Setting

The DEW study is housed at the Centre for Addiction and Mental Health (CAMH) -an academic mental health provider- within the Child, Youth and Family Mood and Anxiety Team,

DEW study recruitment and participants

Youth aged 12–21 with current or currently remitted clinically diagnosed DSM-5 MDD [25] were recruited for the DEW longitudinal digital phenotyping study. Participants are referred to the DEW study if among their presenting problems is a clinical diagnosis of ‘depression’ (both current and/or past). To be included in the study, participants had to meet the criteria for a current/past diagnosis of DSM-5 MDD, as assessed by a trained Research Coordinator with the DIAS-C structured psychiatric interview [26] (see below). Youth are excluded from participation into the DEW study if they have current or past substance use disorder whose severity is rated more than moderate by DSM-5 criteria, active psychotic symptoms, bipolar disorder, epilepsy, autism spectrum disorder, multiple sclerosis, paraplegia or spinal cord injury, juvenile rheumatoid arthritis or other major autoimmune disease, chronic renal failure, inherited metabolic disorders, or active cancer. Following clinicians’ referral, youth were recruited from: (a) the Mood and Anxiety clinic within CAMH’s Child, Youth and Emerging Adult Division, (b) online classified advertising including our hospital’s research recruitment page, and (c) referrals by professional health care providers in the greater Toronto area.

Participants were assessed for capacity to consent. Capacity of each potential participant to provide consent was assessed by a study team member trained by the PIs. Following the consent discussion, questions were asked to the participants to ensure they understood all study procedures, risks and benefits, as well as their rights as a volunteer in this research study. If the potential participant did not demonstrate such capacity during the informed consent process, the assumption of capacity to consent was not validated and the informed consent discussion did not continue. In this case, the potential participant was not considered eligible for the study.

In addition, non-CAMH recruited participants were asked for consent for DEW research personnel to contact their health professional to confirm their diagnosis of MDD, or having experienced an episode of MDD. CAMH participants’ diagnoses were confirmed through inspection of their medical charts.

Data collection

Participants into the DEW study are administered clinical diagnostics at baseline, including: the DIAS-C (a semi-structured interview originally designed for familial-genetic studies that covers an extensive array of diagnoses) interview [26] to confirm their clinical MDD diagnosis. At baseline and at every following quarterly follow-up (named ‘arm’ in the present paper, taking place every 3 months), measures are collected via the following rating scales/questionnaires: the Patient Health Questionnaire adapted for adolescents (PHQ-9-A [22]),, and the Ruminative Response Scale (RRS [27, 28]). The PHQ-9-A consist of 9 items that correspond to the 9 DSM-IV criteria for MDD; it has comparable sensitivity and specificity to longer depression measures [22], and it was derived from the PHQ-9 [23], which is a popular clinical instrument to assess depressive symptoms both in psychiatric and in primary practice. The RRS was originally derived from the Response Style Questionnaire (RSQ [27]),, which included a self-report measure of rumination. A shorter, 8 item version of the RRS [24] showed good ability to measure within-person variation in rumination, so that a shorter version of the RRS was adopted in the DEW study too.

Following the administration of the aforementioned psychometric measures, 4 weeks of passive WA measurements are collected according to the DEW protocol. After the 4-week WA data collection, participants return the wearable device and complete the PHQ-9-A online using a REDCap survey [29]. The whole cycle of 4 weeks of passive data collection and PHQ administration is then repeated at every DEW quarterly follow-up.

The present report is based on the first 72 participants in the DEW study, recruited over the first 31 months of the study, and on the PHQ-9-A and the RRS scores filled at the recruitment of the DEW study (i.e., at baseline). Given the 2-year longitudinal nature of the DEW study, the maximum number of quarterly follow-ups is 8 (with baseline considered as point 1).

Devices and technology

For each of the quarterly WA collections of the DEW study, the GENEActiv triaxial accelorometer original (Activinsights, Cambridge, United Kingdom) device is worn on the non-dominant wrist for 4 weeks (30 days) https://activinsights.com/technology/geneactiv/. The GENEActiv device has been validated against reference methods and proved reliable and valid by several studies [30,31,32,33]. Data were collected in person at CAMH, and Research staff were available to assist participants in clarifying questionnaires while completing study measures.

Geneactiv devices were set up in person with the participant. Geneactiv devices are lightweight and waterproof so that participants are able to wear them continuously throughout the day. Participants were given an instruction page about the device and were also encouraged to contact study staff if they had any problems with the devices, or if they were not able to wear them for the expected time frame. If participants contacted study staff, any issues or disruptions were noted.

Data related to sleep and activity were collected through self-report measures answered via mobile phones (EMA data) while participants were wearing the Geneactiv devices.

