Abstract
OBJECTIVE
Circadian rhythms play a key role in metabolic health. Rest–activity rhythms, which are in part driven by circadian rhythms, may be associated with diabetes risk. There is a need for large prospective studies to comprehensively examine different rest–activity metrics to determine their relative strength in predicting risk of incident type 2 diabetes.
RESEARCH DESIGN AND METHODS
In actigraphy data from 83,887 UK Biobank participants, we applied both parametric and nonparametric algorithms to derive 13 different metrics characterizing different aspects of rest–activity rhythm. Diabetes cases were identified using both self-reported data and health records. We used Cox proportional hazards models to assess associations between rest–activity parameters and type 2 diabetes risk and random forest models to determine the relative importance of these parameters in risk prediction.
RESULTS
We found that multiple rest–activity characteristics were predictive of a higher risk of incident diabetes, including lower pseudo-F statistic (hazard ratio [HR] of quintile 1 [Q1] vs. Q5 1.27; 95% CI 1.09–1.46; Ptrend < 0.001), lower amplitude (HRQ1 vs. Q5 2.56; 95% CI 2.21–2.97; Ptrend < 0.001), lower midline estimating statistic of rhythm (HRQ1 vs. Q5 2.59; 95% CI 2.24–3.00; Ptrend < 0.001), lower relative amplitude (HRQ1 vs. Q5 4.64; 95% CI 3.74–5.76; Ptrend < 0.001), lower M10 (HRQ1 vs. Q5 3.82; 95% CI 3.20–4.55; Ptrend < 0.001), higher L5 (HRQ5 vs. Q1 1.88; 95% CI 1.62–2.19; Ptrend < 0.001), and later L5 start time (HRQ5 vs. Q1 1.20; 95% CI 1.04–1.38; Ptrend = 0.004). Random forest models ranked most of the rest–activity metrics as top predictors of diabetes incidence, when compared with traditional diabetes risk factors. The findings were consistent across subgroups of age, sex, BMI, and shift work status.
CONCLUSIONS
Rest–activity rhythm characteristics measured from actigraphy data may serve as digital biomarkers for predicting type 2 diabetes risk.
Graphical Abstract
Introduction
Circadian rhythms play a pivotal role in regulating the 24-h cycle of many behavioral and physiologic processes and are essential to maintaining optimal health (1). The rest–activity rhythm, which encompasses the 24-h pattern of motor activities, sedentary behaviors, and sleep, is in part driven by circadian rhythms (2). Growing research has linked disrupted rest–activity rhythms to a wide range of mortality and morbidity outcomes, from incidence of cardiovascular disease to death from cancer (3). Because of the increasing accessibility to wearable devices for activity tracking, there is growing interest in investigating the potential of digital behavioral biomarkers, such as rest–activity parameters, for disease risk prediction and health monitoring.
One area in which rest–activity rhythm characteristics may be used to improve risk prediction is metabolic health, specifically type 2 diabetes. Global prevalence of type 2 diabetes has escalated from ∼3,800 per 100,000 individuals in 1990 to 6,000 in 2017 (4). This growing epidemic imposes a substantial economic burden on society (5), requiring effective prevention and management strategies. Previous studies have consistently reported that rest–activity rhythm characteristics, such as reduced amplitude, altered rhythmic phase, higher fragmentation, and lower overall rhythmicity, are associated with a higher risk of type 2 diabetes (6–8). However, evidence from large cohorts is still limited, warranting deeper investigation. Moreover, rest–activity rhythms can be characterized by both parametric and nonparametric algorithms, each offering a unique set of measures of various aspects of the rhythmic pattern. However, earlier studies focused on either parametric (6,7) or nonparametric (8) variables alone and were thus unable to provide a comprehensive assessment of the relationship between rest–activity rhythms and diabetes risk. Finally, no prior study has compared the predictive value of rest–activity parameters, either among the parameters themselves or with other conventional diabetes risk factors, which is a critical step toward developing risk models based on these potential behavioral biomarkers.
Therefore, the aims of the current study were to 1) investigate associations of actigraphy-derived parametric and nonparametric rest–activity rhythm parameters with risk of incident type 2 diabetes in UK Biobank participants and 2) evaluate the relative importance of different rest–activity rhythm parameters in predicting incident type 2 diabetes using machine learning.
Research Design and Methods
Study Population
Data were obtained from the UK Biobank, which enrolled >500,000 participants aged 40–69 years from 2006 to 2010 (9). The UK Biobank collected data on participants’ phenotypic and genotypic details, with ongoing follow-up for a wide range of health outcomes (9). Participants who had provided a valid e-mail address were invited to wear an accelerometer for 7 days between May 2013 and December 2015 (10,11). Devices were sent to 106,053 participants, and of these, 103,660 had sufficient recordings to derive acceleration data summarized at 5-s epochs (10). Among these participants, 15,183 were excluded because their recordings were determined to have one of eight quality issues (e.g., interrupted recording periods, data recording errors, daylight saving crossover, data calibration issues). We further excluded 2,622 individuals who had <3 valid days of data (valid day was defined as no missing epoch data after imputation and <4 h of imputed data) and 1,968 individuals with prevalent diabetes at the time of accelerometer data collection. The final analytic sample included 83,887 participants.
Rest–Activity Rhythm
Accelerometer data were collected using wrist-worn Axivity AX3 triaxial accelerometers and processed as detailed previously (10,11). Briefly, participants were asked to wear these devices on their dominant wrist for 7 continuous days. The accelerometer data were calibrated to local gravity to ensure different devices provided similar output under similar conditions. The devices captured triaxial acceleration data at 100 Hz with a dynamic range of ±8g. To derive the vector magnitude for each 5-s epoch, the Euclidean norm of the acceleration across all axes was calculated, followed by the removal of gravity and machine noise. Nonwear time (defined as consecutive stationary episodes lasting for at least 60 min where all three axes had SD <13.0 mg) was imputed using the average vector magnitude of the same time period from days with wear data (11). For the current analysis, we further consolidated the 5-s data to 1-min epochs to reduce noise.
