Accelerometer-derived moderate-to-vigorous physical activity and incident nonalcoholic fatty liver disease | BMC Medicine

Study design and participants

As previously described [15], the UK Biobank is a large, observational, population-based cohort recruiting half a million adult residents, aged 37–73 years, from 1 of 22 assessment centers across the UK (England, Wales, and Scotland) between 2006 and 2010. At baseline, participants were asked to complete a comprehensive questionnaire assessing sociodemographic, lifestyle, and health-related information, receive physical examinations, and provide biological samples. The UK Biobank was approved by the North West Research Ethics Committee (11/NW/0382) and all participants signed an informed consent.

The current analysis was restricted to a sub-sample of 103,661 participants who responded to emails for the accelerometer sub-study between 2013 and 2015. Individuals with insufficient accelerometer data quality or less than a full week of available acceleration data (n = 12,175) were excluded. Additionally, we excluded participants who had NAFLD or severe liver diseases or other liver diseases or alcohol/drug use disorders at/before the time of accelerometer measurements (n = 2829), resulting in a final analysis of 88,656 participants (Additional file 1: Figure S1).

Exposure assessment

Between February 2013 and December 2015 (Additional file 1: Figure S2), participants who provided a valid email address to UK Biobank were invited at random to wear a wrist-worn accelerometer (Axivity AX3). Participants were instructed to wear the device on their dominant wrist continuously for 1 week while continuing with their usual activities. The accelerometer captures triaxial acceleration data over 7 days at 100 Hz with a dynamic range of ± 8 gravity. Proportions of time spent across sleep (non-awake behavior), sedentary behavior (SB) (awake behavior at ≤ 1.5 metabolic equivalent of task (METs), such as driving or watching television), light physical activity (LPA) (awake behavior at < 3 METs not meeting the sedentary behavior definition, such as cooking or self-care), and MVPA (awake behavior at ≥ 3 METs, such as walking the dog or jogging) per day were identified from raw accelerometer data using a previously published random forest and hidden Markov model machine-learning methods that were trained using wearable cameras and time-use diaries among 152 individuals in free-living conditions [16]. Briefly, accelerometer data was annotated with activities from the Compendium of Physical Activities. [17] A balanced Random Forest (RF) with 100 decision trees was trained to classify the behavior in 30-s time windows using 50 rotation-invariant time and frequency domain features of the accelerometer signal. Then, a hidden Markov model was employed to use time sequence information to improve the RF-assigned label sequence.

Because the optimal level of accelerometer-measured MVPA for prevention of incident NAFLD is unknown [18], in the primary analyses, based on the guideline-based threshold (≥ 150 min/week) [6], individuals were classified as active WW (equal to or above the MVPA threshold and ≥ 50% of total MVPA achieved within 1–2 days) [13], regularly active (equal to or above MVPA threshold but not active WW), and inactive (below MVPA threshold) MVPA patterns.

Covariates assessment

Detailed information on covariates at baseline was available through standardized questionnaires at baseline, including age, sex, ethnicities, Townsend deprivation index (TDI), income, education levels, employment, smoking, and alcohol drinking (Additional file 1: Figure S2). Body mass index (BMI) was measured and calculated as weight (in kilograms (kg)) divided by the square of height (in meters (m)).

Study outcome assessment

The primary outcome was incident NAFLD [19], ascertained through links to hospital inpatient data and death register records. In accordance with the Expert Panel Consensus statement, NAFLD (including non-alcoholic steatohepatitis (NASH)) was identified as ICD-10 K76.0, K75.8, and ICD-9 571.8. Hospital admissions data were available until September 30, 2021, for centers in England, July 31, 2021, for centers in Scotland, and February 28, 2018, for centers in Wales, and mortality data were available until October 2021(Additional file 1: Figure S2). The follow-up person-time for each participant was calculated from the final date of accelerometer wear until the date of death, the first date of outcome diagnosis, the date of loss to follow-up, or the end of follow-up, whichever came first.

Since NAFLD is the most important cause of cirrhotic complications, hepatocellular carcinoma, and liver-related mortality [20], to avoid missing NAFLD events that can lead to adverse liver outcomes, we used incident severe liver diseases (a composite of liver cirrhosis, liver failure, hepatocellular carcinoma, and liver-related death) and incident liver cirrhosis as secondary outcomes (Additional file 1: Table S1).

