Accelerometer-derived physical activity and mortality in individuals with type 2 diabetes


Study design and population

The datasets used in this study were obtained from the UK Biobank (Application: 79095), which received approval from the North West Multi-centre Research Ethics Committee (R21/NW/0157), and the Biomedical Research Ethics Committee of Hangzhou Normal University (200400001). All participants gave written informed consent and have signed the informed consent to be linked to national electronic health-related datasets32. The UK Biobank is a large, population-based prospective cohort, in which more than 0.5 million participants aged 37–73 years attended 1 of the 22 assessment centers across England, Scotland, and Wales to complete touchscreen questionnaires, physical examinations, and biological sample collections between 2006 and 2010 (baseline of the UK Biobank). Then between 2013 and 2015 (baseline of this study), 240,000 invitations were randomly sent to them for PA measurement by accelerometers, with a response rate of 44%. Devices were dispatched for 106,053 participants and, of these, data were received from 103,666. In this study, participants with T2D at baseline were identified through an integration of multiple data sources: self-reported T2D, measured random glucose level ≥11.1 mmol/L or glycated hemoglobin (HbA1c) level ≥48 mmol/L according to the American Diabetes Association criteria, and code E11 in the electronic health records (England and Wales: Health Episode Statistics; Scotland: Scottish Morbidity Records) according to the 10th Revision of the International Classification of Diseases (ICD-10). The algorithms for the capture and adjudication of prevalent T2D was consistent with previous research9.

Device-measured physical activity

Participants were asked to wear Axivity AX3 (Newcastle upon Tyne, UK) triaxial accelerometers on their dominant hands for 7 days, when the sensor captured the acceleration at 100 Hz with a dynamic range of ±8 g (unit of gravity)13. Participants with insufficient wear time (<72 h), poor device calibration, or daylight saving time shift during their wear period were excluded in this study. Minutes per week of LPA, MPA, and VPA were determined as the time spent in 30–125 milligravities (mg), >125–400 mg, and >400 mg intensity activity, respectively25. The duration of MVPA was estimated as the sum of MPA and VPA.

Outcomes ascertainment

The primary outcomes of interest in this study were all-cause, cancer, and CVD mortality. Information on the date and cause of death (C00-C97 for cancer and I00-I99 for CVD according to ICD-10) were obtained from death certificates held by the National Health Service (NHS) Information Center (England and Wales) and the NHS Central Register (Scotland). The follow-up time began at the accelerometer completion and ended at the occurrence of death or the end of follow-up (November 12, 2021), whichever came first.

Covariates ascertainment

Covariates were selected based on a prior-defined directed acyclic graph (Supplementary Fig. 3). Our study finally included age (years between birth date and the start of accelerometry), sex (acquired from central registry at recruitment; Field ID: 31), ethnicity, education, smoking status, alcohol intake (product of intake frequency and gram of each type of alcohol per one standard drink)33, diet quality, sleep quality, self-rated health status, history of cancer, CVD, hypertension, long-standing illness, disability, or infirmity, and experience of serious illness, injury, bereavement, stress in last two years derived from the questionnaires, body mass index (BMI, kg/m2, weight in kilograms divided by height in meters squared) and waist circumstance (cm) measured during the initial visit, season and wear duration recorded by the accelerometers, and diabetes duration (years between the first occurrence of diabetes and the start of accelerometry). Diet scores (0–7) were by assigning 1 point for a healthy frequency and 0 point for an unhealthy frequency of consumption of fruits, vegetables, fish, processed meat, unprocessed red meat, whole grains, and refined grains, with higher scores indicating a healthier diet quality34,35. Similarly, sleep scores (0–5) were generated by incorporating five sleep factors: chronotype, sleep duration, insomnia, snoring, and excessive daytime sleepiness, with higher scores indicating better sleep quality36.

Statistical analyses

Baseline characteristics of the participants according to different volumes of PA were described as means and standard deviations (SDs) for continuous variables and numbers (percentages) for categorical variables.

