Associations of “Weekend Warrior” Physical Activity With Incident Disease and Cardiometabolic Health

The UK Biobank is a prospective cohort of 502 629 participants enrolled between 2006 and 2010.⁹ Briefly, 9.2 million individuals 40 to 69 years of age living within 25 miles of 22 assessment centers in the United Kingdom were invited, and 5.4% participated in the baseline assessment. Within the physical activity measurement substudy, 103 695 participants submitted data from an Axivity AX3 (Newcastle upon Tyne, United Kingdom) wrist-based triaxial accelerometer worn for 1 week.¹⁰ The sensor captured continuous acceleration at 100 Hz with dynamic range ±8g. As described previously, acceleration signals were calibrated to gravity.¹⁰ Sample data were combined into 5-s epochs, each representing average vector magnitude. Non–wear-time was identified as consecutive stationary episodes ≥60 min in which all 3 axes had SD<13.0 mg.¹⁰ Epochs representing non–wear-time were imputed on the basis of the average of similar time-of-day vector magnitude and intensity distribution data points on different days. We excluded individuals whose wear-time was insufficient to support imputation (ie, ≥1 1-hour period of the 24-hour cycle [eg, 0900–1000] during which no wear data were ever accrued), whose signals were insufficient for calibration or MVPA estimation, whose mean acceleration values were previously adjudicated as nonphysiological, and who contributed less than a full week of activity data.^8,11

MVPA was classified using a published machine-learning–based method developed to categorize a broad range of activities (eg, walking, jogging, stationary cycling, elliptical, and others) and validated in a United Kingdom–based sample.¹¹ For our primary analysis, we defined active status at the guideline-based threshold (≥150 minutes/week).^2,4,8 Individuals were classified as inactive (below MVPA threshold), weekend warrior (at or above MVPA threshold and ≥50% of total MVPA over 1 to 2 days⁶), and regular (at or above MVPA threshold but not weekend warrior). Because optimal MVPA levels using wrist-based accelerometers are unclear,¹² in secondary analyses, we assessed alternative thresholds (outlined below).

We investigated associations between physical activity pattern and incident disease using a phenome-wide association study approach, using Cox proportional hazards models with physical activity pattern (ie, weekend warrior, regular activity, or inactive) as the exposure of interest, and adjusted for age at activity measurement, sex, ethnic background, tobacco use, Townsend Deprivation Index, alcohol intake, educational attainment, employment status, self-reported health, and diet quality (definitions in Table S1). Covariates were selected a priori on the basis of known or suspected associations with physical activity and disease reported in previous literature as well as previous analyses of activity patterns in the UK Biobank.^1,8,11 In our models, the effect of physical activity pattern on a given disease was estimated as the hazard ratio and P value corresponding to the respective activity category (weekend warrior or regular) compared with the inactive referent. Subsequently, to assess whether risk of disease might differ for weekend warrior versus regular activity, we calculated hazard ratios and P values corresponding to weekend warrior activity using regular activity as the referent. Given baseline imbalances in total MVPA time between the weekend warrior and regular activity groups (Figure 1), and our intent to identify the potential effects of physical activity pattern for a given physical activity volume, all models directly comparing weekend warrior versus regular activity were in addition adjusted for total MVPA volume.⁸ In secondary analyses, to assess whether the observed imbalance in MVPA volume may itself possess relevance for disease risk, we compared weekend warrior versus regular activity without adjustment for MVPA volume. The functional form of continuous covariates was assessed by fitting penalized splines (Supplemental Methods), and secondary models were fit to accommodate nonlinear effects when present (see below). All models used a complete case analysis (exclusions for missing data are shown in Figure 1).

