Night-to-night sleep variability and subsequent cognitive performance: a UK Biobank cohort study

Abstract Background Dimensional measures of night-to-night sleep variability may provide information about cognitive performance that is not captured by the Sleep Regularity Index (SRI). We examined associations with subsequent cognition and conditional follow-up performance. Methods This exploratory UK Biobank study included 6,081 participants with wrist accelerometry and two subsequent cognitive assessments. Exposures were within-person standard deviations (SDs) of sleep onset, waking time, sleep duration and movement-defined interruption count, their equally weighted Composite Variability Score (CVS), and Sleep Regularity Index 1 (SRI1). Primary models used type 3 heteroscedasticity-consistent (HC3) inference and adjusted for estimated age at accelerometry, sex, within-country regional socioeconomic scores, country, body mass index, smoking, ever drinking, mean sleep duration and assessment timing; longitudinal models additionally adjusted for earlier same-domain cognition. Benjamini–Hochberg correction covered six exposures and two outcomes within each design. Results At the first subsequent assessment, all four component SDs and CVS were associated with lower fluid intelligence (FI); for CVS, β = −0.1444 points per exposure SD (95% confidence interval (CI) − 0.1970 to − 0.0918; q < 0.0001). In longitudinal models, waking-time SD, duration SD and CVS had similar small negative estimates (β = −0.0591, − 0.0483 and − 0.0506), but their 12-test q-values were 0.0610, 0.0704 and 0.0704. Longitudinal added information beyond SRI1 was small (increase in explained variance, ΔR² ≤ 0.00079) and did not meet the corrected threshold. Selection-weighted estimates were weaker. Conclusions Greater sleep variability was consistently associated with lower FI at the first subsequent assessment. Longitudinal estimates were small and directionally similar, but the evidence was sensitive to multiplicity correction and selection adjustment. These findings do not establish prediction of cognitive decline or clinical superiority over SRI.