Based on the 1-min epoch data, we derived 13 rest–activity rhythm parameters using both parametric and nonparametric methods. The parametric method was the extended cosine model, from which we derived six variables: 1) pseudo-F statistic, 2) amplitude, 3) midline estimating statistic of rhythm (mesor), 4) up mesor, 5) down mesor, and 6) acrophase. Furthermore, we derived seven nonparametric parameters, including 1) interdaily stability (IS), 2) intradaily variability (IV), 3) activity during the least active 5 h of the day (L5), 4) activity during the most active 10 h of the day (M10), 5) relative amplitude (RA), 6) L5 start time, and 7) M10 start time. Detailed descriptions of all rest–activity rhythm parameters are presented in Supplementary Table 1. All parameters were divided into quintiles, with the quintile presumed to have the lowest risk of diabetes chosen as the reference (i.e., quintile 1 [Q1] for IV, L5, and L5 start time; Q3 for up mesor, down mesor, acrophase, and M10 start time; Q5 for the other seven parameters). Because the distribution of the L5 midpoint included midnight, we expressed this variable as hours from midnight (e.g., 2300 h is expressed as −1 h, and 0200 h is expressed as 2 h) to avoid discontinuity in variable values.
Incident Diabetes
We applied published algorithms by Eastwood et al. (12) to the UK Biobank first occurrences data set (13) to define both prevalent and incident type 2 diabetes. Briefly, the UK Biobank first occurrences data set reported the date of the first recorded health-related outcomes across read code information in the primary care data, ICD-9 and ICD-10 codes in the hospital inpatient data, ICD-10 codes in death register records, and self-reported medical condition codes at the baseline or subsequent UK Biobank assessment center visit (13). Participants with any form of diabetes at baseline or recorded before the last day of accelerometer data collection were categorized as prevalent cases. Incident cases of type 2 diabetes were defined as non–insulin-dependent diabetes after the accelerometer data collection.
Covariates
Covariates included age, sex, ethnicity, household income, educational or professional qualifications, smoking, shift work, sleep duration, total physical activity as measured by overall acceleration average per day, and BMI. All covariates except for total physical activity were self-reported.
Statistical Analysis
To assess associations between rest–activity rhythm parameters and risk of incident diabetes, Cox proportional hazards models were used to derive hazard ratios (HRs) and 95% CIs. The follow-up time was calculated as the number of days between the end of accelerometer data collection and the earliest date of the following: 1) the first occurrence of type 2 diabetes, 2) death, or 3) the last date reported in the UK Biobank first occurrences data set (1 February 2023) at the time of our analysis. We presented results from the following models: 1) the minimal model (model 1), adjusted for age and sex, and 2) the main model (model 2), adjusted for confounders, including age, sex, ethnicity, average total household income before tax, educational or professional qualifications, smoking status, alcohol drinking status, and night-shift work status. We also conducted two sensitivity analyses by additionally adjusting for total physical activity and sleep duration (model 3) or BMI (model 4). The objective of model 3 was to assess the degree to which the associations were explained by specific behavioral components of the rest–activity rhythm (i.e., physical activity and sleep). The objectives of model 4 were to account for any confounding by BMI but also to provide insight into BMI as a potential mediator. To test for trend, quintiles of rest–activity rhythm parameters were modeled as continuous, and statistical significance was assessed by the Wald test. In the subgroup analysis, we stratified by age, sex, night-shift work status, and BMI, and the interaction between rest–activity rhythm parameters and the variables used to define subgroups was tested by the likelihood ratio test comparing a model with the cross-product term with one without.
The relative importance of rest–activity variables in predicting incident diabetes was assessed by a random forest model, a regression-based machine learning approach that averages across a large number (in our case, 500) of tree predictor models, each constructed using a bootstrap sample (14). The model simultaneously included all rest–activity rhythm parameters (continuous) and covariates in the main model. The importance of each variable in predicting incident diabetes was assessed by the difference in Gini impurity index comparing models with and without the variable. A larger decrease in Gini impurity index values when a certain predictor was left out suggested that the inclusion of the variable in the model led to a more effective classification of participants as incident diabetes cases or noncases. Variables with larger decreases were considered to have greater importance in making a more accurate classification. The random forest model was performed in the overall sample and in subgroups of age, sex, shift work status, and BMI. All analyses were performed in R (version 4.1.2). Parametric and nonparametric variables were generated using R packages ActCR and nparACT, respectively. The random forest model was performed using the R package randomForest.
Results
Baseline characteristics of the participants are listed in Table 1, grouped by quintiles of the pseudo-F statistic. Participants in the lower quintiles, representing less robust overall rhythmicity, were more likely to be male, report current smoking, and have a history of any shift work and night-shift work specifically, a higher BMI, and lower total physical activity.
Table 1.