To define outcomes, we used v1.2 of the Phecode Map,¹³ comprising 1867 disease definitions organized into clinically relevant categories and defined using standardized sets of codes from the International Classification of Diseases, Ninth Revision and Tenth Revision. These Phecode definitions can be accessed at https://phewascatalog.org/. Sources of diagnostic codes encompassed hospital data obtained through linkage with national health-related datasets and outpatient general practitioner visit data accessed through electronic health records. To prevent model instability, only diseases with ≥120 events were tested (ie, ≥10 events per covariate), resulting in a total of 678 conditions included in association testing.¹⁴ Given our interest in incident disease and the age distribution of the sample, we excluded Phecodes corresponding to pregnancy-related conditions and congenital anomalies. Significance thresholds were corrected for multiplicity to target a false discovery rate (FDR) of 0.01.¹⁵ The FDR process corrects for multiple testing by controlling the expected proportion of incorrectly rejected null hypotheses (ie, false discoveries or type 1 errors). Here, we used a stringent threshold (FDR 0.01), which would be expected to result in only one false discovery per 100 associations detected. Further details are provided in the Supplemental Methods.

For analyses of incident disease, the assessment period commenced at the conclusion of physical activity measurement with accelerometry and concluded at the occurrence of an event, death, or last follow-up, whichever transpired first. Individuals were excluded from analysis of a particular incident disease if they met criteria for that disease at the time of accelerometer wear. The duration of follow-up was contingent upon the availability of linked health data, with a defined end date of March 31, 2021, for participants enrolled in England and Scotland, and February 28, 2018, for those enrolled in Wales. For reference, we tabulated the 5-year cumulative incidence, and person-time incidence rates, for all conditions included in association testing.

To assess the robustness of our findings, we performed several secondary analyses. First, we repeated disease association testing using the sample median MVPA (≥230.4 minutes) as the threshold for activity, with the median chosen on the basis of United Kingdom national health surveys reporting that roughly half of individuals are physically active.¹⁶ Second, to explore whether our findings may have been driven by reverse causation, we repeated association testing while excluding all events occurring within 2 years of physical activity measurement. Third, to quantify the robustness of observed associations to residual confounding, we calculated E-values, which represent the minimum strength of association on the risk ratio scale that a potential confounder would need to have with the exposure and outcome to nullify the observed association.¹⁷ Fourth, we compared weekend warrior and regular activity without adjustment for MVPA volume. Fifth, we tested the following alternative definitions of weekend warrior activity: (1) ≥150 minutes of MVPA/week with ≥50% of total MVPA achieved over 1 or 2 consecutive days (rather than any 1 or 2 days of the week), (2) ≥150 minutes of MVPA/week with ≥50% of total MVPA achieved over 1 or 2 true weekend days, (3) ≥150 minutes of MVPA/week with ≥75% of total MVPA achieved over 1 or 2 days, and (4) ≥150 minutes of MVPA/week with ≥75% of total MVPA achieved over 1 or 2 consecutive days. Sixth, we performed analogous association testing for selected cardiometabolic factors of specific interest (ie, hypertension, diabetes, or obesity) using alternative definitions combining Phecodes with relevant anthropometric measurements, laboratory values, and medication use. Seventh, we fit models accounting for potential nonlinear relations between alcohol intake and MVPA with disease outcomes using spline terms. Eighth, to assess associations between MVPA volume and disease, we fit Cox proportional hazards models with 4 exemplar cardiometabolic factors of specific interest (ie, hypertension, diabetes, sleep apnea, and obesity) as outcomes and MVPA quartile as the main exposure, with adjustment for the same covariates as the primary analyses. Further details are provided in the Supplemental Methods.

Analyses were performed using R v4.0. The proportional hazards assumption was assessed by performing the Grambsch-Therneau test of correlation¹⁸ and inspecting smoothed fits of Schoenfeld residuals versus time (Supplemental Methods).

Data processing scripts used to perform the analyses described herein using UK Biobank data are available at https://github.com/shaankhurshid/weekend_warrior_phewas. All UK Biobank data are available for researchers by application (https://www.ukbiobank.ac.uk/enable-your-research/apply-for-access).

