Calibration of an Accelerometer Activity Index among Older Women and Its Association with Cardiometabolic Risk Factors

Abstract

Purpose:

Traditional summary metrics provided by accelerometer device manufacturers, known as counts, are proprietary and manufacturer specific, making them difficult to compare studies using different devices. Alternative summary metrics based on raw accelerometry data have been introduced in recent years. However, they were often not calibrated on ground truth measures of activity-related energy expenditure for direct translation into continuous activity intensity levels. Our purpose is to calibrate, derive, and validate thresholds among women 60 years and older based on a recently proposed transparent raw data based accelerometer activity index (AAI), and to demonstrate its application in association with cardiometabolic risk factors.

Methods:

We first built calibration equations for estimating metabolic equivalents (METs) continuously using AAI and personal characteristics using internal calibration data (n=199). We then derived AAI cutpoints to classify epochs into sedentary behavior and intensity categories. The AAI cutpoints were applied to 4,655 data units in the main study. We then utilized linear models to investigate associations of AAI sedentary behavior and physical activity intensity with cardiometabolic risk factors.

Results:

We found that AAI demonstrated great predictive accuracy for METs (R²=0.74). AAI-based physical activity measures were associated in the expected directions with body mass index (BMI), blood glucose, and high density lipoprotein (HDL) cholesterol.

Conclusion:

The calibration framework for AAI and the cutpoints derived for women older than 60 years can be applied to ongoing epidemiologic studies to more accurately define sedentary behavior and physical activity intensity exposures which could improve accuracy of estimated associations with health outcomes.

Introduction

Accelerometers have been widely used to measure physical activity (PA) in biomedical studies in the past decades (Yang and Hsu 2010, Bai, Di et al. 2016). Traditionally, the common data output consists of summary measures over user-defined epochs (e.g. 1 minute) (Chen, Janz et al. 2012). For example, both ActiGraph (ActiGraph, Pensacola, FL) and Actical (Phillips Respironics, Bend, OR) software use proprietary algorithms to calculate “counts” per epoch (John and Freedson 2012), which is an aggregated measure of the acceleration magnitude over a given time epoch. Though easy to use, the proprietary nature of counts makes comparison between studies using different devices difficult because the counts defined by different manufactures can have very different meanings (Rowlands 2018).

As technology advances in recent years, several research grade accelerometers allow direct measurement and access to high resolution raw acceleration data (10–100 Hz), which contain much richer information than proprietary counts on epoch-level resolutions. This enables researchers to use well-defined, open-source, reproducible summaries of the data to compare and combine studies that collect raw accelerometry data (Karas, Bai et al. 2019). To overcome limitations of counts, efforts have been made to establish summary metrics from the raw data with transparent formulas. A few raw accelerometry data-based metrics have been proposed, such as an accelerometer activity index (AAI) (Bai, Di et al. 2016), Euclidean norm minus one (ENMO) (van Hees, Gorzelniak et al. 2013), mean amplitude deviation (MAD) (Aittasalo, Vaha-Ypya et al. 2015), and a monitor-independent movement summary (MIMS) (John, Tang et al. 2019). In general, all these newly proposed metrics aim to quantify and summarize the magnitude of acceleration during a given epoch. For these metrics to be useful and interpretable, it is important to calibrate them against ground truth PA energy expenditure (e.g, intensity of PA). The gold standard measure of PA energy expenditure, or intensity, is oxygen uptake directly measured using calorimetry procedures during performance of sedentary behaviors (SB) and PA tasks (LaMonte, Ainsworth et al. 2006). Calibration studies are useful in translating PA metrics into intensity-specific categories relevant to specified populations (Jago, Zakeri et al. 2007) and because for some health outcomes PA intensity may be a more relevant exposure than the amount of PA performed (Kesaniemi, Danforth Jr. et al. 2001). To the best of our knowledge, little existing work on calibration of raw accelerometry data-based metrics exists, especially among older adults. It is particularly important to calibrate the intensity-specific cutpoints among adults 60 years and older because the cutpoints for this group may differ dramatically from those derived in samples of younger adults (Evenson, Wen et al. 2015) given the increase in absolute energy cost of movement associated with aging (Ortega and Farley 2007).

