Exposome factors predict the association between sleep health and depression

Background: Sleep health and depression are multidimensional constructs, yet their shared determinants remain obscure. We aim to identify the key exposome factors in predicting sleep-related depression (SRD). Methods: We integrated regularized canonical correlation analysis with machine-learning models within a nested cross-validation to identify a shared phenotype (i.e., sleep-related depression (SRD)) and to predict SRD from exposome factors in the UK Biobank (n=64,781). The best-performing model was validated in an independent subsample at baseline and follow-up (n=8,139) and in a clinical depression subsample (n=3,046) to assess generalizability. Subsequently, we conducted complementary analyses to assess the exposome's role in other latent phenotypes of sleep and depression. Findings: We identified a robust multivariate association between sleep and depression in GP1 (canonical r = 0.42). Exposome factors moderately predicted SRD (r = 0.25; 95% CI [0.25, 0.26]; R2 = 0.06; 95% CI [0.06, 0.07]; rMSE = 1.11; 95% CI [1.10, 1.11]). The top predictors were less frequency of confiding in others, less vigorous physical activity, passive smoking exposure, and more sedentary television viewing. Out-of-sample validation of the predictive model showed similar patterns in GP2 at baseline, at follow-up, and in clinical depression subsamples. Similarly, less frequency of confiding in others and greater sedentary television viewing were the main predictors of other depression-related profiles, whereas less walking frequency and less time spent outdoors in winter predicted poor sleep-related profiles. Interpretation: Our generalizable predictive model identifies critical modifiable predictors of the association between sleep health and depression that could serve as potential targets for personalized interventions.