We performed incident disease phenome-wide association testing in 89 573 UK Biobank participants (62±8 years of age; 56% women) who underwent 1 week of physical activity measurement with a wrist-based accelerometer between June 8, 2013, and December 30, 2015 (Figure 1). Median follow-up was 6.3 years (quartile 1: 5.7, quartile 3: 6.8), and 94.2% of participants had ≥5 years of follow-up. Stratified at the guideline-based threshold of ≥150 minutes MVPA/week, a total of 30 228 participants were in the inactive group (33.7%), 37 872 were in the weekend warrior group (≥50% of total MVPA within 1 or 2 days; 42.2%), and 21 473 were in the regular group (no 1- or 2-day period with ≥50% of total MVPA; 24.0%). Total weekly MVPA achieved was highest for the regular group (median 418 minutes, quartile 1: 302, quartile 3: 605), followed by weekend warrior (288 minutes, quartile 1: 216, quartile 3: 432) and inactive (72 minutes, quartile 1: 29, quartile 3: 115; Figure 2). Weekend warriors were substantially more active on their 2 most active days of the week versus the remaining 5, whereas activity was more evenly distributed in the regular activity group (Figure 2). Baseline characteristics of the analysis sample are displayed in Table 1.

Using multivariable Cox proportional hazards models testing 678 conditions (with ≥120 events, disease incidence rates provided in Table S2) and evaluated at a FDR of 0.01 (or 1%), both weekend warrior activity (267 significant associations [39% of all conditions tested], 264 with lower disease risk [99% of significant associations]; hazard ratio [HR] range, 0.35–0.89) and regular activity (209 associations [31%], 205 with lower disease risk [98%]; HR range, 0.41–0.88) were broadly associated with lower risk of incident disease (Figure 3; Table S3). Most of the significant associations represented circulatory (weekend warrior: 16.5% of all significant associations, regular: 14.4% of all significant associations), metabolic (14.2%, 13.9%), and digestive (12.0%, 13.4%) conditions, although associations spanned all 16 categories evaluated (Figure 3; Table S3). For both physical activity patterns, the small number of associations with higher disease risk primarily represented musculoskeletal conditions, injuries, and dermatologic conditions (eg, disorders of muscle, ligament, and fascia; Tables S3 and S4). Compared directly, there were no conditions for which risk differed significantly for weekend warrior versus regular activity at an FDR of 0.01, whether or not MVPA was included as an adjustment variable (Tables S5 and S6). Individual association test results for the weekend warrior and regular activity patterns, including individual effect sizes and E-values, are displayed in Tables S3, S5, and S6.

Across all conditions tested, associations between both weekend warrior and regular activity patterns and incident cardiometabolic conditions were particularly prominent (Figure 3). For example, among diseases with the strongest associations (ie, smallest P values), we observed similarly lower risks of incident hypertension (weekend warrior: HR, 0.77 [95% CI, 0.73–0.80], P=1.2×10^-27; regular: HR, 0.72 [95% CI, 0.68–0.77], P=4.5×10^-28), diabetes (weekend warrior: HR, 0.57 [95% CI, 0.51–0.62], P=3.9×10^-32; regular: HR, 0.54 [95% CI, 0.48–0.60], P=8.7×10^-26), obesity (weekend warrior: HR, 0.55 [95% CI, 0.50–0.60], P=2.4×10^-43, regular: HR, 0.44 [95% CI, 0.40–0.50], P=9.6×10^-47), and sleep apnea (weekend warrior: HR, 0.57 [95% CI, 0.48–0.69], P=1.6×10^-9; regular: HR, 0.49 [95% CI, 0.39–0.62], P=7.4×10^-10; Table 2; Tables S3 and S7). The 5-year cumulative risks of these 4 exemplar cardiometabolic conditions were similar for both regular activity and weekend warrior activity, and substantially lower than those observed among inactive individuals (Figure 4). The relative hazard of all 4 conditions was progressively lower with increasing quartile of MVPA (Figure S1). Results were similar when conditions were ranked by effect size, with comparably large associations with lower risk observed for cardiometabolic conditions (eg, chronic renal failure: weekend warrior: HR, 0.64 [95% CI, 0.58–0.72], P=6.8×10^-15; regular: HR, 0.65 [95% CI, 0.56–0.74], P=5.4×10^-10), as well as a range of other diseases (eg, cholelithiasis: weekend warrior: HR, 0.64 [95% CI, 0.56–0.73], P=7.2×10^-12; regular: HR, 0.57 [95% CI, 0.48–0.67], P=5.9×10^-12). The top 5 associations by disease category are shown in Table S8.

