Daily activity patterns from sleep and wake data in two large groups

The first step in the analysis (procedure shown in Fig. ), principal component analyses (PCA) for data reduction of the 36 metrics, yielded the eight principal components (described above, Tables –). K-means cluster analysis on these principal components found nine clusters to be the optimal representation of circadian rhythm metrics in participants in both cohort studies. This decision was based on examination of statistical criteria (Supplementary Figs. and) and interpretability of clusters. S3 S4 S6 S4 S5

Clusters were the same in both cohort studies, unless otherwise specified. Participants in cluster 1 (RAR++/PA++) had the most robust (++) RAR—characterised by highest amplitude, less fragmented and more stable rhythm —and active (++) pattern both in terms of light-intensity PA (LIPA) and moderate-to-vigorous PA (MVPA) while those in cluster 9 (RAR−−/PA−−) had the poorest (−−) RAR and least (−−) active pattern. In UKB, cluster 9 (RAR−−/PA−−/Chronotype−) was also characterised by delayed chronotype. Participants in clusters 2 and 3 had robust (+) RAR but opposite sleep profiles with cluster 2 characterised by shorter sleep duration (in UKB), more inneficient and fragmented sleep (Sleep−) and cluster 3 by longer sleep duration, more efficient, and less fragmented sleep (Sleep+). In addition, in WII, cluster 2 (RAR+/LIPA+/Sleep−) was also characterised by more LIPA and in UKB by both more LIPA and MVPA (RAR+/PA+/Sleep−). In both WII and UKB, cluster 3 (RAR+/LIPA+/Sleep+) participants had daytime activity characterised by light-intensity PA (LIPA+). Participants in cluster 4 (MVPA++) had a RAR pattern close to the mean, but with higher relative amplitude that was characterised by higher values in metrics related to high-intensity PA, including MVPA features and intensity gradient (IG) slope.

Participants in clusters 5, 6, 7, and 8 had poor (−) RAR pattern. Cluster 5 (RAR−/Chronotype−−) had the most delayed (−−) chronotype with later sleep period and later activity during the day. Participants in clusters 6 (RAR−/PA−/Sleep+) and 7 (RAR−/PA−/Sleep−−) had both low daytime activity but differed in sleep. Participants in cluster 8 (RAR−/PA+/Restless sleep) showed restless (agitated) sleep where mean acceleration during sleep was the highest (>1 SD and 2 SD from the sample mean in WII and UKB, respectively). In WII, participants in cluster 8 (RAR−/PA+/Restless sleep) also presented shorter and less efficient sleep. In this cluster, relative amplitude was low and cosinor mesor high, resulting from both more daytime PA and more nighttime movements (measured by mean acceleration) compared to participants in other clusters.

Clusters 3 (RAR+/LIPA+/Sleep+) and 6 (RAR−/PA−/Sleep+) in both cohort studies were the largest (15.7/17.1% (WII/UKB) and 17.7/16.1%, respectively) and clusters 8 (RAR−/PA+/Restless sleep) and 9 (RAR−−/PA−−/(Chronotype−, in UKB)) the smallest (5.9/3.4% and 6.5/6.9%, respectively). The size of clusters was similar in WII and UKB, with the largest difference observed for cluster 4 (MVPA++) (9.5 and 13.0%). Data on subjective chronotype preference in UKB (Table ) showed 8.2% of participants to report themselves as being 'Definitely an evening person' and this was particularly the case for participants in clusters identified with delayed chronotype using accelerometer data—clusters 5 (RAR−/Chronotype−−) and 9 (RAR−−/PA−−/Chronotype−) (21.9% and 16.1%, respectively, compared to 6.0% in cluster 3 (RAR+/LIPA+/Sleep+), p < 0.001). S9

Among behavioural factors, we found current smokers to be more likely to be in clusters 2 (RAR+/LIPA+/Sleep−), 5 (RAR−/Chronotype−−), 7 (RAR−/PA−/Sleep−−), and 9 (RAR−−/PA−−(/Chronotype− in UKB)); this was also the case for cluster 6 (RAR−/PA−/Sleep+) in WII and 8 (RAR−/PA+/Restless sleep) in UKB. Individuals who consumed less fruit and vegetables were more likely to be in clusters 7 (RAR−/PA−/Sleep−−) and 9 (RAR−−/PA−−(/Chronotype− in UKB)), and in UKB also in clusters 5 (RAR−/Chronotype−−) and 6 (RAR−/PA−/Sleep+). Exposure to light was higher in cluster 1 (RAR++/PA++) that had the most robust RAR and highest PA and lower in clusters characterised by poorer RAR and less PA (clusters 5, 6, 7, and 9); all p < 0.001.

