Neighborhood Disadvantage and Sleep Phenotypes In Adolescence: A Multi-Stage Analytical Framework Integrating LCA and Machine Learning

Background and Purpose:

Adequate sleep is vital for the physiological and psychological well-being of youth, serving as a primary mechanism through which environmental conditions influence developmental outcomes. Sleep health is a multidimensional construct deeply embedded in socioecological contexts. While traditional research often relies on single indicators or subjective reports, this study addresses existing gaps by identifying objective sleep phenotypes and examining their structural determinants. Using a data-driven approach, this study aimed to: (1) identify distinct sleep pattern subgroups among adolescents using actigraphy-derived indicators, and (2) determine the relative predictive importance of neighborhood-level versus individual-level factors in defining these sleep phenotypes.

Methods:

Data were drawn from the Adolescent Brain Cognitive Development Study (N=2,059), focusing on the 4-year follow-up when participants reached mid-adolescence (ages 13–14). In Stage 1, Latent Class Analysis (LCA) was conducted using eight Fitbit-derived objective indicators, including total sleep time, efficiency, fragmentation, and sleep stage proportions (Deep, REM, Light). In Stage 2, tree-based machine learning algorithms (Decision Tree, Random Forest, XGBoost, and CatBoost) were benchmarked to predict membership in a ‘Sleep Issue Group’. Candidate predictors included 33 geocoded neighborhood features (e.g., area deprivation, nighttime noise, social composition) and 10 individual-level factors (e.g., race, household income, mental health). The Boruta algorithm was used for feature selection, and the final CatBoost model was optimized for Recall to prioritize the identification of at-risk youth. LCA was performed using Stata 18.0, while the machine learning pipeline was implemented in Python via Jupyter Notebook.

Results:

LCA identified four distinct sleep phenotypes: (1) Non-restorative sleepers (5.4%), (2) Effective short sleepers (27.3%), (3) Typical sleepers (59.4%), and (4) Fragmented long sleepers (8.0%) . The predictive modeling revealed a striking finding: neighborhood-level predictors exerted a more dominant influence on sleep vulnerability than individual-level demographics or mental health indicators. Specifically, the final model (Recall=0.967) identified nighttime noise, area deprivation, foreign-born density, and digital exclusion (lack of computer access) as the top predictors. SHAP analysis further revealed powerful compounding effects, where the interaction between area deprivation and nighttime noise emerged as the single most influential predictor of sleep disturbances.

Conclusions and Implications:

The findings provide empirical evidence that adolescent sleep health is a structural outcome of neighborhood equity rather than a purely individual behavioral issue. By identifying specific environmental stressors, particularly the synergistic impact of noise and socioeconomic deprivation, this study offers a roadmap for targeted community-level interventions. For social work, these results advocate for a shift from individual-centered 'sleep hygiene' education toward macro-level policies and practice that address environmental injustices and improve physical infrastructure in marginalized neighborhoods. Stable sleep should be redefined as a fundamental prerequisite for developmental equity, requiring integrated housing and public health strategies.