Association results were similar in analyses using the sample median (≥230.4 minutes) as the threshold for physical activity (Figure S2; Tables S9 through S11), and when events within 2 years of physical activity measurement were blanked (Tables S12 and S13). Across increasingly stringent definitions of weekend warrior activity, both the number of individuals meeting criteria for a weekend warrior pattern and the number of significant disease associations identified at FDR 0.01 decreased (≥50% of total MVPA on 1 or 2 consecutive days: n=26 032, with  218 associations; ≥50% of total MVPA on 1 or 2 weekend days: n=8002, with 79 associations; ≥75% of total MVPA on 1 or 2 days: n=6333, with 19 associations; ≥75% of total MVPA on 1 or 2 consecutive days: n=2978, with 1 association; Tables S14 through S21). Yet effect sizes for the 4 exemplar cardiometabolic diseases remained consistent, suggesting fewer associations identified at our stringent multiplicity threshold were a result of decreasing statistical power rather than changes in the effect of weekend warrior activity across varying definitions (Table 2). Using each of the alternative weekend warrior definitions tested, there were again no diseases for which effects differed significantly between weekend warrior and regular activity at an FDR of 0.01 (Table 2). Our observations about cardiometabolic conditions remained consistent when incorporating physical measurements taken at systematic assessments after physical activity measurement such as blood pressure, glycated hemoglobin, and body mass index into outcome definitions of hypertension, diabetes, and obesity (Table S22). Associations were similar in models including spline terms to account for potential nonlinearity in the effects of alcohol and MVPA (Tables S23 through S25).

Several limitations should be taken into consideration. First, this study is observational and cannot be used to infer causal relationships. Nevertheless, we adjusted for multiple potential confounders and took steps to mitigate reverse causation (eg, 2-year blanking analysis), in which our observations remained robust. Furthermore, E-values for key outcomes suggested that the effects of unmeasured confounders would need to be very large to nullify observed associations. Second, physical activity was assessed over a single week, and participants may have altered their behavior during observation. Third, our methodology relied on a previously validated MVPA classification method encompassing various activities like walking, jogging, stationary cycling, elliptical, and others,^10,11 yet the accuracy of MVPA classification may vary depending on physical activity type. Fourth, although optimal MVPA thresholds using wrist-based accelerometers remain uncertain, our results remained consistent when using the 150 minutes/week threshold endorsed in consensus guidelines, as well as the sample median threshold and alternative definitions of weekend warrior activity. Fifth, the UK Biobank is predominantly White, which may limit the generalizability of our findings to other ethnic or racial groups. Sixth, we defined diseases using Phecodes, which are a validated diagnosis code–based framework for ascertaining hundreds of conditions. We submit that use of a publicly available system for principled association testing will facilitate reproduction and comparison of our results to future analyses, and, to this end, we plan to return our Phecode-based definitions to the UK Biobank for future use. At the same time, we acknowledge that accuracy for individual conditions is likely lower than curated disease-specific definitions drawing from multiple sources, which are impractical to develop and apply across all conditions. Reassuringly, we observed similar findings for key cardiometabolic outcomes when incorporating physical measurements, medications, and laboratory results into our definitions. Seventh, although detailed hospital diagnoses were available for the whole cohort, primary care data were only available in roughly 45% of participants and only through 2017, which may limit sensitivity for certain conditions less likely to result in the generation of inpatient codes (eg, obesity). It is important to note that incidence rates for other major cardiometabolic conditions (eg, hypertension or diabetes) appeared comparable with estimates from other relatively healthy research samples,^23–26 and results remained similar when additional data (ie, body mass index) were used to identify obesity. Eighth, follow-up time is limited in duration, and therefore, the long-term impact of habitual exercise on health outcomes is incompletely elucidated. Ninth, we focused on quantifying disease associations. Future work is warranted to assess the potential value of measured activity pattern within a prediction model framework.

In a cohort of nearly 90 000 individuals with wrist-based physical activity tracking, we found that both physical activity concentrated within 1 or 2 days and more regular activity patterns were each associated with a similarly lower risk of >200 diseases spanning the full spectrum of disease. Both physical activity patterns had particularly prominent associations with lower risk of cardiometabolic conditions. Future studies are warranted to assess the potential value of concentrated physical activity interventions to improve public health.

Supplemental Methods

Figures S1–S2

Tables S1–S25

References 27–32