Participants with higher BMI were less likely to be in clusters with higher PA (1 and 4), although the association was not statistically significant in WII (p = 0.29), and more likely to be in clusters with poorer RAR (clusters 5, 6, 7, 8, and 9), all p < 0.001 except for cluster 8 (RAR−/PA+/Restless sleep) in WII. Participants with diabetes were less likely to be in clusters characterised by MVPA (cluster 1 in UKB and cluster 4 in both studies), and more likely to be in the less active clusters in UKB (6, 7, and 9), while associations were in the same direction in WII although not statistically significant (p = 0.08, p = 0.31, p = 0.23, respectively). In UKB, participants with diabetes were also more likely to be in clusters 5 (RAR−/Chronotype−−) and 8 (RAR−/PA+/Restless sleep). No strong associations were found for hypertension and hyperlipidaemia, besides higher odds (1.26 (95% CI = 1.15–1.38)) of hypertension among participants in cluster 9 (RAR−−/PA−−) compared to cluster 3 (RAR+/LIPA+/Sleep+) in UKB. Participants with at least one chronic disease were more likely to be in cluster 5 (RAR−/Chronotype−−) and cluster 9 (RAR−−/PA−−(/Chronotype− in UKB)) in both studies, and in UKB also in clusters 6 (RAR−/PA−/Sleep+) and 7 (RAR−/PA−/Sleep−−). Use of CNS medication was less likely among participants in cluster 1 (RAR++/PA++) and more likely in clusters with poorer RAR (5 to 9).

The WII study is an ongoing prospective cohort study established in 1985–1988 on 10,308 British civil servants, consisting of a clinical examination every 4 to 5 years. An accelerometer measure was added to the 2012–2013 wave of data collection for 4880 men and women (aged 60–83 years) seen at the London clinic and those living in the South-Eastern regions of England who underwent clinical examination at home²³. Written informed consent for participation was obtained at each contact. Research ethics approval was obtained from the University College London ethics committee at each wave (latest reference number 85/0938).

The UKB is a prospective population-based cohort study on over 500,000 men and women aged between 40 and 69 years, registered with the UK National Health Service. Baseline measurements took place between 2006 and 2010. The accelerometer sub-study was undertaken on 236,488 individuals between 2013–2015³⁷. Participants aged ≥60 at accelerometer sub-study were retained in the present analysis as our focus was circadian rhythm clusters at older ages. Approval was received from the National Information Governance Board for Health and Social Care and the National Health Service North West Centre for Research Ethics Committee (reference number 11/NW/0382). All participants gave written consent for participation. Access to UKB data in our study was under application number 96856.

In WII, individuals wore a GENEActiv (Activinsights Ltd, Kimbolton, UK) triaxial accelerometer on their non-dominant wrist and completed a sleep diary over 24 h, for nine consecutive days. Accelerometer data were sampled at 85.7 Hz with the acceleration expressed relative to gravity (1 g = 9.81 m/s²), and processed using GGIR R package version 3.1–7³⁸. Raw acceleration was calculated with the metric Euclidean Norm Minus One (ENMO), negative values rounded to zero. Then these values were corrected for calibration error and non-wear time³⁹. Data from days 2 to 9 (penultimate day) were retained in the analysis, comprising a total of seven consecutive full day windows (daytime waking period and following sleep period). Sleep onset and waking up times were detected using an algorithm, guided by a sleep diary³⁹. The daytime waking periods were defined as the time between waking up to start the day and sleep onset at night, and sleep periods between sleep onset and waking up to start the day. Participants were included in the analyses if they had data from at least five valid full day windows (defined as wear times ≥2/3 of both waking and following sleep periods)⁴⁰.

