representing standardized deviations from the population mean for both the GMM and HMM solutions. For ease of interpretation, Tables S1 and S2 in Multimedia Appendix 1 present the HMM solution in raw units for the sleep and physical activity clusters, respectively. This was only done for the HMM, as the GMM and HMM solutions showed high comparability.

Finally, the demographic and clinical characteristics of the states were identified. Descriptive statistics for baseline and follow-up variables were calculated at the observation level, meaning that participants contributing multiple follow-ups appeared multiple times. No formal statistical tests were performed due to the nonindependence of repeated observations. These analyses were exclusively reported on the HMM solution, as the HMM considers repeated measures, and the obtained solutions showed high comparability with those derived from the GMM model.

For the HMM, transition probabilities between states were estimated as part of the model fitting. These probabilities represent the probability of transitioning from one state at a given observation to another state at the subsequent observation within the same participant. Transition estimates were derived from all available consecutive 3-month time points per participant, allowing the model to characterize within-person movement between behavioral states over time. Participants contributed transitions only when sequential observations were available.

Figure S1 in Multimedia Appendix 1 presents the mean LL scores across the 5 folds evaluated on the unseen test data for a given number of clusters. In the GMM solution, LL increased substantially up to 4 clusters, with more modest improvements thereafter. Although the curve continued to rise slightly beyond 4 clusters, the 4-cluster solution was selected to balance model fit, parsimony, and interpretability. Similarly, the HMM solution showed substantial improvement up to 4 states, followed by progressively smaller gains with additional states, indicating diminishing returns beyond this point. Taken together, these patterns support the selection of the 4-cluster solution across both the GMM and HMM solutions.

The 4-cluster GMM and 4-state HMM solutions have similar descriptive characteristics across the digital sleep features used to define the model. The 4-cluster solution for the GMM and the 4-state solution for the HMM are visualized in Figure 2, presenting deviations from the population mean. Table S1 in Multimedia Appendix 1 presents the findings in raw units without scaling for the HMM solution. Overall, these results indicate the following:

Table 4 describes the demographic and clinical characteristics of each state. The clusters presented similar characteristics, with only slight differences observed across variables. For state 1, efficient early sleepers had a slightly higher proportion of female participants than the overall sample. In contrast, in state 4, variable late sleepers had the highest proportion of male participants. When considering employment, state 4, variable late sleepers, had a slightly lower proportion of people employed or furloughed and a higher proportion of unemployed or sick leave individuals compared to the overall. The prevalence of physical health comorbidities varied across states; state 1 (efficient early sleepers) had the least, while state 4 (variable late sleepers) had the highest presence of physical comorbidities. No substantial differences in depression severity and anxiety were observed; all states fell within the moderate range on the IDS-SR, though a 10-point difference was observed between state 1 (efficient early sleepers; 27.1) and state 4 (variable late sleepers; 38.4), highlighting potentially clinically relevant variation within the same severity category. States 2 (efficient late sleepers) and 4 (variable late sleepers) exhibited marginally higher GAD-7 anxiety scores compared to the other states and were slightly more affected by functional disability, indicating significant impairments.

Figure S2 in Multimedia Appendix 1 presents the mean LL scores across the 5 folds evaluated on the unseen test data, for the physical activity models. For the GMM, a clear inflection point is observed at 3 clusters. This suggests that a 3-cluster solution provides the best fit for the data. The HMM LL curve shows an inflection point at 3 and 7 states. Based on these findings, the 3-state GMM and HMM solutions were selected as the optimal number of clusters. The 7-state HMM was also explored and is presented in Table S4 in Multimedia Appendix 1 for further reference.

The 3-cluster solutions for the GMM and HMM models presented similar solutions. Figure 3 describes these differences, presenting deviations from the population mean. Table S2 in Multimedia Appendix 1 presents the findings in raw units without scaling for the HMM solution.

Their demographic and clinical characteristics were explored to examine the distinctions among the 3 delineated activity groups. The states varied only marginally across characteristics. Table 5 presents these differences, showing state 1 (high activity) to have the highest proportion of male participants (n=198, 34%) and instances of no physical health comorbidity (n= 335, 57.6%) compared to any other cluster or the overall sample. State 3 (low activity) had a slightly higher proportion of retired participants than the overall sample (n= 126, 27.6% and n= 383, 21.9%, respectively). Clinical differences are minor, with all states presenting moderate levels of depression (scores between 26 and 38). State 3 (low activity) presented the highest moderate IDS-SR score compared to the other 2 groups. All states presented some functional impairment (scores between 11 and 20) and mild anxiety scores between 5 and 9.

The human sleep-wake cycle is governed by the interplay between sleep homeostasis and circadian processes. Dysregulation of this system, for example, when an individual’s internal circadian rhythm and external behavior do not align, can cause sleep disruptions [46]. With approximately 90% of patients affected, sleep disturbance is one of the most commonly reported symptoms of depression [47,48]. This study aimed to examine these differences via clustering of sleep RMT data, identifying 4 subtypes in the RADAR-MDD sample.

