Accelerometry‐Derived Activity and Sleep Patterns in the NIH All of Us Cohort: Insights and Predictive Potential for Inflammatory Arthritis

Abstract

Objective

The relationship between physical activity and sleep with inflammatory arthritis (IA) is understudied, and existing research has relied largely on self‐report or short‐term assessments. The NIH All of Us database provides long‐term accelerometry data, enabling more precise estimation of the association between lifestyle behaviors and IA.

Methods

Participants from the All of Us database who shared electronic health record and Fitbit data were included. Daily activity and sleep metrics were compared between individuals with and without IA using multiple linear regression. Cox proportional hazards regression was used to examine the association of activity and sleep patterns with incident IA in a 10‐year follow‐up period.

Results

A total of 23,855 participants were included, 200 of whom had IA. Participants with IA took fewer daily steps (P < 0.001) and had greater sleep variability (P < 0.001) compared to those without IA. 122 individuals had incident IA. Every 1,000 extra daily steps were associated with a 7% lower risk of IA (hazard ratio [HR] 0.93 [95% confidence interval (CI) 0.87–0.99], P = 0.02). Compared to those who walked <5,000 steps daily, those who walked 5,000 to 10,000 steps and 10,000+ steps had a 41% (HR 0.59 [95% CI 0.39–0.91], P = 0.02) and 50% (HR 0.50 [95% CI 0.29–0.86], P = 0.01) reduction in risk of IA.

Conclusion

Individuals with IA had reduced step counts and more sleep variability compared to those without IA, highlighting how physical activity and sleep contribute to IA. Additionally, increased daily step count was associated with decreased risk of incident IA, suggesting a possible research intervention for those at high risk of IA.

INTRODUCTION

Rheumatoid arthritis (RA) and psoriatic arthritis (PsA) are immune‐mediated disorders characterized by articular pain, swelling, and stiffness. If left untreated, these types of inflammatory arthritis (IA) can lead to permanent joint damage and decreased function. IA negatively affects lifestyle behaviors, with patients showing decreased physical activity and higher sedentariness. ¹ Although these behaviors were traditionally attributed to disease symptoms, emerging evidence suggests a bidirectional relationship. For example, increased physical activity and reduced sedentary time can mitigate arthritis symptoms such as pain and fatigue, ² whereas sedentary behavior is linked to increased inflammation in the general population. ³ Highlighting this bidirectional relationship, the 2022 American College of Rheumatology guidelines on RA recommend exercise as an adjuvant therapy to immunomodulators for RA. ⁴ Furthermore, sleep issues, particularly poor sleep quality, are also common among patients with IA. ⁵ However, research using objective data from wearables to assess sleep patterns remains scarce.

Understanding physical activity and sleep in IA, and its potential to intercept or modify disease course, is crucial. The National Institutes of Health (NIH) All of Us Research Program (AoU) offers a unique opportunity to address this knowledge gap, combining electronic health record (EHR) data with wearable accelerometry (Fitbit) data from up to 60,000 participants. Using this data set, we aimed to characterize physical activity and sleep patterns in IA and examine their ability to identify early IA or even predict IA incidence.

PATIENTS AND METHODS

Study population

This study used data from the AoU research program, which enrolls individuals aged 18 and older at various US clinical centers. The AoU Institutional Review Board approves all program aspects, and the research protocol has been published previously. ⁶ Participants enroll digitally via the program's website or smart phone application, complete the electronic consent process to share EHR data, and fill out baseline health surveys. Participants who own a Fitbit can choose to share their data with the AoU program. This study used the AoU Controlled Tier Data set version 8. Data analyzed in this study included day‐level Fitbit information, demographics, physical measurements, enrollment surveys, and EHR data.

IA ascertainment

IA was defined using at least one International Classification of Diseases, Tenth Revision (ICD‐10) codes for RA (M06.9) or PsA (L40.5). Participants were categorized as having prevalent IA if they had an IA diagnosis on or before the last day of Fitbit data collection. Incident IA was defined as a new diagnosis of RA or PsA by ICD‐10 code occurring after the Fitbit data period, up to 10 years of follow‐up captured in EHR.