Demographic characteristics of included participants by quintile of pseudo-F statistic
| Pseudo-F statistic* | |||||
|---|---|---|---|---|---|
| Q1 n = 16,277 | Q2 n = 16,277 | Q3 n = 16,278 | Q4 n = 16,276 | Q5 n = 16,277 | |
| Pseudo-F statistic | 18.9 (4.6, 32.2) | 68.3 (56.7, 79.7) | 114.4 (102.6, 127.0) | 171.1 (154.9, 189.7) | 280.7 (239.7, 353.2) |
| Age, years | 56.0 (49.0, 61.0) | 57.0 (50.0, 62.0) | 58.0 (51.0, 63.0) | 58.0 (51.0, 63.0) | 57.0 (49.0, 63.0) |
| Female sex | 8,275 (51) | 8,900 (55) | 9,391 (58) | 9,950 (61) | 10,105 (62) |
| Ethnicity | |||||
| White European | 15,746 (97) | 15,832 (98) | 15,765 (97) | 15,737 (97) | 15,707 (97) |
| South Asian | 98 (0.6) | 79 (0.5) | 113 (0.7) | 114 (0.7) | 139 (0.9) |
| African Caribbean | 125 (0.8) | 112 (0.7) | 111 (0.7) | 149 (0.9) | 139 (0.9) |
| Mixed or other | 238 (1.5) | 204 (1.3) | 232 (1.4) | 216 (1.3) | 250 (1.5) |
| Average total household income before tax, £ | |||||
| <18,000 | 2,085 (13) | 2,079 (13) | 2,116 (13) | 2,127 (13) | 2,112 (13) |
| 18,000–30,999 | 3,353 (21) | 3,514 (22) | 3,635 (22) | 3,543 (22) | 3,561 (22) |
| 31,000–51,999 | 4,270 (26) | 4,272 (26) | 4,169 (26) | 4,223 (26) | 4,138 (25) |
| 52,000–100,000 | 3,943 (24) | 3,718 (23) | 3,643 (22) | 3,590 (22) | 3,561 (22) |
| >100,000 | 1,096 (6.7) | 1,082 (6.6) | 1,016 (6.2) | 1,029 (6.3) | 1,141 (7.0) |
| Do not know/prefer not to answer | 1,530 (9.4) | 1,612 (9.9) | 1,699 (10) | 1,764 (11) | 1,764 (11) |
| Qualifications† | |||||
| College or university degree | 7,405 (45) | 7,010 (43) | 6,967 (43) | 7,001 (43) | 6,940 (43) |
| A levels/AS levels or equivalent | 2,160 (13) | 2,080 (13) | 2,142 (13) | 2,111 (13) | 2,182 (13) |
| O levels/GCSEs or equivalent | 3,087 (19) | 3,305 (20) | 3,406 (21) | 3,401 (21) | 3,339 (21) |
| CSEs or equivalent | 712 (4.4) | 668 (4.1) | 615 (3.8) | 581 (3.6) | 594 (3.6) |
| NVQ, HND, or HNC or equivalent | 894 (5.5) | 933 (5.7) | 815 (5.0) | 788 (4.8) | 791 (4.9) |
| Other professional qualification | 715 (4.4) | 801 (4.9) | 841 (5.2) | 859 (5.3) | 846 (5.2) |
| None of above/prefer not to answer | 1,304 (8.0) | 1,480 (9.1) | 1,492 (9.2) | 1,535 (9.4) | 1,585 (9.7) |
| Smoking status | |||||
| Never | 9,219 (57) | 9,269 (57) | 9,349 (58) | 9,495 (58) | 9,555 (59) |
| Previous | 5,622 (35) | 5,858 (36) | 5,853 (36) | 5,769 (36) | 5,686 (35) |
| Current | 1,385 (8.5) | 1,113 (6.9) | 1,027 (6.3) | 969 (6.0) | 995 (6.1) |
| Alcohol drinking status | |||||
| Never | 403 (2.5) | 429 (2.6) | 444 (2.7) | 461 (2.8) | 562 (3.5) |
| Previous | 427 (2.6) | 424 (2.6) | 407 (2.5) | 380 (2.3) | 470 (2.9) |
| Current | 15,431 (95) | 15,406 (95) | 15,410 (95) | 15,419 (95) | 15,231 (94) |
| BMI, kg/m2 | 26.2 (23.7, 29.1) | 26.1 (23.7, 29.0) | 25.9 (23.5, 28.8) | 25.8 (23.4, 28.6) | 25.5 (23.1, 28.3) |
| Shift work status | |||||
| Never/rarely | 14,585 (90) | 14,897 (92) | 15,075 (93) | 15,135 (93) | 15,215 (93) |
| Sometimes | 711 (4.4) | 629 (3.9) | 557 (3.4) | 573 (3.5) | 525 (3.2) |
| Usually | 219 (1.3) | 149 (0.9) | 136 (0.8) | 129 (0.8) | 141 (0.9) |
| Always | 762 (4.7) | 602 (3.7) | 510 (3.1) | 439 (2.7) | 396 (2.4) |
| Night shift work status | |||||
| Never/rarely | 15,430 (95) | 15,590 (96) | 15,703 (96) | 15,751 (97) | 15,801 (97) |
| Sometimes | 472 (2.9) | 391 (2.4) | 354 (2.2) | 320 (2.0) | 298 (1.8) |
| Usually | 142 (0.9) | 107 (0.7) | 75 (0.5) | 86 (0.5) | 66 (0.4) |
| Always | 233 (1.4) | 189 (1.2) | 146 (0.9) | 119 (0.7) | 112 (0.7) |
| Overall acceleration average, mg | 26.7 (22.1, 31.9) | 26.7 (22.1, 32.0) | 26.8 (22.4, 32.2) | 27.3 (22.9, 32.6) | 28.9 (24.1, 34.6) |
| Sleep duration, h | 7.0 (7.0, 8.0) | 7.0 (7.0, 8.0) | 7.0 (7.0, 8.0) | 7.0 (7.0, 8.0) | 7.0 (7.0, 8.0) |
Data are given as n (%) or median (interquartile range).
A level, Advanced Level; AS level, Advanced Subsidiary Level; CSE, Certificate of Secondary Education; GCSE, General Certificate of Secondary Education; HNC, Higher National Certificate; HND, Higher National Diploma; NVQ, National Vocational Qualification; O level, Ordinary Level.
*Higher pseudo-F statistic indicates stronger overall rhythmicity.
†CSE qualifications were secondary school qualification available between 1965 and 1987, typically taken at age 15 or 16 years, and were less rigorous than O-level secondary school qualifications, which were replaced in 1988 by GCSEs. A-level and AS-level qualifications are typically taken by students aged 16–18 years. NVQ and HND qualifications are vocational qualifications.
