Subjective and objective sleep and circadian parameters as predictors of depression-related outcomes: A machine learning approach in UK Biobank

Sleep and circadian disruption are associated with depression onset and severity, but it is unclear which features (e.g., sleep duration, chronotype) are important and whether they can identify individuals showing poorer outcomes. Within a subset of the UK Biobank with actigraphy and mental health data (n = 64,353), penalised regression identified the most useful of 51 sleep/rest-activity predictors of depression-related outcomes; including case-control (Major Depression (MD) vs. controls; postnatal depression vs. controls) and within-case comparisons (severe vs. moderate MD; early vs. later onset, atypical vs. typical symptoms; comorbid anxiety; suicidality). Best models (of lasso, ridge, and elastic net) were selected based on Area Under the Curve (AUC). For MD vs. controls (n(MD) = 24,229; n(control) = 40,124), lasso AUC was 0.68, 95 % confidence interval (CI) 0.67–0.69. Discrimination was reasonable for atypical vs. typical symptoms (n(atypical) = 958; n(typical) = 18,722; ridge: AUC 0.74, 95 % CI 0.71–0.77) but poor for remaining models (AUCs 0.59–0.67). Key predictors across most models included: difficulty getting up, insomnia symptoms, snoring, actigraphy-measured daytime inactivity and lower morning activity (~8 am). In a distinct subset (n = 310,718), the number of these factors shown was associated with all depression outcomes. Analyses were cross-sectional and in middle-/older aged adults: comparison with longitudinal investigations and younger cohorts is necessary. Sleep and circadian measures alone provided poor to moderate discrimination of depression outcomes, but several characteristics were identified that may be clinically useful. Future work should assess these features alongside broader sociodemographic, lifestyle and genetic features.