Sleep health dimensions from wearables and transdiagnostic mental health in young adolescents

Abstract

Importance:

Sleep behavior markedly shifts in adolescence, increasing vulnerability to mental health disorders. While sleep health is understood to be multidimensional, adolescent-specific sleep health dimensions have not been empirically validated, and their relevance to transdiagnostic mental health outcomes is unknown.

Objective:

To identify sleep health dimensions using Fitbit devices in a large sample of young adolescents and assess concurrent and prospective associations between sleep health dimensions and transdiagnostic mental health outcomes.

Design:

A longitudinal cohort study using data from the Adolescent Brain Cognitive Development Study (data release 5.1, collected 2018–2020) including early adolescents (age 11–13 years) within the United States. We used exploratory factor analysis (EFA) to identify sleep health dimensions and confirmatory factor analysis (CFA) to confirm the factor structure in an independent subsample. We used linear mixed-effects models to test concurrent and prospective associations between sleep dimensions and mental health outcomes at 1-year follow-up.

Setting:

Multicenter community setting.

Participants:

3,393 participants from the Adolescent Brain Cognitive Development Study.

Exposures:

Objective sleep data collected for up to 21 days (range 7–21) using wearable Fitbit devices.

Main Outcomes and Measures:

Transdiagnostic mental health outcomes assessed via the Child Behavior Checklist and Brief Problem Monitor (internalizing and externalizing symptoms), Prodromal Questionnaire-Brief Child Version (psychotic-like symptoms), and 10-Item Mania Scale (mania symptoms).

Results:

3,393 participants were split into EFA and CFA subsamples. We identified 6 factors using EFA: Irregularity, Timing, Social Jetlag, Duration, Weekend Oversleep, and Continuity. CFA confirmed this factor structure. Greater sleep Irregularity was associated with transdiagnostic mental health symptoms cross-sectionally, but not prospectively. Shorter Duration was associated with total, internalizing, externalizing and attention symptoms cross-sectionally, and total, attention and psychotic symptoms 1 year later.

Conclusions and Relevance:

Wearable Fitbit data provide empirical support for multidimensional frameworks of sleep health in adolescence. Although effect sizes were small, sleep Irregularity and Duration emerged as key dimensions with relevance to mental health. These findings establish a foundation for myriad future investigations, including examining within-person patterns of the 6 dimensions; extending to older adolescence; investigating associations with other health outcomes; replicating with research-grade actigraphy devices; and suggesting potential targets for pediatric sleep interventions.

Introduction

Adolescence is a period of alarming increases in mental health risk^1–3. Half of all mental health conditions emerge by age 14, and by age 15 mental illness contributes to more disability than any other illness class in the United States^2,4. Adolescence is also a time of dynamic changes in habitual sleep health patterns. Biologically-driven delayed bedtimes converge with early wake times during the school week, which can reduce sleep duration, increase sleep timing variability, and amplify weekend–weekday shifts in both duration (‘weekend oversleep’) and timing (‘social jetlag’)^5–10. These co-occurring changes in sleep and mental health, combined with the modifiable nature of sleep¹¹, highlight sleep health as an important risk marker for detection of transdiagnostic mental illness during adolescence¹².

Despite this potential, the specific behavioral sleep features that relate to transdiagnostic adolescent mental health symptoms are not well understood. Adolescent sleep studies often focus on single or limited sleep metrics¹³, yet sleep health is a multidimensional construct with dimensions occurring simultaneously within an individual¹⁴. In adults, sleep health is captured by the Ru-SATED model, comprising Regularity, Satisfaction, Alertness, Timing, Efficiency and Duration¹⁴. Ru-SATED has been validated using self-reported sleep measures^15–17 and modified versions were validated with objective actigraphy data in middle-aged and older adults^18–20. It was also adapted for pediatric samples by incorporating an expanded definition of Regularity to sleep-related Behaviors - emphasizing the influence of regular bedtime routines, parent-child interactions, and environmental factors such as caffeine and technology use on youth sleep patterns (the Peds B-SATED model²¹). However, behavioral sleep health dimensions have not been empirically determined in adolescents.