Table 2 reports the associations between rest–activity rhythm parameters and risk of incident diabetes. Of the 13 rest–activity parameters, seven showed a significant association: 1) lower pseudo-F statistic (model 2: HRQ1 vs. Q5 1.27; 95% CI 1.09–1.46; Ptrend < 0.001), 2) lower amplitude (HRQ1 vs. Q5 2.56; 95% CI 2.21–2.97; Ptrend < 0.001), 3) lower mesor (HRQ1 vs. Q5 2.59; 95% CI 2.24–3.00; Ptrend < 0.001), 4) lower RA (HRQ1 vs. Q5 4.64; 95% CI 3.74–5.76; Ptrend < 0.001), 5) lower M10 (HRQ1 vs. Q5 3.82; 95% CI 3.20–4.55; Ptrend < 0.001), 6) higher L5 (HRQ5 vs. Q1 1.88; 95% CI 1.62–2.19; Ptrend < 0.001), and 7) later L5 start time (HRQ5 vs. Q1 1.20; 95% CI 1.04–1.38; Ptrend = 0.004). No association was found for up mesor, down mesor, acrophase, IS, or M10 start time. Most associations were attenuated after additional adjustment for total physical activity and sleep duration in model 3 (Supplementary Table 2) or BMI in model 4 (Supplementary Table 3), but the associations for amplitude, mesor, RA, L5, and M10 remained statistically significant in both models. Results from stratified analyses by age, sex, status of any shift work, status of night-shift work, and BMI (Supplementary Tables 4–8) were generally similar across subgroups. Although we noticed that some interactions exhibited nominal statistical significance, these may have been chance findings resulting from multiple comparisons.
Table 2.
Association between quintiles of rest–activity rhythm parameters and risk of type 2 diabetes
| Model* | Q1 | Q2 | Q3 | Q4 | Q5 | P trend |
|---|---|---|---|---|---|---|
| Pseudo-F statistic | ||||||
| 1 | 1.24 (1.08–1.44) | 1.15 (0.99–1.33) | 1.07 (0.92–1.24) | 1.07 (0.92–1.24) | Ref. | 0.002 |
| 2 | 1.27 (1.09–1.46) | 1.18 (1.02–1.37) | 1.08 (0.93–1.26) | 1.08 (0.93–1.25) | Ref. | <0.001 |
| Amplitude | ||||||
| 1 | 2.55 (2.20–2.95) | 1.65 (1.40–1.93) | 1.49 (1.26–1.75) | 1.09 (0.91–1.30) | Ref. | <0.001 |
| 2 | 2.56 (2.21–2.97) | 1.71 (1.45–2.01) | 1.55 (1.31–1.83) | 1.12 (0.94–1.34) | Ref. | <0.001 |
| Mesor | ||||||
| 1 | 2.51 (2.17–2.90) | 1.65 (1.41–1.93) | 1.30 (1.10–1.54) | 1.16 (0.98–1.37) | Ref. | <0.001 |
| 2 | 2.59 (2.24–3.00) | 1.74 (1.48–2.03) | 1.35 (1.15–1.60) | 1.18 (1.00–1.40) | Ref. | <0.001 |
| Up mesor | ||||||
| 1 | 1.12 (0.96–1.30) | 1.04 (0.90–1.20) | Ref. | 1.02 (0.88–1.17) | 1.15 (1.00–1.32) | 0.667 |
| 2 | 1.06 (0.91–1.23) | 1.04 (0.90–1.20) | Ref. | 1.00 (0.87–1.15) | 1.05 (0.91–1.21) | 0.800 |
| Down mesor | ||||||
| 1 | 1.02 (0.89–1.18) | 0.93 (0.81–1.07) | Ref. | 0.98 (0.85–1.14) | 0.99 (0.85–1.15) | 0.890 |
| 2 | 0.93 (0.81–1.07) | 0.88 (0.77–1.02) | Ref. | 0.98 (0.85–1.14) | 0.98 (0.85–1.14) | 0.213 |
| Acrophase | ||||||
| 1 | 1.01 (0.88–1.17) | 0.99 (0.86–1.15) | Ref. | 0.91 (0.78–1.06) | 1.09 (0.95–1.26) | 0.691 |
| 2 | 0.91 (0.79–1.05) | 0.96 (0.83–1.11) | Ref. | 0.91 (0.79–1.06) | 1.03 (0.89–1.19) | 0.214 |
| IS | ||||||
| 1 | 0.91 (0.78–1.07) | 1.07 (0.93–1.24) | 1.15 (1.00–1.33) | 1.02 (0.88–1.18) | Ref. | 0.505 |
| 2 | 1.00 (0.85–1.17) | 1.18 (1.01–1.36) | 1.23 (1.06–1.42) | 1.07 (0.93–1.24) | Ref. | 0.513 |
| IV | ||||||
| 1 | Ref. | 1 (0.87–1.16) | 0.97 (0.84–1.12) | 1.12 (0.98–1.29) | 1.01 (0.87–1.17) | 0.389 |
| 2 | Ref. | 1.05 (0.91–1.21) | 1.04 (0.90–1.20) | 1.21 (1.06–1.40) | 1.10 (0.95–1.27) | 0.036 |
| RA | ||||||
| 1 | 5.18 (4.18–6.42) | 3.51 (2.82–4.38) | 2.44 (1.94–3.07) | 1.77 (1.39–2.24) | Ref. | <0.001 |
| 2 | 4.64 (3.74–5.76) | 3.41 (2.74–4.26) | 2.38 (1.89–2.99) | 1.76 (1.38–2.23) | Ref. | <0.001 |
| L5 | ||||||
| 1 | Ref. | 1.18 (1.00–1.39) | 1.33 (1.13–1.56) | 1.54 (1.32–1.80) | 2.14 (1.85–2.48) | <0.001 |
| 2 | Ref. | 1.17 (0.99–1.38) | 1.31 (1.11–1.53) | 1.48 (1.26–1.73) | 1.88 (1.62–2.19) | <0.001 |
| M10 | ||||||
| 1 | 3.86 (3.25–4.59) | 2.46 (2.05–2.96) | 1.99 (1.65–2.41) | 1.61 (1.32–1.96) | Ref. | <0.001 |
| 2 | 3.82 (3.20–4.55) | 2.54 (2.11–3.05) | 2.06 (1.70–2.49) | 1.64 (1.35–2.00) | Ref. | <0.001 |
| L5 start time | ||||||
| 1 | Ref. | 0.94 (0.81–1.09) | 1.09 (0.94–1.26) | 1.05 (0.91–1.21) | 1.23 (1.07–1.42) | <0.001 |
| 2 | Ref. | 0.97 (0.83–1.12) | 1.14 (0.99–1.32) | 1.08 (0.93–1.25) | 1.20 (1.04–1.38) | 0.004 |
| M10 start time | ||||||
| 1 | 1.06 (0.92–1.22) | 1.02 (0.88–1.17) | Ref. | 1.06 (0.92–1.22) | 1.04 (0.89–1.21) | 0.976 |
| 2 | 1.00 (0.87–1.15) | 0.99 (0.86–1.14) | Ref. | 1.05 (0.91–1.21) | 0.98 (0.84–1.14) | 0.881 |
Data are given as HR (95% CI).
