0309 Linking Social, Emotional, Structural, and Financial Determinants of Health in the US Population Associated with Sleep Health Using Machine-Learning

Abstract Introduction Previous research has studied various determinants of health in relation to sleep health, but few leverage machine learning for parsimonious modeling to identify the most impactful predictors and corresponding trends. Methods Data from the 2022 Behavioral Risk Factor Surveillance System were used (N=193,225). Health determinants included availability of emotional support, social isolation, food insecurity, food stamps, housing insecurity, inability to pay utilities, employment insecurity, lack of transportation, household income, education, and employment status. Sleep duration was examined, with age, sex, race/ethnicity, and body mass index as covariates. Model building used LASSO to develop the most parsimonious model, followed by a generalized ordered logit model using the variables included in the LASSO model. Results A LASSO regression with EBIC selection was conducted while forcing in age, sex, and race/ethnicity as partialed covariates. The EBIC-optimal solution used λ=345.32. Variables with nonzero lasso coefficients included employment, education, income, housing, transportation, employment, and food insecurity, food stamps, utilities, emotional support, social isolation, and stress. The general logit model demonstrated a strong overall fit (Log-likelihood=−289,027.24, p 0.001). Stress was consistently the strongest predictor (Rarely stressed: β≈0.89-1.79; Never stressed: β≈1.11-1.79), with higher stress consistently linked to lower sleep duration. Worse emotional support strongly predicted lower sleep duration, particularly at lower cutpoints (Rarely supported: β=–0.28 to –0.34; Never supported: β=–0.21 to –0.44). Isolation showed opposite directional effects at different cutpoints; at low cutpoints, isolation predicted higher sleep duration but at high cutpoints it predicted less sleep. Hardships were consistently negative across thresholds (utilities insecurity: β≈–0.13 to –0.20; transportation insecurity: β≈–0.05 to –0.17; housing insecurity and food stamps: β≈–0.04 to –0.06). Lower SES (lower income and education) predicted lower sleep duration in early thresholds and higher duration in upper thresholds, again requiring a generalized model. Non-Hispanic Whites had highest sleep duration; minority groups showed lower sleep across most thresholds. Conclusion Machine learning provides a way of identifying high impact targets for intervention. In addition to highlighting the importance of socioeconomic and structural factors for sleep health interventions, the model suggests that emotional factors such as stress may be the most impactful to address. Support (if any) R01MD011600, R01MH135978