The large and diverse Adolescent Brain Cognitive Development (ABCD) study²² provides an opportunity to identify population-level sleep health dimensions in early adolescence (ages 11–13). Focusing on this age window – which coincides with escalating mental health risk – could yield important insights and guide treatment targets for existing or novel interventions. This would extend prior work in the adolescent sleep-mental health field, which has generally used broad age ranges^23,24 that may obscure identification of developmentally-specific sleep health dimensions during a time of dynamic shifts in sleep and mental health. The ABCD study includes a 21-day Fitbit protocol – a commercial wearable activity tracking device that enables longer-term objective assessment of sleep-wake behaviors at scale. Such longer-term monitoring facilitates the distinction of day-to-day irregularity versus weekend-weekday shifts in sleep timing and duration.

Our aims were to use the ABCD study to: (1) empirically identify clinically meaningful and interpretable early adolescent behavioral sleep health dimensions measured via Fitbit; and (2) comprehensively map associations between behavioral sleep health dimensions and transdiagnostic mental health during this developmentally sensitive age window. For Aim 1, we hypothesized at least four behavioral sleep health dimensions reflecting regularity, timing, efficiency and duration, consistent with objective dimensions of the Ru-SATED framework identified in adults and older adults and suggested in the Peds B-SATED model^18–21. Given our early adolescent sample, we also considered measures of weekend-weekday shifts in sleep duration and timing (weekend oversleep and social jetlag), although we did not have a priori hypotheses regarding whether they may contribute to the regularity dimension or form one or more distinct dimensions. To accomplish this aim, we used exploratory factor analysis (EFA) to empirically identify the dimensionality of the Fitbit sleep features and confirmatory factor analysis (CFA) to subsequently validate the sleep health dimensions identified through EFA. For Aim 2, we characterized associations between the identified sleep health dimensions and transdiagnostic mental health outcomes, both concurrently and prospectively at one year follow-up. We examined a broad spectrum of both self- and guardian-reported psychopathology outcomes to capture the most comprehensive and replicable psychopathology symptoms associated with behavioral sleep health dimensions. We hypothesized that worse sleep health would be associated with worse transdiagnostic mental health, with the strongest effect sizes for self-reported mental health measures (consistent with prior work²⁵).

Methods

Participants

The ABCD study is a longitudinal cohort of 11,878 children aged 9–10 years at baseline, recruited from 21 U.S. study sites. Annual assessments collect a range of mental health, cognitive, physiological, and neurocognitive measures. During the 2-year follow-up (age 11–13 years), participants were invited to participate in a 21-day Fitbit protocol. We used data from this assessment for our analysis (release 5.1, collected 2018–2020). N=5,108 participants completed the Fitbit protocol. During processing, we excluded data collected outside the 21-day collection period (N=2,254 nights across 2,110 participants) and entries that were invalid or incorrect (N=8,175 nights across 3,679 participants). After processing, we excluded individuals with insufficient data (<2 weekend days, <5 weekday days, or <7 total days; N=1,667), resulting in a final sample of N=3,393 (median 13 [range 7–21] days of Fitbit data, Table 1). See eFigure 1 in Supplement for participant flow chart, and eTable 1 for comparison to the full ABCD sample. Study procedures were approved by a central institutional review board at the University of San Diego, California. Parents/guardians provided written informed consent and children provided written assent. Study reporting followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines²⁶.

Table 1.

Sample Description.