Accelerometry data

Fitbit‐derived metrics included: average daily step count, weekly moderate‐to‐vigorous physical activity (MVPA) (using “fairly active” and “very active” minutes), daily sedentary time (defined by minutes classified as inactive by Fitbit), and nightly sleep duration calculated only for Fitbit‐designated primary sleep episodes (excluding naps). Analyses were performed only on valid days of wear. For activity metrics, a valid day was defined as a day with at least 100 steps recorded. For sleep metrics, days with sleep >1,440 minutes and participants with <4 hours of sleep on >30% of days were excluded. Activity and sleep metrics were averaged across their respective valid day sets over the one‐year wear period.

Statistical analysis

Cross‐sectional analysis was used to compare accelerometry‐derived physical activity and sleep metrics between people with and without IA. Baseline characteristics were described using means and SDs. Univariate comparisons using a t‐test for continuous variables and chi‐square tests for categorical variables were performed. We then performed multivariate linear regression analysis for each accelerometry metric as outcome, adjusted for the demographic and clinical covariates noted earlier, including history of IA as a binary (yes or no) variable.

To examine the association between accelerometry metrics and risk of incident IA, we performed a Cox proportional hazards regression analysis separately for each accelerometry metric. Each Cox regression model was adjusted for age at Fitbit start (continuous), sex, race and ethnicity, body mass index (BMI), hypertension, and diabetes. Hypertension and diabetes were defined by ICD‐10 code and/or use of antihypertensive and antidiabetes medications, respectively. ⁷ The time to event outcome was computed as the number of days between the start of follow‐up (the day after one year of Fitbit wear) and the date of incidence of IA ascertained from the EHR. Participants without IA were censored at last EHR contact in the 10‐year follow‐up period.

All statistical analyses were performed in Python (statsmodels), and P < 0.05 was considered statistically significant.

RESULTS

Prevalent IA

Participant characteristics

The final cohort included 23,855 participants (mean ± SD age 52 ± 15 years; 70% female; 79% White; Table 1). A total of 200 participants had prevalent IA (82% female and 85% White). Compared to those without IA, participants with IA were significantly older (mean ± SD 58 ± 12 years vs 52 ± 15 years; P < 0.001) and had a higher prevalence of hypertension (73% vs 30%; P < 0.001) and diabetes (40% vs 13%; P < 0.001) in univariate analysis. BMI was also higher in the IA group (mean ± SD 31 ± 7 vs 29 ± 5; P < 0.001).

Table 1.

Baseline characteristics of patients with and without history of IA^*

Characteristic	IA, n = 200	No IA, n = 23,655	P value
Age, mean (SD), y	58 (12)	52 (15)	<0.001
Female, n (%)	163 (82)	16,480 (70)	<0.001
White, n (%)	170 (85)	18,684 (79)	0.046
Body mass index, mean (SD)	31 (7)	29 (5)	<0.001
Hypertension, n (%)	146 (73)	7,168 (30)	<0.001
Diabetes, n (%)	80 (40)	3,097 (13)	<0.001

^*IA, inflammatory arthritis.

Accelerometry metrics

Participants with IA had a mean ± SD 317 ± 53 usable days of data, similar to those without IA (320 ± 55 days). In univariate analysis, participants with IA had lower mean levels of activity compared to those without, with fewer daily steps (mean ± SD 6,703 ± 3,188 vs 7,733 ± 3,339; P < 0.001), fewer weekly minutes of MVPA (mean ± SD 202 ± 199 minutes vs 241 ± 206 minutes; P = 0.01), and greater daily sedentary time (mean ± SD 14.7 ± 3.6 hours vs 14.1 ± 3.3 hours; P = 0.04). Although sleep duration was not significantly different between those with and without IA (mean ± SD 397 ± 56 minutes vs 400 ± 51 minutes, respectively; P = 0.25), sleep variability was higher in the participants with IA, with a mean ± SD coefficient of variation of 25.1% ± 9.4% compared to 23.8% ± 8.9% in those without IA (P = 0.005).