Dose–response associations of PA with all-cause and cause-specific mortality were evaluated using restricted cubic splines fitted in the Cox proportional hazards models (“rms” package in R). The reference values were set at the 1st percentile, and knots at the 5th, 35th, 65th, and 95th percentiles of the PA distribution. Potential non-linearity was tested by Wald tests. Then, the duration of PA was categorized into four levels: <1750, 1750–2099, 2100–2449, and ≥2450 min/week for LPA, <150, 150–299, 300–449, and ≥450 min/week for MPA, 0, 1–14, 15–29, and ≥30 min/week for VPA, and <275, 275–449, 450–624, and ≥625 min/week for MVPA after considering recommendations in the current PA guideline (≥150 min/week of MPA or ≥75 min/week of VPA), the inflection points of the dose–response relationship in this study (about 1750, 400, 15, and 375 min/week for LPA, MPA, VPA, and MVPA), and the sample size of each group. Cox proportional hazard models were performed to estimate the HRs and 95% CIs of all-cause and cause-specific mortality risk. Linear trends were examined by entering the median value of each category of PA as a continuous variable into the models. Three multivariable-adjusted models were constructed to account for potential confounding: model 1 was the crude model that adjusted for age (continuous, in years), sex (male or female), ethnicity (white or others), education (college/university or others), season (spring, summer, autumn, or winter) and duration of the accelerometry (continuous); model 2 was further adjusted for smoking status (never, former, or current), alcohol intake (continuous, in gram/day), diet scores (categories, 0–7), and sleep scores (continuous, 0–5); model 3 was the main model that additionally adjusted for BMI (continuous, in kg/m2), waist circumference (continuous, in cm), self-rated health (excellent, good, fair, or poor), long-standing illness, disability or infirmity (yes or no), illness, injury, bereavement, stress in last 2 years (yes or no), history of cancer or CVD (yes or no), history of hypertension (yes or no), and diabetes duration (continuous, in years). In the primary analyses, age, season at the time of accelerometry recording, and history of cancer or CVD were treated as time-varying covariates by constructing interaction terms between covariates and time. The proportional hazard assumption for Cox models was checked with Schoenfeld residuals and no violation was observed (Supplementary Table 27). Furthermore, a risk matrix was used to investigate the joint associations of different-intensity PA with all-cause and cause-specific mortality. The combination of the least active groups was set as the reference group. The PAFs were calculated to estimate the percentage of all death cases that would be prevented if individuals in the less active group were as active as the most active one37,38. Missing values in covariates with a missing rate <1% (e.g., smoking status and BMI) were completely excluded, while those with a missing rate >1% (e.g., education and diet scores) were coded as an additional category for categorical variables or mean values for continuous variables.

Stratified analyses were conducted according to age (<60 years and ≥60 years), sex (male and female), BMI (<25 kg/m2 and ≥25 kg/m2), waist circumference (<102 cm for male and <88 cm for female, and other), smoking status (never and ever), alcohol intake (<28 g/day for male and <14 g/day for female, and other), diet scores (<4 and ≥4), sleep scores (<3 and ≥3) and history of hypertension (yes and no) to examine whether the associations varied by these factors. Interaction terms were tested by a wald test.

Several sensitivity analyses were carried out to assess the robustness of our results. First, participants with cancer or CVD at baseline were excluded when assessing the association between PA and corresponding cause-specific mortality. Second, participants with poor self-rated health status were excluded, considering that they were more likely to undertake less PA and have higher risks of mortality. Third, we excluded participants who died within the first 2 and 4 years of follow-up to avoid the potential risk of reverse causation. Fourth, we performed stratified analyses by the number of diabetes severity factors, including glycated hemoglobin level (HbA1c) ≥ 7.0%, diabetes duration ≥10 years, and insulin medication use. Fifth, these factors were treated as covariates in the adjusted models. Sixth, T2D-related complications including related eye disease were additionally adjusted in the model. Seventh, LPA, MPA, and VPA variables were mutually adjusted to evaluate whether the associations were attributable to other intensity. Eighth, we stratified the cancer deaths by T2D/Obesity-related cancers and T2D/Obesity-independent cancers. Ninth, we repeated the main analysis after imputing the missing values of the covariates using chained equation multiple imputations (Supplementary eMethods). Tenth, we employed the Fine and Gray sub-distribution hazard model to examine the associations between PA and CVD mortality, accounting for cancer deaths as a competing risk. Similarly, we considered CVD deaths as a competing risk when analyzing the relationship between PA and cancer mortality through competing risk analyses. Finally, total PA in MET-min/week was estimated as the sum of three PA modes (LPA, MPA, and VPA) according to guidelines for data processing of the International Physical Activity Questionnaire39. Additionally, we repeated the analysis in all-cause mortality after stratified the participants into two distinct groups: those who meet the current guideline PA recommendations and those who did not align with the recommendations.

All the analyses were conducted using STATA 16 statistical software (Stata Corp LLP, College Station, TX) and R software (version 4.1.3). The statistical significance was set as P < 0.05 (two-sided test).

Reporting summary

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


Leave a Reply

Your email address will not be published. Required fields are marked *