Figure 1 depicts the study procedures. Study procedures were identical across participants regardless of case status. Participants completed the protocol across a mean time span of 7 weeks.

The sample comprised 36 individuals (75% female [n=27], mean age 19.50 y, SD 3.92 y, range 14‐27 y) recruited from the Boston, MA, area (Table 2). The majority of the sample identified as White (67% [n=29]), with 2 individuals identifying as Hispanic or Latino, 4 individuals identifying as Black or African American, and 1 individual identifying as Asian. Twenty-three individuals had a MDD and were recruited from outpatient psychiatry services at Boston Children’s Hospital. Providers requested consent from the participant and their family for the research team to contact them about the study, and research staff followed up to explain the study and, if appropriate, arrange for a consent visit to take place. The control group (N=13) was recruited from the community via flyers and advertisements placed in Boston. Exclusion criteria for individuals with MDD included a severe neurodevelopmental disorder or other impairment that impacts the participant’s ability to provide the required information for the study (eg, symptom reports), a substance or medication-induced affective disorder, an affective disorder secondary to a medical condition, and a current Axis I psychotic or bipolar disorder. Exclusion criteria for controls included a severe neurodevelopmental disorder or a current or past psychiatric diagnosis as defined by DSM-5 (Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition) criteria.

This study was reviewed and approved by the Boston Children’s Hospital Institutional Review Board (BCH IRB-P00031981). The study involved the collection and analysis of data from human subjects, and all procedures were conducted in accordance with the IRB-approved protocol. All participants provided informed consent/assent using forms approved by the institutional review board at Boston Children’s Hospital. This ensured that participants were fully informed about the study’s purpose, procedures, potential risks, and benefits before agreeing to participate. To maintain participant confidentiality, all collected data were stored in secure databases and deidentified prior to analysis. All data included in the present manuscript or any supplementary materials in Multimedia Appendix 1 are completely anonymous. There was no cost or fee associated with study participation. Participants were compensated for their time and effort in completing the study procedures. They were eligible to receive up to US $144, with payments provided via ClinCard, a secure electronic payment system. Payment breakdown was as follows: US $20 for Structured Clinical Interview for DSM-5; US $20 for questionnaire completion; US $14 (US $1 per day) for FitBit wearing; US $14 for FitBit return; US $14 for usability questionnaire; and US $56 (US $2 per check-in for EMA). Participants were paid for all aspects of the study that they completed, including individual EMA check-ins. No identifiable images of individual participants are included in this manuscript or supplementary materials in Multimedia Appendix 1.

The presence or absence of major depression diagnoses (and comorbid disorders) was confirmed using the Structured Clinical Interview for DSM-5 [42] (with additional modules from the Kiddie Schedule for Affective Disorders and Schizophrenia interview [43] if<18 years of age) in order to confirm past or current MDD (and other DSM diagnoses). Depressive symptoms were assessed at study intake using the patient health questionnaire (PHQ-9) [44]. Participants were also administered the Brief Psychiatric Rating Scale [45], generalized anxiety disorder (GAD-7) [46], and the Pittsburgh Sleep Quality Index [47]. Parents of probands were administered the Hollingshead Index of Socioeconomic Status [48], and 25 parents agreed to complete the scale.

EMA data were recorded using LifeData [49]. LifeData is a smartphone-based EMA program that runs on both Android and iPhone smartphones. Check-ins are delivered via the LifeData mobile app RealLifeExp. EMA lasted 2 weeks, during this time participants received a push notification twice per day (6 AM and 4 PM), reminding them to complete a check-in. Notifications were sent 4 times (at 1 h intervals) per check-in, allowing each participant a 4-hour window to complete the check-in (6-10 AM and 4-10 PM).