Moreover, liver magnetic resonance imaging (MRI) was performed between January 2016 and February 2020 in the UK Biobank imaging sub-study (Additional file 1: Figure S2), and the proton density fat fraction (PDFF) and iron-corrected T1 mapping (cT1) was extracted as a measurement of liver steatosis and liver fibrosis, respectively [21]. In this sub-study, liver steatosis, defined as PDFF ≥ 5.5% [22], and liver fibrosis, defined as cT1 ≥ 800 ms (ms) [21], were also identified as secondary outcomes to capture undiagnosed relatively mild cases of chronic liver disease. In the current sub-analysis, 15,455 participants had available data of PDFF, while 12,393 participants had available data of cT1 (Additional file 1: Figure S1).

Statistical analysis

Population characteristics were presented as mean (SD) for continuous variables or proportions for categorical variables. Difference of characteristics according to incident NAFLD (yes vs. no) were tested by t-tests and chi-square tests for continuous and categorical variables, respectively.

To test the actual association of MVPA in relation to other movement behaviors, a compositional data analysis (CODA) approach was used [16]. For CODA, the activity composition was created by expressing the time spent on each activity (i.e., MVPA, sleep, SB, and LPA) as a proportion of a 24-h day. The activity composition was then expressed as isometric log-ratio (ilr) coordinates to account for the interdependency of the activity domains. Then, Cox proportional hazards regression models estimating survival were built using the corresponding set of three ilr coordinates for MVPA.

Restricted cubic spline (RCS) Cox regression was performed to test for linearity and explore the shape of the dose–response relationship of total MVPA with incident NAFLD. A two-piecewise Cox regression model was used to examine the threshold effect of total MVPA on incident NAFLD using a smoothing function. The inflection point (i.e., threshold) was determined using the likelihood-ratio test and bootstrap resampling methods. Cox proportional hazards models were used to estimate the relationship of total MVPA or MVPA patterns with study outcomes, except for liver MRI-related outcomes which was estimated using binomial regression models. In multivariable models, several potential confounders were controlled for, including demographics (age, sex, ethnicities, recruitment center, TDI, educational attainment, household income, employment), anthropometric and lifestyle factors (smoking status, alcohol consumption, and BMI), and the total time and season of accelerometer wear. The proportional hazards assumption was checked using the Schoenfeld residuals, and no violation was found. Percentages of missing values of covariates were less than 1% except for income (10.3%). Missing data were coded as a missing indicator category for categorical variables and with mean values for continuous variables.

To test the robustness of our findings, several sensitivity analyses were also performed for the association between MVPA patterns and primary outcome. Firstly, the MVPA patterns was defined using threshold derived from the threshold effect analyses (≥ 208 min/week) at which the rapid decline in NAFLD risk lessened or leveled off as MVPA increased. Secondly, we assessed alternative definitions of the WW pattern, including ≥ 50% of total MVPA over 1–2 consecutive days and ≥ 50% of total MVPA over 1–2 weekend days. Third, all participants within 2 years of follow-up were excluded to minimize reverse causation. Fourth, participants with missing covariates were excluded. Fifth, we further adjusted for pre-existing hypertension and diabetes, defined as baseline self-reported medical history or health records taken before the time of accelerometer measurement, and healthy diet scores, evaluated using a more recent dietary recommendation for cardiovascular health which considered adequate consumption of fruits, vegetables, whole grains, fish, shellfish, dairy products, and vegetable oils, and reduced consumption of refined grains, processed meats, unprocessed meats, and sugar sweetened beverages. Sixth, we further limited the main analysis to participants with low-to-intermediate predicted NAFLD risk as estimated by the Dallas Steatosis Index (DSI). DSI is a superior tool to predict NAFLD as inferred using MR spectroscopy. Based on DSI, NAFLD can be excluded with a negative predictive value of 80% at a threshold of < 50% risk [23, 24]. Seventh, as NAFLD is an important cardiovascular risk factor, we also assessed the association between MVPA pattern and cardiovascular disease.

As additional exploratory analyses, possible modifications of the association of MVPA patterns with incident NAFLD were also assessed for the following variables: age (< 60 or ≥ 60 years), sex (females or males), BMI (< 25 or ≥ 25 kg/m²), smoking status (never or ever), and alcohol drinking (< 1 or ≥ 1 times/week).

A two-tailed P < 0.05 was considered to be statistically significant in all analyses. Analyses were performed using R 4.1.1 software (