We conducted a novel analysis based on an AAI, a recently proposed transparent metric to summarize high-dimensional raw accelerometry data, in women 60 years and older. Calibration equations were derived for METs continuously using AAI and AAI cutpoints were obtained to classify epochs into SB and distinct PA intensity categories, by using data from a calibration study on older women. We then utilized the calibrated AAI to investigate associations of SB and PA measures with cardiometabolic risk factors. To investigate simultaneous associations of multiple intensity categories, we also adopted isotemporal models to quantify the potential substitutional effects of reallocating time between two activity intensity categories on cardiometabolic risk factors.

Materials and Methods

Study Design and Participants

The Women’s Health Initiative (WHI) is an ongoing study of the determinants of morbidity and mortality in aging postmenopausal women. WHI enrolled 161,808 postmenopausal women, aged 50–79 years, into one of three randomized clinical trials or a prospective observational study across 40 U.S. study sites from 1993 to 1998 (Anderson, Manson et al. 2003). During the WHI 2010–2015 extension study, home examinations were completed in a sub-cohort of 7,875 women also enrolled into the ancillary Long Life Study (2012–2013). The Objective Physical Activity and Cardiovascular Health (OPACH) Study (LaCroix, Rillamas-Sun et al. 2017) was ancillary to the Long Life Study and consisted of ambulatory community-dwelling women aged 63–99 years. Women were recruited to wear an ActiGraph GT3X+ accelerometer to measure free-living PA and SB from May 2012 to April 2014, and relate these measures to cardiovascular disease incidence during follow-up. Overall, 7,048 women were sent the accelerometer and a sleep log, among whom 6,721 (95.4%) women returned accelerometers and 6,489 (92.0%) women had data for at least one day. Among them, 6,078 women had valid raw data for computing AAI, and 5,870 had data for at least 4 adherent day using the common definition of at least 10 hours of accelerometer wear time while awake. Study protocols were approved by the Fred Hutchinson Cancer Research Center (WHI Coordinating Center) and all women provided informed consent in writing or by phone.

Free-living Accelerometry Data Collection

OPACH women wore an ActiGraph GT3X+ triaxial accelerometer over the right hip 24 hours-per-day for up to 7 consecutive days except when they risked submerging the accelerometer in water (e.g., during bathing or swimming). Sleep time was identified using self-reported in-bed and out-of-bed times from sleep diaries that were filled out each night of accelerometer wear (Rillamas-Sun, Buchner et al. 2015). If their sleep log data were missing, their in-bed and out-of-bed times were imputed using person-specific means, if available, or the mean over the full OPACH sample. ActiLife version 6 software was used to process raw data (30 Hz) into counts per 15-seconds epochs. Vector magnitude (VM) counts were derived by taking the square root of the sum of the three axes squared to capture movement in all three axes. Non-wear periods were identified using the Choi algorithm on VM counts (Choi, Liu et al. 2011, Choi, Ward et al. 2012). We then applied count-based intensity-specific cutpoints that were determined using the same sample as the present study (Evenson, Wen et al. 2015), i.e., moderate-to-vigorous PA (MVPA; >518 counts per 15-seconds), high light PA (HLPA; >225 counts per 15-seconds and <=518 counts per 15-seconds), low light PA (LLPA; >18 counts per 15-seconds and <=225 counts per 15-seconds), and SB (<=18 counts per 15-seconds).

Health outcomes

Questionnaires ascertained participant age, race and ethnicity (Black, White, Hispanic/Latina), and educational attainment (high school equivalent or lower, some college, college graduate). During the Long Life Study home examinations, trained study staff measured height (m) and weight (kg) with a tape measure and calibrated scale, respectively, and calculated BMI (kg/m²). Fasting (12 hours) blood samples were obtained and cardiometabolic risk factors including serum glucose and HDL were measured using standardized Clinical Laboratory Improvement Act-approved methods at the University of Minnesota (LaMonte, Lewis et al. 2017).