Ref., reference group. *Model 1 adjusted for age (continuous), sex (male, female). Model 2 adjusted for age (continuous), sex (male, female), ethnicity (White, Asian or Asian British, Black or Black British, mixed), average total household income (£) before tax (<18,000, 18,000–30,999, 31,000–51,999, 52,000–100,000, >100,000), qualifications (college or university degree, Advanced Levels/Advanced Subsidiary Levels, Ordinary Levels/ General Certificate of Secondary Education, Certificate of Secondary Education, National Vocational Qualification or Higher National Certificate or Diploma, other professional qualifications), smoking status (never, previous, current), alcohol drinking status (never, previous, current), night-shift work status (never, sometimes, usually, always), sleep duration (continuous), and total physical activity as measured by overall acceleration average per day (continuous). Detailed descriptions of rest–activity rhythm parameters are presented in Supplementary Table 1.
Figure 1 shows the relative importance of each variable in predicting incident diabetes using a random forest model, with a higher mean difference in the Gini index suggesting a higher effectiveness of the variable in predicting incident diabetes. Of all the variables included in the model, BMI emerged as having the highest importance or predictive value in our sample, followed by the pseudo-F statistic. A cluster of 12 second-tier predictors included most of the rest–activity variables and total physical activity. We found similar patterns when random forest analyses were stratified by age, sex, BMI, and night-shift status (Supplementary Figs. 1–4); in all subgroup samples, BMI was consistently the most important predictor, and most of the rest–activity parameters ranked among the top predictors.
Figure 1.
Variable importance in predicting incident diabetes from random forest model. Importance of each variable in predicting incident diabetes was assessed by difference in Gini impurity index comparing models with and without the variable. Larger decrease in Gini value when certain predictor is left out suggests inclusion of the variable in model leads to more effective classification of participants as incident diabetes cases or noncases. Variables with larger decreases in Gini indicate higher importance of variable in making more accurate classification.
Conclusions
In the UK Biobank, our analyses showed that rest–activity characteristics, including lower pseudo-F-statistic, IS, amplitude, mesor, RA, M10, later L5 start time, and higher L5, were associated with higher risks of incident diabetes. Random forest models consistently ranked BMI as the most important predictor of incident diabetes, followed by many of the rest–activity parameters, which ranked higher than several other conventional diabetes risk factors. The results were generally similar across different subgroup analyses by age, sex, shift work status, and BMI.
Several previous studies have examined the prospective relationship between rest–activity parameters and diabetes risk (7,8). For example, in our study in a cohort of older men, we used the parametric algorithm to derive rest–activity parameters based on an extended cosine model. Our analysis found that the lowest quartile of amplitude derived from the extended cosine model was associated with a 3.8-fold higher diabetes risk when compared with the highest quartile (7). Moreover, a recent study by Wu et al. (8) examined the relationship between five of the seven nonparametric rest–activity parameters (i.e., IS, IV, RA, L5, and M10) and diabetes risk in UK Biobank participants. The study found that the lowest quartiles of RA and M10 were associated with 2.1- and 1.3-fold higher risks of type 2 diabetes, respectively. Our study extended the existing research by conducting a comprehensive assessment, including both parametric and nonparametric parameters. The parametric method assumes the presence of a cosine or cosine-like waveform and therefore may not provide accurate estimates for individuals whose activity patterns deviate from this assumption. The nonparametric method does not rely on such an assumption and captures a different set of parameters, each assessing a specific aspect of the rhythmic pattern. These two methods are complementary, and by applying both, our study offers a more comprehensive assessment of the rest–activity rhythm. More importantly, it enabled us to use machine learning methods to make direct comparisons across multiple different yet correlated rest–activity parameters and rank their relative importance with respect to diabetes risk.
Of the rest–activity parameters examined in our study, amplitude, mesor, RA, and M10 exhibited the largest effect sizes in relation to incident diabetes. Participants in the lowest quintile for these parameters were 2.5–4.6 times more likely to develop incident diabetes when compared with those in the highest quintile. These findings are consistent with multiple previous studies that focused on rest–activity rhythms and diabetes. These parameters are largely influenced by total and daytime physical activity levels, and therefore, their inverse association with diabetes incidence may be primarily driven by the well-established benefit of physical activity in type 2 diabetes prevention. However, it is worth noting that the magnitude of these associations seemed larger than that reported for physical activity (15). For example, a meta-analysis reported a 25–35% reduction in diabetes risk comparing the highest with the lowest level (the definition of levels varied across studies but roughly corresponded to quartiles and quintiles in many studies) of physical activity, which is considerably lower than the effect size we observed (16). This may be due to the fact that the calculation of most of these parameters took into account not only the daytime activity levels but also the 24-h rhythm, including the resting period. Notably, in our random forest model analysis, amplitude and M10 ranked higher than total physical activity with respect to their importance in predicting incident diabetes. Finally, in the sensitivity analysis with additional adjustment for total physical activity, the associations between these rest–activity parameters and type 2 diabetes incidence were attenuated but remained statistically significant, suggesting that the observed results could not be fully explained by the effects of total activity level. Taken together, these parameters, which represent a better characterization of the overall rest–activity pattern, may be strong predictors of diabetes risk independent of physical activity.