	EFA Sample N = 1,697	CFA Sample N = 1,696	Effect sizea
Age (median, (Q1-Q3))	12.00 (11.42–12.50)	12.00 (11.42–12.50)	0.016
Sex assigned at birth (N, %)			0.022
Male	849 (50%)	871 (51%)
Female	847 (50%)	825 (49%)
Intersex-Male	1 (<0.1%)	0 (0%)
Race (N, %)			0.018
Asian	48 (2.9%)	53 (3.2%)
Black/African American	125 (7.5%)	133 (7.9%)
Multi/Other	282 (17%)	264 (16%)
White	1,219 (73%)	1,224 (73%)
Unknown	23	22
Hispanic ethnicity (N, %)			0.021
Yes	338 (20%)	312 (19%)
No	1,329 (80%)	1,367 (81%)
Unknown	30	17
Maternal Education (N, %)			0.057
Elementary school	77 (4.5%)	53 (3.1%)
Secondary school	125 (7.4%)	118 (7.0%)
Partial college/associate degree	393 (23%)	456 (27%)
Bachelor’s degree	575 (34%)	534 (32%)
Graduate degree	525 (31%)	532 (31%)
Unknown	2	3
Puberty (median, (Q1-Q3))	3.00 (2.00–3.00)	3.00 (2.00–3.00)	0.015
Unknown	60	71
Any Fitbit tracking during school break (N, %)			0.004
No	1,340 (79%)	1,333 (79%)
Yes	357 (21%)	363 (21%)
Weekday:Weekend days (median, (Q1-Q3))	71.43 (66.67–76.47)	71.43 (66.67–76.47)	0.016
Number of Fitbit tracking days (median, (Q1-Q3))	13.00 (10.00–16.00)	13.00 (10.00–16.00)	0.005

^aEffect sizes calculated using Cramer’s V (categorical variables) and Wilcoxon signed-rank test (continuous variables). Abbreviations: Q1, first quartile; Q3, third quartile.

Measures

Fitbit data collection & processing

Participants wore a Fitbit Charge HR2 continuously for up to 21 days, which collected data at a one-second sampling rate²². Data were uploaded to the participant’s or guardian’s phone and stored on the Fitbit mobile application when the device was in close proximity to the phone. Data were then transferred to an online database, where ABCD investigators extracted 10 daily sleep-related variables²². From these, we computed 24 summary Fitbit variables (i.e., characterizing sleep across the 21-day protocol) that were i) commonly used in the field; ii) clinically and/or scientifically meaningful^{13,23,27–31}, iii) potential targets for sleep health interventions³²; iv) potentially valid using Fitbit measurement³³; v) potentially relevant within the Peds B-SATED²¹ and Ru-SATED¹⁴ frameworks (eTable 2; further details in eMethods). We conducted several processing steps to remove implausible values (see eMethods), and coded all Fitbit variables as ordinal (10 deciles) to minimize the effect of outliers and skewed distributions in the EFA and CFA models, consistent with prior work^18,19.

Psychopathology symptoms

We obtained raw scores from the parent-reported Child Behavior Checklist (CBCL³⁴), including the total score and eight syndrome subscales: anxiety/depression, withdrawal, somatic complaints, thought problems, social problems, aggression, attention problems and rule breaking. We assessed self-report psychopathology using the Brief Problem Monitor (BPM³⁵), including the total score and three subscales: internalizing, externalizing and attention problems. We used the parent-reported ten-item Mania Scale³⁶, derived from the Parent General Behavior Inventory³⁷, to assess mania symptoms. We used the total severity score from the self-report Prodromal Questionnaire-Brief (Child version)³⁸ to assess psychotic-like symptoms. Due to non-normality of sample distributions, we transformed all mental health variables with square-root transformation prior to modeling, and confirmed normality by examining residual Q-Q plots. We also removed sleep-related items from specific CBCL subscales to reduce tautology given Fitbit sleep dimensions as our main predictors of interest (details in eMethods).

Covariates

Caregivers reported on sex assigned at birth, ethnicity and race. We collapsed the race variable from fifteen possible categories into four (black/African American, Asian, other/multiple and white). We used maternal education to assess socioeconomic status. We assessed pubertal status using the self-reported Pubertal Development Scale³⁹. Higher scores indicate more advanced pubertal development. We also derived the following variables from the Fitbit data and included them as additional covariates: number of Fitbit tracking days, ratio of weekday to weekend tracking days, and whether any of the Fitbit tracking occurred during a school break (binary variable).