Calibration Study

The OPACH study included a calibration sub-study, where 200 women aged 60 to 91 years old were invited to participate in one laboratory session to calibrate accelerometry to energy expenditure during SBs and various PA tasks differing in known energy costs. The participants were asked to perform several standardized tasks while simultaneously wearing a hip-worn accelerometer, a wrist-worn accelerometer, a heart rate monitor, and a portable indirect calorimeter to measure oxygen uptake (VO₂). The VO₂ measures were expressed in METs which represent multiples of metabolic energy cost defined as the ratio of activity-related energy expenditure to resting energy expenditure (LaMonte, Ainsworth et al. 2006). A portable breath-by-breath metabolic device, Oxycon Mobile (CareFusion, Rolle, Switzerland), was used to measure (VO₂) and a chest-worn POLAR heart rate monitor (Polar Electro, Lake Success, NY, USA) was used to measure heart rate continuously during the PA tasks. The intensity of selected tasks of the calibration study varied from sedentary (<1.5 METs) to light intensity (1.5 – 3.0 METs) and moderate intensity (3.0–6.0 METs) for older women. Participants were asked to rest 2 minutes between activities to allow their heart rates to return to within 10 beats/minute of their resting heart rate. The duration of tasks was seven minutes, except for the usual pace 400-meter walk, so that participants achieved steady rate metabolism for measurement of task-specific VO₂. The participants performed the following tasks: watching DVD while sitting quietly (DVD), wash/dry dishes while standing (DISHES), laundry (removing towels from basket and folding) while standing (LAUNDRY), 400-meter walk (WALK), assemble puzzle while sitting (PUZZLE), and dust mopping while standing (MOPPING). The Oxycon Mobile measured VO₂ was converted to average energy expenditure during each activity in METs, by dividing the resting oxygen intake by 3.0 mL/ (kg*min). The 3.0 mL/ (kg*min) is a departure from the conventional value 3.5 mL/ (kg*min), but given the age-related decline in resting metabolic rate (Tzankoff and Norris 1977), a lower value is a more accurate estimate of resting metabolic rate for older adults as observed by other investigators (Reidlinger, Willis et al. 2015). Using too large a value for resting metabolic rate (e.g., 3.5) in the denominator of a MET introduces a downward bias (underestimation) of the participant’s activity-related MET value (Kozey, Lyden et al. 2010). More details about these measurements and protocol can be found elsewhere, including the collection of descriptive characteristics, weight, and height (Evenson, Wen et al. 2015).

Statistical Analysis

Summary statistics described distributions of participant characteristics for the OPACH main study and calibration study. Using the equation in Bai et al. (Bai, Di et al. 2016) (implemented in an R package “ActivityIndex”), we first calculated AAI per second for each participant based on 30 Hz raw data from hip-worn accelerometers. In particular, we calculate the AAI in relative scale (their formula 2) with the sigma value (systematic noise standard deviation) estimated based on raw acceleration signals when the device is not moving (0.002559424). They were then aggregated into AAI per 15-seconds by taking the sum of AAI per second within the 15-second epoch. The AAI was shown to have desirable mathematical properties including additivity and rotational invariance, and thus was not affected by variation in device orientations among participants. In the OPACH calibration study, we used data during minutes 3–7 for activities lasting 7 minutes (DVD, DISHES, LAUNDRY, PUZZLE, and MOPPING). For the 400-meter walk, we used data from minute 3 to the end of the walk. Oxygen uptake during these activities measured by Oxycon Mobile devices were used to calculate activity-specific intensity defined in units of METs, which represent the ratio of activity energy expenditure to resting energy expenditure (we used 3.0 mL·min^{− 1}·kg^{− 1}, the median value measured in our sample while sitting quietly watching a DVD) (Evenson, Wen et al. 2015). After deriving AAIs, histograms were plotted to show their distributions for each of the six activity types.

Linear regression was used to study the relationship between AAI and METs and to build calibration equations to predict continuous METs. Three types of transformations of AAI were explored: original, square root, and logarithm scales. When modeling the relationship between METs and AAI per 15-seconds, the fitted line was forced to go through the point (0,1) to reflect prior knowledge that a zero AAI per 15-seconds generally implies that a person stays still with little movement, so the corresponding activity intensity is exactly 1 MET. We fit two models, one with AAI only and the other allowing an additional interaction with age groups (selected based on quartiles to ensure equal group sizes). The latter yielded age-specific calibration equations. Five-fold cross-validation was used to estimate root mean square error and R² to evaluate the predictive performance of these models (Hastie, Tibshirani et al. 2017).

