Aims/hypothesis: Suboptimal sleep health is linked to higher risks for incident type 2 diabetes. We aimed to assess the clinical utility of adding self-reported sleep traits to a type 2 diabetes prediction model.

Methods: In this cohort study, we used UK Biobank data and Cox proportional hazards models to examine how self-reported sleep duration and insomnia symptoms were associated with incident type 2 diabetes risk. Harrell's C statistic and net reclassification improvement (NRI) were used to assess whether sleep traits improved the incident type 2 diabetes discrimination and predictive utility achieved using QDiabetes variables, with and without including a type 2 diabetes polygenic risk score (PGS). Independent replication was explored in the Nurses' Health Study, the Nurses' Health Study II and the Health Professionals Follow-up Study.

Results: Extremes of sleep duration and occasional or frequent insomnia symptoms were associated with higher risks for incident type 2 diabetes. In the UK Biobank and replication cohorts, adding sleep traits to the QDiabetes risk score did not improve type 2 diabetes prediction (C statistic: QDiabetes alone 0.8933; QDiabetes + sleep duration 0.8939; QDiabetes + insomnia 0.8931; QDiabetes + sleep traits 0.8935). The corresponding total NRI values were: 0.08 (95% CI -0.18, 0.33), 0.04 (95% CI -0.08, 0.16) and 0.04 (95% CI -0.10, 0.18). Inclusion of PGS data marginally improved the type 2 diabetes risk prediction achieved using The QDiabetes calculator, with or without the inclusion of sleep traits in the model (QDiabetes + PGS: C statistic 0.8945; total NRI 0.20 [95% CI 0.12, 0.28]; QDiabetes + PGS + sleep traits: C statistic 0.8946; total NRI 0.18 [95% CI 0.09, 0.27]).

Conclusions/interpretation: While sleep duration and insomnia symptoms were associated with type 2 diabetes risk, they are not useful for improving type 2 diabetes prediction beyond QDiabetes model performance. Inclusion of a type 2 diabetes PGS marginally improved prediction but lacked clear clinical utility.