We then ran multiple linear regression models with accelerometry metrics as outcome and IA as our primary exposure adjusted for age, self‐reported sex, race, and ethnicity (Supplementary Table 1). Participants with IA averaged 808 (95% confidence interval [CI] 350–1,270) daily steps lower than those without IA (P < 0.001). In subgroup analysis separating individuals with RA and PsA, individuals with RA had 949 (95% CI 400–1,490, P < 0.001) fewer daily steps than those without IA, whereas those with PsA had similar steps compared to those with IA. Sedentary time and MVPA were not significantly different between individuals with and without IA, although duration of weekly vigorous activity was lower in those with IA by 18.2 (95% CI 0.0–37.8, P = 0.040) minutes compared to those without IA. Among sleep metrics, sleep duration was not different by IA status; however, sleep variability was still significantly higher in people with IA by 1.6 (95% CI 0.8–2.5, P < 0.001) percentage points in adjusted analysis (Supplementary Table 1).

Incident IA

Participant characteristics

We analyzed three separate cohorts for steps, sedentary time and MVPA, and sleep, as each of these three had separate preprocessing and filtering criteria as outlined in the Patients and methods section. The steps cohort had 21,044 participants with usable Fitbit step count data and follow‐up data from their EHR without a previous diagnosis of IA. Among these participants, 122 participants had incident IA during the 10‐year follow‐up period. The breakdown of incident IA was 88 new RA cases (72%) and 34 new PsA cases (28%). Participant characteristics for this cohort are documented in Table 2. In unadjusted analysis, the participants who developed IA compared to those who did not develop IA had a larger proportion of females (84% vs 70%; P = 0.002), higher BMI (mean ± SD 31 ± 6 vs 29 ± 5; P = 0.001), and higher rates of hypertension (44% vs 31%; P = 0.002) and diabetes (24% vs 13%; P = 0.001). The participant characteristics of the sedentary time or MVPA and sleep cohorts were similar to the steps cohort.

Table 2.

Baseline characteristics for incidence cohort^*

Characteristic	Incident IA, n = 122	No incident IA, n = 20,922	P value
Age, mean ± SD, y	53 ± 13	52 ± 15	0.510
Female, n (%)	102 (84)	14,662 (70)	0.002
White, n (%)	92 (75)	16,776 (80)	0.046
Body mass index, mean (SD)	31 (6)	29 (5)	0.001
Hypertension, n (%)	54 (44)	6,509 (31)	0.002
Diabetes, n (%)	29 (24)	2,785 (13)	0.001

^*IA, inflammatory arthritis.

Incidence by step count

Using a Cox regression adjusted for demographic characteristics, every additional 1,000 daily steps was associated with a 7% lower risk of developing IA (HR 0.93 [95% CI 0.87–0.99], P = 0.02) (Figure 1). Male sex had a 53% lower risk of IA (HR 0.47 [95% CI 0.29–0.77], P = 0.002); no other demographic characteristics were associated with IA. Evaluating steps as a categorical variable, compared to daily step count of <5,000 (n = 4,068; n_IA = 31 [0.8%]), participants who recorded 5,000 to 10,000 steps (n = 12,119; n_IA = 69 [0.6%]) and 10,000+ steps (n = 4,857; n_IA = 22 [0.5%]) per day had 41% (HR 0.59 [95% CI 0.39–0.91], P = 0.02) and 50% (HR 0.50 [95% CI 0.29–0.86], P = 0.01) lower risk of developing IA. Notably, when we included BMI, history of hypertension, diabetes, and smoking into our Cox models, daily step count was no longer associated with incident IA (Supplementary Table 2). BMI (per 5 kg/m²; HR 1.04 [95% CI 1.01–1.08], P = 0.03) and history of hypertension (HR 1.76 [95% CI 1.18–2.63], P = 0.006) were both associated with incident IA.

Figure 1.

Step count hazard ratios for inflammatory arthritis incidence. Cox regression model adjusted for age, sex, race, and ethnicity. Step count per 1,000 steps is continuous. *Step count 5,000 to 10,000 and 10,000+ are compared to the reference of less than 5,000 steps per day. CI, confidence interval.