Statistical Analysis

Factor analysis

We used a split-sample approach to allow for data-driven exploration of the factor structure followed by validation. Specifically, we randomly split the analytic sample into two equal halves: one for exploratory factor analysis (EFA; N=1,697) and the other for confirmatory factor analysis (CFA, N=1,696). We conducted all analyses in R, version 4.5.2. We performed EFA using the fa() function within the psych package, and CFA using the cfa() function within the lavaan package. For EFA, we used minimum residual estimation (appropriate for ordinal distributions⁴⁰) and oblimin rotation, ensured the data were appropriate for EFA using Bartlett’s Test of Sphericity (p<.05), and estimated sampling adequacy using the Kaiser-Meyer-Olkin test (minimum >0.6)⁴¹. To identify the optimal number of factors, we used visual inspection of scree plots and parallel analysis. Parallel analysis estimates the optimal number of factors by comparing the eigenvalues between real and simulated data of the same size⁴². We assigned a variable to a factor if it had a loading >0.40⁴³.

To validate the factor structure empirically identified in the EFA, we fit the CFA using maximum likelihood estimation with robust (Huber-White) standard errors, assessing goodness of fit based on root mean square error of approximation (RMSEA), Comparative Fit Index (CFI), Tucker-Lewis Index (TLI), and standardized root mean square residual (SRMR). RMSEA and SRMR values <0.08, and CFI and TLI values >0.95, indicate excellent fit⁴⁴.

Associations with psychopathology symptoms

Prior to modeling, missing covariate data (6.54% of participants) were multiply-imputed using the mice package⁴⁵ (M=20). We followed the ‘missing imputation, then deletion’ approach, which is recommended over methods that retain the imputed outcome data⁴⁶. Accordingly, we imputed covariates and mental health outcomes for the full sample (N=3,393), then removed the imputed outcome data. Primary analyses used complete cases (fully observed or imputed covariates, factor scores and outcomes), yielding consistent sample sizes across models (N=2,779). See eTable 3 for comparisons of included versus excluded participants.

We used linear mixed-effects models within the full sample (EFA and CFA samples combined) to examine associations between EFA-derived factor scores and psychopathology symptoms. We regressed each psychopathology outcome on all factor scores simultaneously, adjusting for covariates, with family nested within ABCD site as random effects. We ran separate models for cross-sectional and longitudinal (1-year follow-up) outcomes. Longitudinal models also adjusted for the concurrent baseline psychopathology outcome. Because covariates were multiply-imputed, we pooled estimates across imputation sets following Rubin’s rules⁴⁷.

We extracted standardized betas for effect sizes and adjusted for the False Discovery Rate (FDR) using the Benjamini and Hochberg method⁴⁸, correcting across 90 comparisons for cross-sectional models (6 dimensions and 15 outcome variables) and 84 comparisons for longitudinal models (6 dimensions and 14 outcome variables; mania symptoms were not assessed longitudinally).

As a secondary approach to enhance clinical accessibility and replication in future research, we selected the highest loading variable from each CFA-derived factor as a representative indicator and used these variables in analyses with psychopathology symptoms (details in eMethods).

Finally, in sensitivity analyses, we i) used all available cases for each mental health outcome, allowing sample sizes to vary by model depending on the availability of outcome data (N=3038–3389); ii) examined each sleep dimension in a separate model when predicting psychopathology, and iii) included quadratic effects of sleep dimensions in predicting psychopathology. We modeled quadratic effects only for sleep dimensions with established empirical support (Timing, Social Jetlag, Duration and Weekend Oversleep^49–51).

Results

Sample descriptives

EFA and CFA subsamples did not significantly differ on almost all demographic variables (Table 1). The analytic sample (N=3,393) was younger, less racially diverse, and had higher socioeconomic status, lower pubertal development, and lower mental health symptom levels than the full ABCD sample (eTables 1&3).

