Actigraphy and subjective sleep predictors of nine-year generalized anxiety disorder

Sleep disturbances have been linked to generalized anxiety disorder (GAD) symptoms. However, cross-sectional studies, linearity assumptions, and limited predictor sets preclude identifying which unique sleep disturbance markers precede GAD symptoms. We thus harnessed machine learning (ML) to determine objective and subjective sleep disturbance predictors of nine-year GAD symptoms.

Community adults (N = 1054) underwent baseline surveys, clinical interviews, and seven-day sleep actigraphy protocols. GAD symptoms were reassessed nine years later. Seven ML models were examined with 44 baseline predictors. Partial dependence and Shapley additive explanation plots were created as interpretable ML approaches with the best-performing random forest model using nested cross-validation. Sensitivity analyses included and excluded GAD sleep items.

The final multivariable predictive algorithm performed well (R² = 69.7 %, 95 % confidence interval [67.3 %-71.9 %]), thus explaining over half the variance in the outcome. These self-reported sleep disturbances predicted GAD symptoms in descending order of relative importance: sleep disturbances, poorer sleep quality, longer sleep onset latency, daytime dysfunction, habitual sleep inefficiency, and sleep medication use. These rest-phase actigraphy markers predicted nine-year GAD symptoms: higher maximum and total activity counts. Longer total sleep time during the sleep phase and higher average sleep bouts during the active phase also predicted nine-year GAD severity.

Outcomes highlight the importance of combining actigraphy and self-report sleep assessments. Future studies should determine the degree to which these patterns extend to the within-person level to develop early prevention, treatment, and precision mental health strategies for individuals at risk of, or with, increased GAD severity.