For each check-in, participants were asked the same 8 questions. Participants rated the items using a slider, which translated to a score (0‐100): “I am sad.”; “I feel bothered by every little thing.”; “I have no interest in things I would usually enjoy (eg, food, TV, games, spending time with friends/family).”; “I do not have enough energy to get going.”; “I have no appetite or feel much hungrier than usual.”; “I feel physically tense and/or jittery.”; “I am nervous, anxious or on edge.”; and “I can’t stop worrying.” Items were designed to be reflective of established measures of depression (PHQ-9) [44] and anxiety (GAD-7) [46], with the adaptation that they could be administered multiple times per day (Table S1 in Multimedia Appendix 1 for an explanation of each item). For a subset of statistical models (see below), the first 5 items were classed as measuring mood and the rest as measuring anxiety. Once per week, during an afternoon check-in, participants completed a checklist of stressful life events. The question reads “Did any of the following happen to you since the last time you completed one of these surveys? Check all that apply”. The participant selects as many of the following options as appropriate: “I argued with a friend or family member.”; “I was not allowed to do something I wanted to do.”; “I got a bad grade in school.”; “My parents have been arguing a lot.”; “Somebody in my family got a serious illness.”; “Someone in my family was arrested.”; “Somebody teased or threatened me.”; “I teased somebody else.”; “I did something that made me feel embarrassed.”; “Someone commented negatively on the way I look.”; “I was excluded from a group event.”; and “I got disciplined or suspended from school.”

Participants were issued a FitBit Charge 3 and were instructed to wear it for the duration of the EMA portion of the study. Twenty-two individuals (11 controls (mean age 22.90, SD 3.27 y, 10 females) and 11 MDD cases (mean age 18.20, SD 2.86 y, 7 females) wore the FitBit with sufficient regularity for data to be analyzed. Data were collected on daily activity (number of steps) and sleep.

Fitbit data were obtained via the Fitbit API, which provides preprocessed JSON files reflecting proprietary algorithms for step counting and sleep [50]. For sleep, each participant had one JSON file that spanned the duration of the study, which included a sleep log with timestamps for sleep onset and offset and a breakdown of sleep stages (eg, light, rapid eye movement [REM], deep). For steps, each JSON file contained minute-level step count data linked to a timestamp. Data in the JSON files were converted to csv using a combination of standard and pandas [51] libraries in PyCharm [52] (Python version 3.9; Python Software Foundation) and rearranged for analysis using tidyr [53] and dplyr [54] in R [55]. Following others’ work [56] in the field, we extracted numerous sleep features from the data collected using the Fitbit device. These are related to sleep architecture, quality, and stability (Table S2 in Multimedia Appendix 1 contains details or individual metrics). A nap was defined as time spent in bed lasting<180 minutes, FitBit does not calculate deep, light, and REM sleep during these shorter sleeps. Therefore, naps were included in analyses focused on total sleep time (TST) and time in bed (TIB) but were excluded from analyses focused on sleep stage. Sleep quality metrics were calculated per all logged sleeps per individual, and sleep stability metrics were derived across all logged sleeps. Steps that were recorded during a documented period of sleep were recoded to 0 as they were likely due to movement in bed rather than steps.

At the end of participation, all participants completed a short questionnaire that evaluated their experience using the EMA app (eg, if the app was easy to use; see in Multimedia Appendix 1 for a list of questions). All but one of the statements were rated from 1 to 7, with 1 being strongly disagree, 4 neutral, and 7 strongly agree. Responses were recorded as disagree (1-3) and agree (5-7). Analysis included frequencies and proportions of responses per item. One item, relating to where participants used the app, included a free text response (unique responses are included in the Multimedia Appendix 1).

All analyses were conducted in R (version 4.2.1; R Foundation for Statistical Computing) [55].

In line with others’ EMA work [57], we opted to examine variability in depressive symptoms both visually and statistically. Time series data were examined using standardized (group-mean-centered) and raw data. Group-mean-centered plots depict the relative variability over time across the sample. Before plotting data were group-mean centered (by subtracting each participant’s mean from each individual score), which results in standardization of the figure such that units are in units of deviation from each individual’s average (set to 0). Per subject, raw score time-series plots depict the fine-grained differences between each individual’s data. Data were plotted using ggplot2 [58].

We used several statistics to quantify variability, including the root mean square of successive differences (RMSSD) [59], coefficient of variation (CoV) [60], and intraclass correlations (ICC). The RMSSD reflects variability over time. The CoV reflects the dispersion of the data and is useful for comparing variability when means differ (eg, in cases vs controls). The ICC reflects the proportion of variance attributable to between-person variability, thus, 1-ICC indicates the proportion of variability attributable to within-person variability. Variability statistics were calculated using a combination of base R and the psych [61] package.