In UKB, participants wore an Axivity AX3 triaxial accelerometer on their dominant wrist over 24 h for seven consecutive days, starting at 10 h on day 1³⁷. Data were sampled at 100 Hz. As a sleep diary was not used, sleep periods were detected using an algorithm that has been described previously³⁹. To maximise measures covering full day windows (sleep period and following waking period), data from the sleep onset of the first night until sleep onset on the last night were retained in the analyses, comprising a total of six consecutive full day windows. Participants were included in our analyses if they had data on at least five valid full day windows⁴⁰.

A total of 36 accelerometer-assessed metrics (Supplementary Table ) were derived to measure four dimensions of circadian rhythm—RAR, daytime activity, sleep, and chronotype. S1

RAR (6 metrics) was based on the day windows and included parametric measures based on the cosinor function fitted to the log of the acceleration signal⁴¹, that yielded three metrics: mesor (average activity according to the function), amplitude (the peak of the function minus the mesor), and R² (goodness of fit)⁴¹. The three non-parametric metrics were: interdaily stability, intradaily variability⁵, and the relative amplitude based on the mean acceleration of M₁₀ and the least active 5 h period (L₅)⁴².

Daytime activity (15 metrics) was based on activity levels during waking periods and classified as SB, LIPA, and MVPA⁴³. For each activity level, we calculated the daytime total duration, number and mean duration of bouts. We also considered: mean acceleration during waking period and during \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{M}}}}}_{10}$$\end{document}M10, transition probabilities (TPs) to switch from activity levels (LIPA or MVPA) to rest (SB), and from rest to activity, during the waking period (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{TP}}}}}_{{{{\rm{ar}}}},{{{\rm{d}}}}}$$\end{document}TPar,d and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{TP}}}}}_{{{{\rm{ra}}}},{{{\rm{d}}}}}$$\end{document}TPra,d)⁵, and two parameters derived from the acceleration distribution (intensity gradient (IG) intercept, and slope)⁴⁴.

Sleep (10 metrics) included: total sleep duration, sleep efficiency during the sleep period, mean acceleration during sleep and during L₅, number and mean duration of sleep bouts, TPs from sleep to wake, or wake to sleep during the night (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{TP}}}}}_{{{{\rm{sw}}}},{{{\rm{n}}}}}$$\end{document}TPsw,n and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{TP}}}}}_{{{{\rm{ws}}}},{{{\rm{n}}}}}$$\end{document}TPws,n), duration of wake after sleep onset (WASO), and mean duration of wake bouts.

Chronotype (5 metrics) was characterised by: time of sleep onset and waking, starting time of M₁₀ and L₅, and the cosinor acrotime (timing of the peak of the function).

All metrics were derived using GGIR R package version 3.1–7.

In WII, covariates were measured at the 2012–2013 wave either using questionnaire, clinical examination, or data from electronic health records, including the Hospital Episode Statistics (HES) and Mental Health Services Data Set. In UKB, covariates were either extracted from the baseline examination in 2006–2010 or extracted from electronic health records (HES).

Socio-demographic factors included age at accelerometer data collection, sex (defined based on the British civil servants record in WII, and from the UK National Health Service, updated in some cases by self-report of the participant, in UK Biobank), ethnicity (white or non-white), education (lower secondary school or less, secondary school, or higher than secondary school), cohabitation status (married/cohabiting or not married/cohabiting), whether in employment (yes or no), and area deprivation index (Index of multiple deprivation in WII⁴⁵ and Townsend deprivation index in UKB⁴⁶).

Behavioural factors included smoking status (never, ex-, or current-smoker), alcohol consumption (0, 1–14, or >14 units/week), fruit and vegetable consumption (less than or ≥ twice daily), and exposure to daylight, estimated as the proportion of waking time exposed to >1000 lux based on the light sensor from the GENEActiv accelerometer in WII⁴⁷ and self-report of the time spent outdoors in UKB³⁴. The light measures were standardised to account for season of accelerometer measurement. In additional analyses, we used self-reported chronotype preference in UKB, categorised as: Definitely a morning person, More a morning than evening person, More an evening than a morning person, and Definitely an evening person. Data on shift work were only available in UKB and were used in additional analyses, categorised as: No and Yes.