Exploratory Factor Analysis

Bartlett’s test of sphericity was significant, indicating the factor analysis could help reduce the data in a meaningful way, and the sampling adequacy was appropriate (Kaiser-Meyer-Olkin test = 0.78). The scree plot and parallel analysis test indicated a six-factor solution as the optimal solution for the EFA (eFigure 2). All variables loaded strongly (≥0.64) onto at least one factor, and all were retained (factor 1 loadings 0.64–0.98; factor 2 0.96–0.98; factor 3 0.95–0.97; factor 4 −0.86–1.01; factor 5 0.68–0.93; factor 6 0.82–0.94; Table 2). EFA factors were weakly correlated (median |r|=0.09, range=-0.19–0.11), indicating a relatively distinct set of sleep health dimensions (eTable 4).

Table 2.

Factor loadings from EFA.

	F1	F2	F3	F4	F5	F6
1. Irregularity
SD sleep duration	0.982	−0.062	−0.014	0.01	−0.01	−0.008
SD time in bed	0.979	−0.058	−0.013	0.011	−0.006	−0.014
SD total sleep time	0.958	−0.054	−0.001	0.012	−0.002	−0.005
SD get-up time	0.721	0.147	−0.04	0.022	0.004	0.008
SD sleep offset time	0.719	0.149	−0.045	0.017	0.007	0.005
SD in-bed time	0.644	0.126	0.097	−0.116	0.009	0.029
SD sleep onset time	0.641	0.124	0.095	−0.117	0.005	0.032
2. Timing
Get-up time	0.053	0.943	−0.013	0.241	0.007	−0.027
Sleep offset time	0.054	0.942	−0.012	0.244	0.007	−0.022
In-bed time	0.031	0.823	0.01	−0.389	0.002	0.04
Sleep onset time	0.031	0.822	0.017	−0.381	0.004	0.056
3. Social Jetlag
Social Jetlag of in-bed time	−0.011	−0.01	0.927	−0.019	−0.251	0.014
Social Jetlag of sleep onset time	−0.008	0	0.918	−0.02	−0.254	0.018
Social Jetlag of get-up time	0.01	−0.004	0.68	−0.028	0.645	0.024
Social Jetlag of sleep offset time	0.015	0	0.678	−0.024	0.647	0.027
4. Duration
Total sleep time	−0.042	0.015	−0.013	0.984	−0.016	0.151
Sleep duration	−0.039	0.01	−0.013	0.957	−0.01	−0.093
Total time in bed	−0.044	0.014	−0.008	0.956	−0.006	−0.089
5. Weekend Oversleep
Social Jetlag of duration	0	0.008	−0.122	−0.008	0.974	0.002
Social Jetlag of total time in bed	0	0.009	−0.124	−0.013	0.972	0
Social Jetlag of total sleep time	0.017	0.013	−0.123	−0.005	0.954	0.006
6. Continuity
Sleep maintenance	0.001	0.007	0.004	0.111	0.003	1.008
Sleep efficiency	−0.019	−0.007	−0.01	0.109	−0.01	0.966
Wake after sleep onset (WASO)	−0.02	0.007	−0.002	0.309	−0.007	−0.864

Variables with meaningful loading (>|0.40|) in bold. Abbreviations: SD, standard deviation.

We labeled the first factor Irregularity; it included the standard deviations (SD) of sleep timing (in-bed time, get-up time, sleep onset time and offset time) and duration (total time in bed, total sleep time, sleep duration). Higher scores indicated greater day-to-day variability. We labeled the second factor Timing; it included the mean timing of sleep onset, offset, in-bed time and get-up time. We labeled the third factor Social Jetlag; it included social jetlag of sleep onset time, sleep offset time, in-bed time, and get-up time. Higher scores indicate later weekday timing relative to weekends. We labeled the fourth factor Duration; it included mean total sleep time, total time in bed and sleep duration. We labeled the fifth factor Weekend Oversleep; it included weekday-weekend differences in total sleep time, time in bed, and sleep duration. Social jetlag of get-up time and sleep offset time also showed cross-loadings with this factor, indicating overlapping constructs. Higher scores on this factor indicate greater weekday-weekend differences in sleep duration, suggesting accumulation of sleep debt during the week. We labeled the sixth factor Continuity; it included sleep maintenance, sleep efficiency and wake after sleep onset.