Incidence by sedentary time and MVPA

62 of 15,719 participants in the sedentary time or MVPA cohort developed IA. Continuous‐valued weekly MVPA (per 60 min/wk; HR 0.94 [95% CI 0.86–1.03], P = 0.18) was not associated with risk of IA (Supplementary Figure 1). However, in categorical analysis for MVPA, participants who got more than the World Health Organization's extended recommendation of 300 minutes of weekly MVPA (n = 4,611; n_IA < 20) had a 65% lower risk of IA compared to those who got fewer than 30 minutes of weekly MVPA (n = 1,048; n_IA < 20) (HR 0.35 [95% CI 0.13–0.94], P = 0.04). This was not significant when including other comorbidities into the model (Supplementary Figure 2). Sedentary time did not have a statistically significant association with risk of IA (HR 1.06 [95% CI 0.99–1.14], P = 0.08) (Supplementary Figure 1).

Incidence by sleep metrics

In the sleep cohort, 125 of 21,230 individuals developed IA. Neither sleep duration nor sleep variability were associated with incident IA (Supplementary Figure 1).

Sensitivity analyses

To address the possibility that incident IA cases may have preceded ICD coding, we performed sensitivity analysis defining our follow‐up period to begin six months after the 1‐year Fitbit assessment period. The results are only shown for those accelerometry metrics that were significantly predictive of IA risk in the previous analyses. For the steps cohort, this reduced the cohort to 19,104 participants, of whom 104 had incident IA. Continuous‐valued daily steps trended toward decreased risk but was no longer significantly associated with incident IA (HR 0.95 [95% CI 0.88–1.01], P = 0.10). However, getting 5,000 to 10,000 steps (HR 0.59 [95% CI 0.37–0.95], P = 0.03) and 10,000+ steps (HR 0.52 [95% CI 0.29–0.95], P = 0.03) per day were associated with similar protective effects against future IA risk as the original analysis. For the MVPA cohort, our population was reduced to 13,794 participants with 52 developing IA. Achieving 300+ minutes of weekly MVPA again trended toward protection but was no longer significantly associated with IA (HR 0.52 [95% CI 0.16–1.75], P = 0.29).

DISCUSSION

To our knowledge, this study using data from the NIH AoU database presents the longest duration of accelerometry data of patients with IA in the literature (average of 46 weeks). We identified a 0.7% combined prevalence of RA and PsA. Patients with IA had lower daily step counts, reduced weekly MVPA, and increased sedentary time, consistent with findings using UK Biobank accelerometry data. ¹ Importantly, however, participants with IA were also significantly older than those without IA and had very high rates of hypertension (73%) and diabetes (40%), all of which may contribute to these activity differences. Monitoring physical activity patterns in IA is crucial as reduced activity can diminish function and quality of life. Additionally, although there is no established threshold for daily steps required, studies have shown that increased steps per day up to 10,000 decreases all cause, cancer, and cardiovascular mortality, ⁸ with benefits accruing even at 7,000 daily steps for certain conditions. ⁹ Therefore, decreased step count (and increase in sedentary behavior) may compound pre‐existing cardiovascular risk and subsequently, increasing physical activity may be a vital intervention in those with IA.

Among 122 participants who developed IA after 1‐year Fitbit data collection, higher step counts were linked to lower IA risk in a dose‐dependent manner, with each additional 1,000 steps per day reducing incidence by 7%. Similar to our findings, a previous study using 7‐day accelerometry data from the UK Biobank found a 5% reduction in RA incidence with every additional 1,000 steps. ¹⁰ However, we need to interpret this data cautiously. Although our overall cohort is large, the number of incident cases is comparatively small leading to possible overestimation of effects. Additionally, patients with incident IA had higher BMIs and high rates of hypertension and diabetes, and when these factors were considered, daily step count was no longer significantly associated with incident IA. This may be due to the multidirectional relationship between cardiometabolic disease, physical activity, and IA. Obesity itself, for example, is linked to higher risk of IA development as well as decreased step count. ¹¹ , ¹² Therefore, it may be difficult to fully parse out the contribution of each factor individually. MVPA and sedentary time showed no association with IA risk, the small sample size (62 participants) limits conclusions, warranting further study.