Data collected through the Geneactiv devices were verified by systematically checking: (a) the length of time participants wore the devices for (i.e., the consistency between the duration of data recording and the amount of time the device was worn); (b) near-body temperatures (these had to be ≥ 27 degrees Celsius as recorded by the Geneactive device, to ensure that the data received were a result of the participant’s actually wearing the device). In order to support the validity of Geneactiv data, nonconformities of sleep duration collected by the Geneactiv were also checked against self-reported EMA data, whereby participants were asked how many hours/day they had slept in any given day. Inconsistent matching between these sources led to data removal.

Whenever a discrepancy attributable to technical (e.g., battery insufficiently charged/dysfunctional) reasons was found, we adopted the Geneactiv Active Insights trouble shooting guide https://activinsights.com/technology/geneactiv/ to solve the issue.

The GGIR Package was used for processing WA data from GENEActiv devices and to extract actigraphy variables [33] listed in Table 1. Activity (in milligravity) was derived by the minute-level accelerometry count [12].

Further details about these parameters can be found in Appendix A.1.

Institutional Research Board (IRB) approval was granted by the CAMH Research Ethics Board.

Table 1 Variables used in the analysesData analyses

Analyses were carried out with the R software version 1.4.1717 [34].

Aim 1: assessing correlation among WA features

We first assessed the correlations between WA variables (Activity, Sleep Duration, Sleep efficiency) by building a correlation matrix among the 3 WA variables’ grand means for each participant, with the Hmisc package in R [https://cran.r-roject.org/web/packages/Hmisc/index.html]. While this oversimplifies the picture by averaging within-individual variability in time, it provides a first, at-a-glance description of the correlations among-variables.

Aim 2: identifying trajectories using latent class analysis

In the light of the heterogeneity of the phenotype of MDD, to define classes of individuals with discrete psychophysiological patterns, we adopted LCA. This type of mixture modelling is often favoured in person-centred approaches like ours, as it is particularly apt at identifying latent subpopulations on the basis of their responses to observed variables. Specifically, LCA is suitable for identifying different classes within a heterogeneous population such as AMDD, and with complex data such as actigraphy data, in which patterns may not otherwise be easily identified.

We ran LCA models for each of the 3 WA variables using the LCMM Package in R ( https://cecileproust-lima.github.io/lcmm/ [34] across all available quarterly assessments. Models were run for solutions between 1 and 6 Latent Classes, with the selection of the best-fitting model based on a comprehensive evaluation of the following parameters: the Bayesian information criterion (BIC), Akaike Information Criterion (AIC) and Lo-Mendell-Rubin ad-hoc adjusted likelihood ratio test (calculated in R by the calc_lrt function), plus a criterion of parsimony, whereby no solutions were deemed acceptable if any cell (class) contained < 5% of subjects. The selected Latent class models were then visualized, with R code adapted from https://mcfromnz.wordpress.com/2011/10/02/latent-class-mixture-modeling-with-graphics/. The model in R can be described as follows: latent class model <- lcmm (dependent.variable ~ day + arm, mixture = ~ day, subject = participant.ID, ng = 2/3/4/5/6). Further details of these procedures can be found in Appendix B.

To estimate the degree of interdependence among the classes of the 3 WA variables, Cramer’s V coefficients were calculated using SPSS 27 [35].

Aim 3: predicting latent class membership through clinical questionnaires

We assessed the ability of PHQ-A’s scores at intake to predict membership into the WA classes by multinomial regression, where: (a) the PHQ-A scores at intake dichotomised into ‘clinical/non-clinical’ (PHQ-A > 14), and: (b) sex, were the predictors, and the most probable membership in every readout’s trajectory was the dependent variable, with age as a covariate.

Since rumination is a frequent associated feature of AMDD, we similarly assessed the ability of RRS scores at baseline to predict membership in WA class as the dependent variable by regression.

To corroborate our findings relative to Aims 2) and 3) (see below) via an approach that is alternative to LCA, we analysed the dataset with Linear mixed effects model [36], using random intercepts and the respective variables as fixed effects predictors. Differently from the LCA, this approach allows for estimating the effects of predictors on variation of accelerometry data per participant in time, without constraining individuals within any specified class.

For every accelerometry variable, we first ran an unconditional model, which only included the ‘arm’ and the ‘night’ variables as the predictors/covariates. Two versions of this unconditional model were run: one with the ‘arm’ variable treated as a linear time predictor (i.e., a version of the model that simply tested whether or not a significant temporal trend was discernible in the data), and another version that treated ‘arm’ as a semi-continuous predictor (i.e., a version of the model that tested whether one or more arms significantly differed from arm 1). Because of these features, one may consider these two versions of the unconditional model as nested one into the other. These two versions were compared one to the other, so that a successive conditional model with: age, sex, PHQ-9 (treated as a dichotomous variable: ‘clinical/non-clinical’ PHQ-A > 14), and RSS, could be run on the version of the model that provided the best fit. The ‘night’ variable was treated as a 30-level categorical covariate, but was not included in the summary.

Based on previous findings in community children [21] and in adults with MDD [12], we hypothesised associations between PHQ scores in the clinical range and: (a) reduced physical activity, and (b) shorter sleep duration.