Confirmatory Factor Analysis

The full 6-factor CFA showed moderate to strong variable loadings (magnitude >|0.41|) and modest RMSEA (0.29) consistent with prior actigraphy factor analyses¹⁹. Although the CFI (0.639) and TLI (0.579) indices were lower than ideal, we did not consider it a major sign of model mis-fit in the broader context of the CFA fit indices and determined that the results were adequate to document the overall 6-factor structure (Table 3). Correlations among factors were weak to moderate (median |r|=0.09, range=-0.35–0.28, eTable 5), confirming that the identified dimensions represent relatively distinct aspects of adolescent sleep health.

Table 3.

Factor loadings from CFA.

	Loadings
Factor 1: Irregularity
SD sleep duration	0.994
SD total time in bed	0.995
SD get-up time	0.644
SD sleep offset time	0.646
SD in-bed time	0.636
SD total sleep time	0.966
SD sleep onset time	0.632
Factor 2: Timing
Get-up time	0.998
Sleep offset time	0.996
In-bed time	0.696
Sleep onset time	0.699
Factor 3: Social Jetlag
Social jetlag of in-bed time	0.999
Social jetlag of sleep onset time	0.992
Social jetlag of get-up time	0.417
Social jetlag of sleep offset time	0.414
Factor 4: Duration
Total sleep time	0.960
Sleep duration	0.994
Total time in bed	0.995
Factor 5: Weekend Oversleep
Social jetlag of sleep duration	0.994
Social jetlag of total time in bed	0.995
Social jetlag of total sleep time	0.974
Factor 6: Continuity
Sleep maintenance	0.994
Sleep efficiency	0.950
Wake after sleep onset (WASO)	0.883
Chi-square test statistic (df)	34,811 (237)
Chi-square p-value	<0.001
Comparative Fit Index	0.639
Tucker-Lewis Index	0.579
Root Mean Square Error of Approximation (90% CI)	0.293 (0.291–0.296)
Standardized Root Mean Square Residual	0.178

Abbreviations: CI, confidence interval; df, degrees of freedom; SD, standard deviation.

Cross-sectional associations between sleep health factors and psychopathology symptoms

Greater sleep Irregularity was concurrently associated with greater severity of almost all assessed outcomes, with the exception of somatic problems and thought problems (β’s 0.06 [95%CI 0.02,0.10]-0.12 [95%CI 0.08,0.16], Figure 1, eTable 6). Shorter Duration was associated with higher BPM total scores as well as attention, internalizing and externalizing problem subscale scores (β’s −0.06 [95%CI −0.10,−0.01] - −0.11 [95%CI −0.15,−0.06]).

Figure 1. Cross-sectional associations between sleep health dimensions and transdiagnostic mental health symptoms.

Figure shows standardized β coefficients for associations between sleep health dimensions (column/panel) and mental health symptoms (row). Shape (circle or triangle) represents point estimate; bars represent 95% confidence intervals. Black coefficients indicate significant associations (uncorrected p<0.05); black circles indicate significant associations after multiple comparison correction (q<0.05); black triangles indicate associations are no longer significant after multiple comparison correction (q≥0.05). All models adjusted for sex, ethnicity, race, maternal education, pubertal status, number of Fitbit tracking days, ratio of weekday to weekend tracking days, and whether any of the Fitbit tracking occurred during a school break. Abbreviations: BPM, Brief Problem Monitor; CBCL, Child Behavior Checklist; PGBI, Parent General Behavior Inventory; PQBC, Prodromal Questionnaire – Brief (Child version).

Prospective associations between sleep health factors and psychopathology symptoms

Shorter sleep Duration was associated with greater following-year psychotic-like symptoms (β=-0.07 [95%CI −0.11,−0.03], Figure 2) and BPM total and attention problems (β’s −0.06 [95%CI −0.09,−0.02]- −0.07 [95%CI −0.1,−0.03]).