A key limitation of this study is the use of one ICD‐10 code to define IA diagnosis and timing, which could not be externally verified. This is particularly important for incidence analysis. Given the low number of new cases, misclassification could have a substantial effect on the results and patients may have undocumented pre‐existing IA. To mitigate this, we performed a sensitivity analysis including only participants diagnosed with IA at least six months after data collection ended, which attenuated the strength of certain associations. Six months may not be a sufficient lag time to completely eliminate reverse causation, and the attenuation suggests some participants may have already had IA during data collection or that step counts decline sharply just before diagnosis, providing opportunities for a wearable‐based early IA detection framework rather than a predictor of disease. Nonetheless, the findings demonstrate that accelerometry data can help identify prodromal or early features of IA months to years before diagnosis.

Sleep measures were less revealing than physical activity. Although sleep duration did not differ significantly between groups, participants with IA showed higher sleep variability, which has been linked to conditions such as hyperlipidemia, obesity, and depression. ¹³ Therefore, addressing sleep variability rather than just hours of sleep, may be a possible target for intervention in IA, although the clinical impact of this is unclear. We used the coefficient of variation to standardize sleep variability across different mean sleep durations, however this metric can be sensitive to outliers. Future studies should consider complementary measures, such as within‐person SD or night‐to‐night successive differences, to further validate these associations. Although neither sleep duration nor variability were predictive of IA development, there are several measurement limitations. For example, factors such as sleep efficiency were not captured. Importantly, the validity of Fitbits for sleep monitoring is not as established as activity measures. With advances in sensors and algorithms to improve sleep monitoring, the role of sleep in IA may need to be re‐assessed and all findings confirmed in future cohorts.

Although these findings are thought‐provoking and hypothesis generating, this study has other limitations in addition to reliance on ICD‐10 codes. First, data on variables that may affect physical activity, such as medication use, were not available. Second, the IA population had very high levels of hypertension and diabetes. However, previous studies have also found significantly elevated rates of these conditions in older adults with RA. ¹⁴ Third, many participants in the AoU Fitbit cohort already owned a Fitbit before enrollment, ¹⁵ suggesting potential self‐selection bias favoring participants with IA with higher digital literacy and technologic familiarity. The AoU Wearables Enhancing All of Us Research (WEAR) initiative, which provides Fitbits to expand participant representation, may improve the generalizability of these findings in the future. ¹⁵ Consequently, because bring‐your‐own‐device and WEAR participants were pooled in this analysis to maintain statistical power, we cannot rule out residual confounding related to socioeconomic status or digital literacy, which may differ between these acquisition pathways. Fourth, not all possible confounders have been addressed in our analysis such as depression, anxiety, fibromyalgia, or osteoarthritis which may contribute to sleep and activity differences. Fifth, our criteria for a “valid day” of activity was based on a minimum step count rather than a formal wear‐time algorithm. Future studies leveraging heart rate data or intraday minute‐level metrics may provide more precise wear‐time validation. Additionally, given our large overall sample size, small differences may be statistically significant, and close attention needs to be paid to the clinical significance of all findings.

In a large, US‐based cohort study, we show the reciprocal relationship between objectively measured physical activity using accelerometry and IA: although IA was associated with reduced physical activity, inactivity may also contribute to a higher risk of IA. Wearable accelerometers could serve as both a monitoring tool and an intervention, helping patients increase activity through real‐time feedback that reinforces behavioral changes. ¹⁶ Future research with well‐defined, high‐risk prospective cohorts (eg, individuals with RA‐related autoantibodies or patients with psoriasis and arthralgia in PsA) is needed to evaluate accelerometry's role in predicting or as a possible biomarker for early IA detection. Overall, our findings emphasize the potential of wearable technology to advance IA research and care, not only for remote management but also for early detection of IA.

AUTHOR CONTRIBUTIONS

All authors contributed to at least one of the following manuscript preparation roles: conceptualization AND/OR methodology, software, investigation, formal analysis, data curation, visualization, and validation AND drafting or reviewing/editing the final draft. As corresponding authors, Drs Barua and Haberman confirm that all authors have provided the final approval of the version to be published and take responsibility for the affirmations regarding article submission (eg, not under consideration by another journal), the integrity of the data presented, and the statements regarding compliance with institutional review board/Declaration of Helsinki requirements.

Supporting information

Disclosure Form:

Data S1 Supporting Information