A higher level of L5 was positively associated with type 2 diabetes incidence. L5 is a measure of activity levels in the least active (i.e., resting) period of the day, and thus, a higher value of L5 may represent disrupted and/or irregular sleep. Therefore, our results may have been partially driven by the relationship between sleep and diabetes. Suboptimal sleep has been consistently linked to type 2 diabetes risk in the literature (17). Specifically, meta-analysis studies and systematic reviews have shown that short and long sleep duration (18), poor sleep quality (19), sleep disorders such as sleep apnea (20), and irregular sleep patterns (21,22) all predicted a higher risk of developing type 2 diabetes. Moreover, previous studies have shown that sleep deprivation was associated with decreased levels of leptin, increased levels of ghrelin, and impaired insulin sensitivity, potentially leading to overeating and weight gain (23,24). Additionally, chronic sleep deprivation contributes to increased stress hormone levels, such as cortisol, which can further impair insulin sensitivity and increase the risk of type 2 diabetes (25). It is worth noting that adjustment for self-reported sleep duration (and total physical activity) in model 3 did not attenuate the relationship between L5 and diabetes risk, and the random forest analysis showed that L5 was a stronger predictor of diabetes risk than sleep duration. This may be because the calculation of L5 was based on per-minute actigraphy data recorded across multiple days, and thus, it reflects the level of sleep fragmentation and sleep regularity, which are important diabetes risk factors that are often not captured by self-reported sleep duration. Therefore, our findings suggest that L5 may be a unique measure of the resting phase with the potential to predict diabetes risk.
Our parametric and nonparametric methods derived multiple variables related to the timing or phase of the rest–activity cycle (i.e., acrophase, down mesor, up mesor, start time of L5, and M10). Of these, only the start time of L5 was significantly and positively associated with diabetes risk. The start time of L5 may be primarily driven by sleep timing, which is influenced by both intrinsic and environmental factors. Chronotype, which measures the propensity to sleep at a certain time of the day, has been previously linked to diabetes risk. Specifically, existing research has been largely consistent in showing an association between a later chronotype and more impaired metabolic health, including poor glycemic control and a higher risk of diabetes (26–30). Thus, our finding regarding the relationship between L5 start time and type 2 diabetes risk corroborates previous findings on sleep timing and metabolic health.
Of the rest–activity parameters, pseudo-F statistic, IS, and IV measure different properties of overall rhythmicity and have been widely used in epidemiologic studies focusing on rest–activity rhythms and health outcomes. Of these, the only parameter that showed a significant association with diabetes risk was the pseudo-F statistic, of which a higher value indicates superior goodness of fit of the extended cosine model and therefore more robust rhythmicity. Our previous studies have also linked a lower pseudo-F statistic with diabetes and impaired glycemic control (6,7), and the results seemed to be similar across different demographic groups in a large, diverse, and nationally representative sample of the US population. Although the relationship between pseudo-F statistic and diabetes incidence was attenuated and became statistically nonsignificant after adjusting for sleep, physical activity, and BMI (model 4), it is worth noting that in the random forest analysis, which included3 all covariates, this variable emerged as the second strongest predictor of diabetes risk, ranking lower only than BMI, in the overall sample and in almost all subgroups (except for groups with age ≥60 years and obesity). This suggests that in the random forest models, a machine learning approach widely used for risk prediction, the pseudo-F statistic played an important and unique role in predicting diabetes risk, even after accounting for other traditional diabetes risk factors. Discrepancies between the results of the sensitivity analysis using Cox regression and those from the random forest models may be explained by different model algorithms.
It was intriguing to observe that rest–activity parameters improved risk prediction of diabetes and ranked higher than many traditional diabetes risk factors in our sample. This finding may be explained by several mechanisms. First, although sleep, physical activity, and sedentary behaviors are the essential components of the rest–activity rhythm, the parametric and nonparametric rest–activity variables may represent unique features of the 24-h behavioral cycle that are not fully captured by conventional measures of sleep and activity (e.g., sleep duration, total physical activity), and therefore, their inclusion may have further improved model performance. Second, characteristics of the rest–activity cycle may serve as proxy measures of the internal circadian rhythm, which plays a central role in regulating human metabolism (31). Mutations in circadian genes lead to impaired β-cell function and diabetes in animal studies (32), and circadian misalignment induced by simulated shift work caused impaired glucose tolerance and reduced insulin sensitivity in humans (33,34). Although the underlying mechanisms remain an important topic for further investigation, our results suggest that rest–activity rhythm characteristics may be strong and consistent predictors of diabetes, and future research should assess whether the incorporation of these variables into risk prediction algorithms for type 2 diabetes leads to improved test performance.
There are several strengths of our study. First, rest–activity rhythms were characterized by multiple parameters derived from multiday actigraphy data using both parametric and nonparametric methods. The comprehensive assessment allowed us to simultaneously evaluate the relative importance of different rest–activity metrics in predicting type 2 diabetes risk, which yielded valuable information that may help develop risk prediction models based on these behavioral biomarkers. Second, the UK Biobank has a large sample size and rich information on a wide range of covariates, which allowed us to adjust for an extensive list of confounding factors and perform subgroup analyses. Third, the UK Biobank has relatively longer follow-up data compared with previous studies, which not only improved statistical power but also reduced the likelihood of reverse causation. Fourth, the use of multiple sources (e.g., self-reported conditions, primary care data, hospital inpatient data, death register records) to capture the occurrence of diabetes provided more accurate data in determining incident and prevalent diabetes cases and healthy individuals, when compared with, for example, use of hospital records only.