Health-related factors included BMI (in kg/m², calculated from measures of height and weight at the clinical examination), diabetes (measured using fasting glucose at the clinical examination ≥7.0 mmol/L, self-reported doctor-diagnosed diabetes, use of anti-diabetic medications, or record in HES), hypertension (defined as systolic/diastolic blood pressure ≥140/90 mmHg, use of antihypertensive drugs, or record in HES), hyperlipidaemia (defined as low-density lipoproteins >4.1 mmol/L or use of lipid-lowering drugs), chronic disease (yes, no out of coronary heart disease, stroke, Parkinson's disease, chronic obstructive pulmonary disease, heart failure, depression, other mental disorders, cancer, liver disease, and arthritis, and extracted from multiple sources including HES records, mental health records, cancer registry, and data collected at the Whitehall clinical examinations in WII, and from HES, clinical measures, and medication data in UKB), use of CNS medication (among anti-depressant, antipsychotic, hypnotic, anxiolytic, or Parkinson medications), reported by participant in the section on medication in the questionnaire.

Mortality data were used in preliminary analysis. Mortality cases were ascertained using linkage to the National Health Service (NHS) mortality register in both cohort studies. Participants were followed from date of accelerometer wear to November 2023 in WII and November 2022 in UKB.

To justify the use of 36 rather than the 10 commonly used metrics (interdaily stability, intradaily variability, relative amplitude, cosinor mesor, cosinor amplitude, cosinor acrotime, M₁₀ and L₅ timing and mean acceleration^6,8,48–50) to characterise circadian rhythm we compared the predictive performance of these two sets of metrics for risk of mortality. We first used principal component analysis (PCA) to extract components to reflect the metrics too strongly correlated to be used in the prediction analysis. The number of components for each set of metrics (10 and 36) was selected using the eigenvalue criterion (≥1) and the cumulated percentage of variance explained (at least 75%). Cox regression analysis was used to examine the association of these components with mortality, and their predictive performances were compared using the C-index.

As the predictive performance of components using 36 metrics was better, subsequent analyses were conducted using the set of 36 metrics in parallel in both cohort studies to identify how circadian rhythm metrics clustered in individuals, and then examine the association of the clusters with covariates. The correlation between the 36 metrics (reflecting RAR, daytime activity, sleep, and chronotype) was examined using Pearson's correlation coefficients. Then, to examine how the 36 metrics were distributed across individuals we undertook cluster analyses; the workflow for steps in these analyses is shown in Supplementary Fig. S3. The procedure to identify clusters in both WII and UKB was made in three steps. First, standardisation of the 36 metrics using scale function to make them comparable. Second, we applied the data reduction method PCA on this standardised metrics using prcomp function. Using eigenvalue ≥ 1 and cumulative variance explained ≥75% of the variance of the 36 metrics, we selected N component. Finally, we used k-means clustering with the kmeans function on the retained components, testing for K = 4 to 12, with 5000 numbers of random starts, and 10,000 maximum number of iterations. To determine the number of cluster K retained, we examined statistical criteria—within cluster sum of squares, Silhouette coefficient, and Davies-Bouldin index—to assess the statistical differences between clustering solutions. As the differences were small, the interpretability of the clusters was further examined. For every K (number of clusters) solution, we looked at the visualisation of the solutions to interpret each cluster. The aim was to determine whether the addition of a new cluster provided new relevant information in terms of differences in the circadian rhythm profiles. We also considered cluster sizes, ensuring that each contained at least N ≥ 200 participants to be suitable for analysis.

In WII, the assumptions for post-hoc Tukey tests were not met and to allow interpretation of the clusters as a function of each metric we used the non-parametric simultaneous rank test procedure. In UKB, the large sample size did not allow non-parametric simultaneous rank test to be used, leading us to use the post-hoc Tukey test with ANOVA⁵¹.

After identification of circadian rhythm clusters, we used multinomial logistic regression to examine associations of participants' characteristics and season of wear (independent variables) with circadian rhythm clusters (dependent variable). The reference category in the multinomial regression was chosen based on size (number of participants), the cluster where none of the metric means was above or below one SD from the population mean, and better health status.

Further information on research design is available in the linked to this article. Nature Portfolio Reporting Summary