Notably, variable late sleepers were characterized by later mean sleep-wake patterns, higher awakenings, and inconsistent sleep onset, offset, and durations over the 7 days compared to the overall population. They typically went to bed around 1:22 AM (the following day) and woke around 9:30 AM, with over 2 hours of variability in these times each week. Among others, the observed variability in sleep patterns may indicate social jet lag, where differing weekday and weekend schedules disrupt sleep cycles [49]. This variability to meet social and work demands can negatively impact circadian rhythms, increasing the risk of poor mental and physical health [49,50].

Previous literature looking at sleep profiles in depression frequently draws a distinction between morning and evening preference types [51,52]. Seo et al [53] describe this as morning-types preferring earlier bedtimes and wake-up times, while evening-types prefer later bedtimes and wake-up times according to self-report measures. While we cannot infer personal preference due to the passive nature of the Fitbit sleep data collection, we identify distinct differences in sleep onset and offset times, which mirror similar patterns.

The efficient late sleepers presented similar later sleep-wake patterns (as seen in variable late sleepers) but efficient sleep with little time awake compared to the overall sample, with an average sleep onset of 00:47 (the next day) and offset at 8:25. Previous works examining later sleep-wake patterns using wearables have found associations with worse outcomes [22,54-56]. Studies using self-report metrics have found that a morning preference may be protective against depression symptoms [57], while evening preference is associated with worse adverse effects [18,58]. Previous literature indicates rumination as a key psychological mediator in the relationship between depression and later sleep onset [59].

In contrast, both disturbed sleepers and efficient early sleepers showed earlier than average sleep onset and offset times, with average variability in these timings. They typically began sleeping around 23:30 and woke around 7:30, around 1 to 2 hours earlier compared to their late sleep counterparts and experienced less than an hour of variation in both onset and offset times each week. They also differ, with efficient early sleepers presenting with high sleep efficiency with little time awake in bed, while disturbed sleepers present with the highest levels of inefficiency and time awake in bed compared to any other state. It is important to highlight that while these differences appear substantial within the context of the overall sample range, they translate to only modest absolute differences, with approximately 2% more time spent awake in the disturbed sleepers group. Despite the modest differences identified, previous literature has also identified poor sleep efficiency and increased wakefulness after sleep onset in patients with depression compared to controls [60].

A few sociodemographic differences were observed across the sleep subtypes, such as a higher prevalence of physical health comorbidities and a higher proportion of unemployment and sick leave in individuals with variable late sleep compared to the overall sample. These differences were only descriptive and not substantial enough to suggest clear, distinguishing features. The absence of such distinct differences indicates that the sleep subtypes may reflect more stable, inherent characteristics of the RADAR-MDD cohort rather than being heavily influenced by other participant characteristics. The finding that participants were most likely to stay in 1 subtype over time further suggests the presence of trait-like differences, indicating that these latent states may reflect underlying behavioral phenotypes rather than short-term fluctuations within the heterogeneous RADAR-MDD population. Nevertheless, transitions that did occur most commonly involved movement from more extreme profiles toward intermediate patterns (eg, high or low activity toward light or some activity; variable late sleepers toward efficient late sleepers), which may partly reflect regression to the mean as well as genuine behavioral change over time.

Depression severity differed across sleep states, although all states fell within the moderate range on the IDS-SR. State 4 (variable late sleepers) scored 10 points higher than state 1 (efficient early sleepers), representing a potentially clinically relevant difference in symptom severity. These findings are consistent with prior work, finding greater variability in sleep timings and later sleep-wake patterns associated with worse depression [11,22,49]. Together, these findings suggest that sleep subtypes may reflect differences in sleep-wake patterns and behavioral regularity within depression, ranging from stable and efficient sleep patterns to delayed and irregular patterns. By categorizing these relationships into possible subtypes, these findings provide an exploratory basis for a framework that identifies consistent sleep-depression profiles, which might mitigate the challenges presented by symptom heterogeneity in research and support more tailored circadian and sleep-focused approaches aimed at improving sleep-wake timing and regularity, such as structured sleep scheduling, sleep hygiene practices, or light exposure interventions.

In terms of physical activity measures, our subtyping analysis identified 3 distinct groups: a high-activity subtype with above-average levels of light, moderate, and vigorous activity, participating in nearly 1 hour of moderate-to-vigorous activity daily; a low-activity subtype with high levels of sedentary time (roughly 1.5 hours more than the high-activity group) and below-average levels across measures of physical activity with around 18 minutes spent in moderate-to-vigorous activity; and, finally, a group that presented a middle ground between these 2 extremes (light or some activity group), spending an average of 30 minutes daily in moderate-to-vigorous activity.

Previous research has identified various associations, including low gross physical activity and low moderate-to-vigorous physical activity with depression [14,61]. Research suggests the presence of a bidirectional link between low mood and physical inactivity, where symptoms of depression, such as feelings of anhedonia and low mood, can result in reduced motivation to engage in physical activity [62], when, in turn, low physical activity also increases the risk of depression [14]. Taken together, these findings suggest that the activity subtypes may reflect differing levels of behavioral activity and daily structure, ranging from highly engaged patterns to reduced activity consistent with behavioral withdrawal. Such patterns may be relevant for interventions aimed at increasing behavioral activation and daily activity. Employment status represents an important contextual factor, as individuals with depression might take leave from work, resulting in reduced activity levels. This study did not identify any meaningful descriptive differences in depression scores between the low, light, and high-activity subtypes, except for employment, where a slightly higher proportion of retired participants was found in the low-activity group compared to the overall sample.