The EMA data were nonindependent, for example, we might expect data collected within the same person to be more strongly related than those collected from 2 different people. This structure is sometimes referred to as multilevel; observations are nested within item type (mood or anxiety), which are nested within the time of day (AM or PM) and day, which are nested within the participant. Therefore, when evaluating the impact of various predictors on EMA scores; we applied linear mixed-effects regression to the data using the package lme4 [62]. Effect sizes for fixed effects (ηp2) were calculated using the effectsize package [63], and were interpreted as small effect: ηp2 ≈ 0.01, medium effect: ηp2 ≈ 0.06, large effect: ηp2 ≈ 0.14. Similarly, observations for sleep architecture and quality were nested within a day (where some individuals logged sleep with onset times within the same day), within the sleep stage, and within the participant. Multilevel modeling is robust to missing data, and therefore we did not impute the small amount of missing data that was present in the EMA, of 1008 expected check-ins, 923 were observed (overall compliance=92%). Prior to testing the effect of various predictors on depression, we tested an intercept-only, or unconditioned model to evaluate the need for a random effect. Significant random effects (determined using the lmerTest [64] package) were retained in subsequent models. Given the different distributions of age and sex in case and control groups, these variables were included as covariates in all models.

First, we asked whether individuals with a MDD diagnosis demonstrated greater symptom severity (indexed using EMA score) and variability than controls. We also evaluated the impact of sex and age on severity, as well as fixed effects of Time of Day and Item Type. Second, we examined whether anxiety and mood covaried over time and whether anxiety or mood at the preceding check-in significantly predicted mood at the following one. Third, we asked whether life stress was associated with increased symptom severity. Fourth, we examined whether activity levels (number of steps) were associated with MDD diagnosis or symptom severity. Fifth, we assessed whether number of minutes asleep, sleep onset, or sleep-waking was associated with MDD diagnosis or symptom severity. Sixth, we examined the results of the usability questionnaire.

The total unique check-ins completed by participants=923 (mean check-ins per participant 25.60, SD 2.78). Assessments were completed across a total of 486 unique days (mean days per participant 13.50, SD 1.25), and yielded 1.90 mean check-ins per participant per day. For the 2 required check-ins per day compliance rate of 91.57%, cases demonstrated slightly lower compliance (89.94%) than controls. Across individual EMA items, this yielded 7336 data points for regression modeling.

Table 3 shows the descriptive and variability statistics for the sum score of all EMA items as well as for the individual items. In terms of reported symptom severity, MDD cases scored higher than controls on all items. Examination of variability statistics (Table 3) indicates that there was considerable variability in both the sum score and the individual items. The RMSSD and CoV suggest greater variability in MDD cases than in controls. Examination of the ICC across items suggests that ≈50% of the variance in item-level responses was due to within-person variance in MDD cases.

Figure 2 shows the group-mean centered time series plot of the sum score. At the group level, scores fluctuated in a saw-tooth pattern, with repeated rises and falls but no clear linear trend of improvement or worsening over time. This pattern reflects short-term variability in EMA data and highlights overall fluctuations in symptom severity across the sample. Item-level data followed a similar pattern (Figure S1 in Multimedia Appendix 1); in MDD cases, variability in symptom reporting did not appear to vary as a function of item. Controls appear to report a greater degree of variability to the irritability and anhedonic items (Bothered and Interest) than they did on mood or anxiety symptoms.

At the individual level, we observe heterogeneity (Figure 3) where some individuals with a MDD diagnosis demonstrate worsening throughout the course of the study (eg, 30,001, 30,015, 30,028, 30,030, 30,044, and 30,048), and others appear to improve (eg, 30,035, 30,036, and 30,047). Interestingly, not all controls demonstrated floor effects. While controls reported more zero responses than cases (Table 3) some endorsed minimal symptoms (eg, 30,011, 30,033, 30,038, 30,046, and 30,057) throughout the course of the study; admittedly these symptoms would likely amount to sub-threshold symptoms for a MDD diagnosis but this suggests that healthy controls can demonstrate variation in depressive symptomatology. Three MDD cases (30,003, 30,012, and 30,019) demonstrated minimal symptoms; diagnoses for these individuals were in line with low symptoms (Other Specified Depressive Disorder [Insufficient Symptoms], Recurrent MDD, and Past MDD Single Episode, respectively).

The unconditioned model indicated significant between-subject nested effects of Participant, Day, and Item Type but not Time of Day (Table S3 in Multimedia Appendix 1). All significant random effects were retained in subsequent modeling.