Our analysis also has a number of limitations. First, UK Biobank participants are predominantly White and have a relatively high socioeconomic status when compared with the background UK population, and therefore, our findings may have limited generalizability to some populations. In particular, although the use of random forest to determine the predictive power of different variables is a strength, it is worth noting that the rankings derived by such methods are sample specific. Therefore, there is a need for future studies to replicate these analyses in samples with sufficient demographic, socioeconomic, and geographic diversity. Second, although we controlled for a variety of confounding factors, unmeasured and residual confounding may still be present. Replication of the results in different populations with different confounding structures or use of other study designs such as Mendelian randomization may help elucidate the causal nature of the observed associations. Third, primary care data were used to define incident and prevalent cases in addition to other data (e.g., hospital inpatient data, death register records, self-reported data, published algorithms), but primary care data were only available in ∼45% of the UK Biobank cohort (35). Therefore, incident diabetes captured in our study may have occurred in individuals with more medically complex issues than if primary care records had been available for all participants. Fourth, the actigraphy data in our study were collected only once over 5–7 days, which limited our ability to capture changes in rest–activity patterns and assess how such changes may predict diabetes risk. Finally, previous studies have reported shifts in temporal trends of sleep, physical activity, and sedentary behaviors (36–38), which are essential behavioral components of the rest–activity rhythm. In particular, the recent COVID-19 pandemic has been shown to have had a substantial impact on rest–activity behaviors (39,40), although it is unclear whether and to what degree these changes may persist. Because our analysis used actigraphy data collected between 2013 and 2015, the observed rest–activity rhythms may not reflect population patterns at a different time point. Given the importance of rest–activity rhythms in health, future studies should monitor trends of key rest–activity metrics in the population for a better understanding of their health impact.
In conclusion, we found that variations in multiple metrics of the rest–activity rhythm predicted risk of incident diabetes in UK Biobank participants. In particular, diabetes incidence was elevated among individuals with rhythmic patterns characterized by lower amplitude and average activity level, reduced activity levels during the active phase and higher levels during the resting phase, a later start time of the resting phase, and impaired overall rhythmicity. The strength of the effect sizes and consistency of these findings across subgroups suggest that these metrics have the potential to be used as digital biomarkers in risk prediction algorithms for type 2 diabetes.
This article contains supplementary material online at https://doi.org/10.2337/figshare.29087657.
Article Information
Duality of Interest. No potential conflicts of interest relevant to this article were reported.
Author Contributions. C.H.C.Y. wrote the initial manuscript draft. C.H.C.Y., M.K.R., and Q.X. interpreted the data. C.H.C.Y. and Q.X. performed the statistical analysis. M.K.R. and Q.X. were responsible for the study concept and design. All authors critically reviewed the manuscript for important intellectual content. Q.X. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Handling Editors. The journal editors responsible for overseeing the review of the manuscript were Elizabeth Selvin and Alka M. Kanaya.
Funding Statement
This work was supported by the National Institute of Aging (R01AG063946), National Heart, Lung, and Blood Institute (R21HL165369), and National Institute for Health and Care Research Manchester Biomedical Research Centre (NIHR203308). This research has been conducted using the UK Biobank resource under Application Number 6818. We would like to thank the participants and research team from the UK Biobank who contributed or collected data.
Supporting information
References
- 1. Patke A, Young MW, Axelrod S.. Molecular mechanisms and physiological importance of circadian rhythms. Nat Rev Mol Cell Biol 2020;21:67–84 [DOI] [PubMed] [Google Scholar]
- 2. Ancoli-Israel S, Cole R, Alessi C, Chambers M, Moorcroft W, Pollak CP.. The role of actigraphy in the study of sleep and circadian rhythms. Sleep 2003;26:342–392 [DOI] [PubMed] [Google Scholar]
- 3. Feng H, Yang L, Ai S, et al. Association between accelerometer-measured amplitude of rest-activity rhythm and future health risk: a prospective cohort study of the UK Biobank. Lancet Healthy Longev 2023;4:e200–e210 [DOI] [PubMed] [Google Scholar]
- 4. Khan MAB, Hashim MJ, King JK, Govender RD, Mustafa H, Al Kaabi J.. Epidemiology of type 2 diabetes: global burden of disease and forecasted trends. J Epidemiol Glob Health 2020;10:107–111 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Bommer C, Sagalova V, Heesemann E, et al. Global economic burden of diabetes in adults: projections from 2015 to 2030. Diabetes Care 2018;41:963–970 [DOI] [PubMed] [Google Scholar]
- 6. Xiao Q, Matthews CE, Playdon M, Bauer C.. The association between rest-activity rhythms and glycemic markers: the US National Health and Nutrition Examination Survey, 2011–2014. Sleep 2022;45:zsab291. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Xiao Q, Qian J, Evans DS, et al.; Osteoporotic Fractures in Men (MrOS) Study Group . Cross-sectional and prospective associations of rest-activity rhythms with metabolic markers and type 2 diabetes in older men. Diabetes Care 2020;43:2702–2712 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Wu H, Liu X, Jiang W, et al. The rest–activity rhythm, genetic susceptibility and risk of type 2 diabetes: a prospective study in UK Biobank. Diabetes Obes Metab 2023;25:3366–3376 [DOI] [PubMed] [Google Scholar]
- 9. Bycroft C, Freeman C, Petkova D, et al. The UK Biobank resource with deep phenotyping and genomic data. Nature 2018;562:203–209 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. UK Biobank . Physical activity monitor (accelerometer), version 1.0. Published 27 January 2016. Accessed 18 June 2025. Available from https://biobank.ndph.ox.ac.uk/ukb/ukb/docs/PhysicalActivityMonitor.pdf