Figure 2. Prospective associations between sleep health dimensions and transdiagnostic mental health symptoms one year later.

Figure shows standardized β coefficients for associations between sleep health dimensions (column/panel) and mental health symptoms (row). Shape (circle or triangle) represents point estimate; bars represent 95% confidence intervals. Black coefficients indicate significant associations (uncorrected p<0.05); black circles indicate significant associations after multiple comparison correction (q<0.05); black triangles indicate associations are no longer significant after multiple comparison correction (q≥0.05). All models adjusted for concurrent mental health symptoms, sex, ethnicity, race, socioeconomic status, pubertal status, number of Fitbit tracking days, ratio of weekday to weekend tracking days, and whether any of the Fitbit tracking occurred during a school break. Abbreviations: BPM, Brief Problem Monitor; CBCL, Child Behavior Checklist; PGBI, Parent General Behavior Inventory; PQBC, Prodromal Questionnaire – Brief (Child version).

Supplementary Analyses

Results were similar when we i) used all available outcome data (eTables 8–9), ii) substituted factor scores with the highest-loading single variable from each factor (eTables 10–11), iii) modeled each sleep dimension in separate univariate models (eTables 12–13) and iv) included quadratic effects (eTables 14–15).

Discussion

In this large, population-based cohort of early adolescents, we empirically derived six sleep health dimensions from wearable Fitbit devices: Irregularity, Timing, Social Jetlag, Duration, Weekend Oversleep, and Continuity. This multidimensional structure broadly aligns with the Ru-SATED and Peds B-SATED conceptual frameworks of sleep health¹⁴ and extends prior work in adults^18,19,20. Crucially, our findings adapt these frameworks for early adolescence by empirically distinguishing three often conflated forms of sleep variability: Irregularity (night-to-night variability in timing and duration), Social Jetlag (weekday–weekend shifts in timing), and Weekend Oversleep (weekday–weekend shifts in duration). Of the six dimensions we identified, greater Irregularity and shorter sleep Duration emerged as important correlates of concurrent mental health symptoms, while Duration was associated with prospective mental health symptoms.

Early adolescent sleep variability was represented by three relatively independent dimensions – Irregularity, Social Jetlag and Weekend Oversleep. Social Jetlag (weekday–weekend timing shifts) and Weekend Oversleep (duration shifts) have been suggested to reflect separate circadian and homeostatic mechanisms⁵², and our findings support their further distinction from day-to-day Irregularity⁵³. The presence of two dimensions reflecting weekday-weekend shifts in sleep behavior – never before identified in adults – highlights the importance of establishing developmentally-specific sleep health frameworks. Further, although weekday–weekend differences in timing peak in late adolescence¹⁰, our data show that these dimensions are already observable in early adolescence (11–13 years). Characterizing the trajectories of these dimensions into late adolescence and adulthood could reveal developmental shifts in the role of social jetlag and weekend oversleep on sleep health.

Among all sleep health dimensions, Irregularity was most consistently associated with concurrent mental health symptoms. These findings extend adult research linking day-to-day sleep irregularity to psychiatric risk^54–56, suggesting it may be a more developmentally salient index of adolescent mental health than traditional sleep measures such as duration or timing^50,57–59. Consistent with previous work^60,61, these findings also suggest that day-to-day sleep irregularity, distinct from social jetlag and weekend oversleep, is uniquely associated with mental health symptoms. The largest effect sizes were relatively small (≤0.12); indicating a 1-hour increase in the standard deviation in sleep midpoint timing corresponds to 0.71 points higher in, for example, attention problems. However, sleep regularity is highly modifiable^62,63, suggesting sleep interventions promoting consistent sleep patterns may represent a feasible, low-cost approach to promote adolescent mental health.

In prospective analyses, shorter duration was associated with higher following-year psychotic-like and attention symptoms, again with a relatively small effect size. Although prior ABCD studies have reported prospective associations between shorter parent-reported sleep duration and internalizing and externalizing symptoms^64–66, parent-reported and Fitbit-derived sleep duration show poor agreement within the ABCD sample⁶⁷. More broadly, subjective and behavioral sleep measures often show inconsistent associations with mental health outcomes, with self-reported measures typically demonstrating stronger associations^30,68,69. It is also plausible that sleep dimensions may act synergistically, rather than independently, to indicate mental health risk^70,71. Combining dimensions into sleep scores or phenotypes^72,73, or examining longitudinal changes, may reveal stronger relationships. For example, recent work using ABCD Fitbit data showed that longitudinal delays in sleep timing parallel increases in impulsive behavior⁷⁴, highlighting the value of dynamic modeling.

Analyses using factor scores were supported by supplemental models using single high-loading indicators for each factor, providing a streamlined set of variables for clinical practice and future sleep health research. Given the modifiability of sleep and accessibility of wearables, these dimensions could inform efforts to enhance mental health screening in clinical, school, or digital health settings, and represent potential targets for future behavioral interventions³². Future work could integrate these variables into predictive models⁷⁵, composite sleep health scores^20,76, or data-driven sleep phenotypes^72,77,78 to advance reproducibility and harmonization efforts in adolescent health research.

This study’s strengths include accessible, biobehavioral sleep measures, addressing limitations in previous studies that have relied on self- or parent-report measures. We leveraged up to three weeks of high-density Fitbit data in a diverse, population-based cohort, enhancing generalizability, and employed data-driven methods to identify sleep health dimensions associated with a broad spectrum of transdiagnostic mental health symptoms. Our work was guided by a strong theoretical framework; future work could apply other unsupervised machine learning methods (e.g., principal component analysis, independent component analysis), providing complementary insights into the dimensionality of adolescent wearable sleep data.

Several limitations warrant consideration. Although Fitbit provides scalable, ecologically valid data, its sensitivity to wake-sleep transitions in youth is limited^33,79. We addressed this through rigorous preprocessing, but some misclassification may remain. Future work should validate these findings using research-grade actigraphy devices. We focused on transdiagnostic self- and parent-report mental health outcomes using validated scales, which should be confirmed with clinician-rated measures. We were not able to validate our findings in an independent cohort; future work should seek to externally validate our findings. Effect sizes were small, but typical of ABCD studies⁸⁰. We did not examine the dimensionality of subjective sleep health, which may differ from the dimensionality of Fitbit sleep and can measure additional sleep health dimensions like Alertness or Satisfaction. Our analytic sample showed sociodemographic and symptomatic differences from the full ABCD sample, which may limit generalizability. Future work should also consider differences by sex, age, and pubertal status; it is possible that associations and effect sizes within key these subgroups may differ than those we observed in the full sample.

This study is the first to empirically establish and validate a multidimensional sleep health structure in early adolescence using wearable data. We identified six dimensions of sleep health – Irregularity, Timing, Duration, Continuity, Weekend Oversleep and Social Jetlag – thereby empirically differentiating weekday–weekend variation from daily inconsistency in sleep behavior. Greater Irregularity emerged as a correlate of transdiagnostic mental health symptoms, whereas shorter Duration was associated with prospective mental health outcomes, indicating their potential as developmentally sensitive, modifiable risk markers of mental health in young adolescents.

Key Points.

Question: What is the dimensionality of Fitbit-measured sleep health in young adolescents, and which dimensions are associated with concurrent and future mental health?

Findings: Using Fitbit data in a longitudinal cohort of 3,393 young adolescents, we identified 6 sleep health dimensions: Irregularity, Timing, Duration, Social Jetlag, Weekend Oversleep, and Continuity. Greater Irregularity was associated with higher concurrent transdiagnostic mental health symptoms, whereas shorter Duration was associated with higher symptoms at 1-year follow-up.

Meaning: These findings demonstrate the multidimensional nature of adolescent sleep health, guide selection of sleep health domains and representative measures to enhance reproducibility, and suggest potential intervention targets.