- 11. Doherty A, Jackson D, Hammerla N, et al. Large scale population assessment of physical activity using wrist worn accelerometers: the UK biobank study. PLoS One 2017;12:e0169649. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Eastwood SV, Mathur R, Atkinson M, et al. Algorithms for the capture and adjudication of prevalent and incident diabetes in UK Biobank. PLoS One 2016;11:e0162388. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. UK Biobank . First occurrence of health outcomes defined by 3-character ICD10 code. Published September 2019. Accessed 18 June 2025. Available from https://biobank.ndph.ox.ac.uk/ukb/ukb/docs/first_occurrences_outcomes.pdf
- 14. Breiman L. Random forests. Machine Learning 2001;45:5–32 [Google Scholar]
- 15. Hayes C, Kriska A.. Role of physical activity in diabetes management and prevention. J Am Diet Assoc 2008;108(Suppl. 1):S19–S23 [DOI] [PubMed] [Google Scholar]
- 16. Aune D, Norat T, Leitzmann M, Tonstad S, Vatten LJ.. Physical activity and the risk of type 2 diabetes: a systematic review and dose-response meta-analysis. Eur J Epidemiol 2015;30:529–542 [DOI] [PubMed] [Google Scholar]
- 17. Antza C, Kostopoulos G, Mostafa S, Nirantharakumar K, Tahrani A.. The links between sleep duration, obesity and type 2 diabetes mellitus. J Endocrinol 2021;252:125–141 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Shan Z, Ma H, Xie M, et al. Sleep duration and risk of type 2 diabetes: a meta-analysis of prospective studies. Diabetes Care 2015;38:529–537 [DOI] [PubMed] [Google Scholar]
- 19. Lee SWH, Ng KY, Chin WK.. The impact of sleep amount and sleep quality on glycemic control in type 2 diabetes: a systematic review and meta-analysis. Sleep Med Rev 2017;31:91–101 [DOI] [PubMed] [Google Scholar]
- 20. Qie R, Zhang D, Liu L, et al. Obstructive sleep apnea and risk of type 2 diabetes mellitus: a systematic review and dose-response meta-analysis of cohort studies. J Diabetes 2020;12:455–464 [DOI] [PubMed] [Google Scholar]
- 21. Zuraikat FM, Makarem N, Redline S, Aggarwal B, Jelic S, St-Onge M-P.. Sleep regularity and cardiometabolic heath: is variability in sleep patterns a risk factor for excess adiposity and glycemic dysregulation? Curr Diab Rep 2020;20:38. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Fritz J, Phillips AJK, Hunt LC, et al. Cross-sectional and prospective associations between sleep regularity and metabolic health in the Hispanic Community Health Study/Study of Latinos. Sleep 2021;44:zsaa218. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Taheri S, Lin L, Austin D, Young T, Mignot E.. Short sleep duration is associated with reduced leptin, elevated ghrelin, and increased body mass index. PLoS Med 2004;1:e62. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Reutrakul S, Van Cauter E.. Sleep influences on obesity, insulin resistance, and risk of type 2 diabetes. Metabolism 2018;84:56–66 [DOI] [PubMed] [Google Scholar]
- 25. Kumari M, Badrick E, Ferrie J, Perski A, Marmot M, Chandola T.. Self-reported sleep duration and sleep disturbance are independently associated with cortisol secretion in the Whitehall II study. J Clin Endocrinol Metab 2009;94:4801–4809 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Kianersi S, Liu Y, Guasch-Ferré M, et al. Chronotype, unhealthy lifestyle, and diabetes risk in middle-aged U.S. women: a prospective cohort study. Ann Intern Med 2023;176:1330–1339 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Reutrakul S, Hood MM, Crowley SJ, et al. Chronotype is independently associated with glycemic control in type 2 diabetes. Diabetes Care 2013;36:2523–2529 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Anothaisintawee T, Lertrattananon D, Thamakaison S, Knutson KL, Thakkinstian A, Reutrakul S.. Later chronotype is associated with higher hemoglobin A1c in prediabetes patients. Chronobiol Int 2017;34:393–402 [DOI] [PubMed] [Google Scholar]
- 29. Hashemipour S, Yazdi Z, Mahabad N.. Association of evening chronotype with poor control of type 2 diabetes: roles of sleep duration and insomnia level. Int J Endocrinol Metab 2020;18:e99701. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30. Tan X, Ciuculete D-M, Schiöth HB, Benedict C.. Associations between chronotype, MTNR1B genotype and risk of type 2 diabetes in UK Biobank. J Intern Med 2020;287:189–196 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Stenvers DJ, Scheer FAJL, Schrauwen P, la Fleur SE, Kalsbeek A.. Circadian clocks and insulin resistance. Nat Rev Endocrinol 2019;15:75–89 [DOI] [PubMed] [Google Scholar]
- 32. Marcheva B, Ramsey KM, Buhr ED, et al. Disruption of the clock components CLOCK and BMAL1 leads to hypoinsulinaemia and diabetes. Nature 2010;466:627–631 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33. Scheer FAJL, Hilton MF, Mantzoros CS, Shea SA.. Adverse metabolic and cardiovascular consequences of circadian misalignment. Proc Natl Acad Sci U S A 2009;106:4453–4458 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34. Qian J, Dalla Man C, Morris CJ, Cobelli C, Scheer FAJL.. Differential effects of the circadian system and circadian misalignment on insulin sensitivity and insulin secretion in humans. Diabetes Obes Metab 2018;20:2481–2485 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35. UK Biobank . Primary care linked data, version 2.0. Published April 2024. Accessed 18 June 2025. Available from https://biobank.ndph.ox.ac.uk/showcase/showcase/docs/primary_care_data.pdf
- 36. Xu J, Luo L, Gamaldo A, et al. Trends in sleep duration in the U.S. from 2004 to 2018: a decomposition analysis. SSM Popul Health 2024;25:101562. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37. Keadle SK, McKinnon R, Graubard BI, Troiano RP.. Prevalence and trends in physical activity among older adults in the United States: a comparison across three national surveys. Prev Med 2016;89:37–43 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38. Yang L, Cao C, Kantor ED, et al. Trends in sedentary behavior among the US population, 2001-2016. JAMA 2019;321:1587–1597 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39. Ceolin C, Limongi F, Siviero P, et al. Changes in sleep duration and sleep timing in the general population from before to during the first COVID-19 lockdown: a systematic review and meta-analysis. Int J Environ Res Public Health 2024;21:583. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40. Okely JA, Corley J, Welstead M, et al. Change in physical activity, sleep quality, and psychosocial variables during COVID-19 lockdown: evidence from the Lothian Birth Cohort 1936. Int J Environ Res Public Health 2020;18:210. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.