Receiver operating characteristic (ROC) curve analysis was conducted to derive AAI cutpoints for classifying a 15-seconds epoch into one of the intensity categories: SB, LLPA, HLPA and MVPA(Jago, Zakeri et al. 2007). We identified cutpoints by balancing the number of false positives and false negatives. The area under the ROC curve (AUC) was estimated using generalized estimating equations with logistic regression (Liang and Zeger 1986) to account for within-person correlations. The AUC represented the predictive accuracy of accelerometer metrics to classify activity intensity categories (Pepe 2004).

In the OPACH main study, we first calculated AAI per 15-seconds during wear periods while awake for each woman and then applied the derived cutpoints to estimate daily average minutes spent in each intensity category. To study the relationships between AAI-derived activity measures and BMI, glucose, and HDL, we fit linear regression models. Two types of linear regression models, single activity linear regression models and isotemporal substitution models, were used to examine the associations between time spent in PA and SB per day and BMI, glucose, and HDL (Mekary, Willett et al. 2009, Stamatakis, Rogers et al. 2015). For the glucose and HDL cholesterol outcomes, both models adjusted for average awake wear time, age, race and ethnicity, education, and BMI groups. For the BMI outcome, we only adjusted for average awake wear time, age, race and ethnicity, and education. For comparison, we repeated the above association analysis with the same statistical methods using counts-based PA variables that are derived with data from the same cohort and same laboratory.

Results

Description of Samples

The analysis sample from the main study included 4,655 OPACH women who had at least four days of complete accelerometry data as well as the cardiometabolic risk factors assessed (Table 1). They had mean age of 78.9 (SD 6.7) years and their BMI was evenly distributed among normal or underweight (33.4%), overweight (30.1%), and obese (36.5%) categories. A total of 199 of the 200 women in the OPACH calibration study had complete raw data (30Hz) available and were our analysis sample. They had mean age of 75.4 (SD 7.7) years and their BMI was also evenly distributed among normal or underweight (36.2%), overweight (30.6%), and obese (33.2%) categories.

Table 1.

OPACH main (N=4655) and calibration (N=199) study participant characteristics

	OPACH Main Study	OPACH Calibration Study
	Mean (SD)	Mean (SD)
Age, years	78.9 (6.6)	75.5 (7.7)
BMI, kg/m2	27.9 (5.7)	28.0 (6.0)
HDL cholesterol, mg/dL	60.5 (15.0)
Glucose, mg/dL	98.2 (27.6)
	N (%)	N (%)
Age, years
60 – 69	461 (9.9)	43 (21.6)
70 – 79	1813 (38.9)	88 (44.2)
80+	2381 (51.1)	68 (34.2)
Race/Ethnicity
White	2441 (52.4)	100 (50.3)
Black	1411 (30.3)	64 (32.2)
Hispanic	803 (17.3)	35 (17.6)
Education
High school or less	940 (20.2)	29 (14.6)
Some college	1778 (38.2)	68 (34.2)
College graduate	1913 (41.1)	102 (51.3)
BMI, kg/m2
Underweight	73 (1.6)	3 (1.5)
Normal	1430 (30.7)	69 (34.7)
Overweight	1672 (35.9)	63 (31.7)
Obese	1387 (29.8)	64 (32.2)

Abbreviations: SD = standard deviation; BMI = body mass index; HDL= high density lipoprotein cholesterol.

Calibration: Derivation of Calibration Equations for AAI

AAIs were calculated and histograms for AAI per 15-seconds are shown in Figure 1. The mean (SD) of measured METs were 1.0 (0.2) for DVD, 1.3 (0.3) for PUZZLE, 1.8 (0.4) for DISHES, 2.0 (0.4) for LAUNDRY, 2.5 (0.6) for MOPPING and 3.7 (0.7) for WALK, while the mean (SD) of corresponding AAI per 15-seconds were 9.2 (27.6), 78.1 (45.1), 139.7 (72.1), 203.3 (71.7), 383.7 (139.9), and 928.3 (258.0), respectively. As expected, epochs with higher intensity activities (METs) had higher AAI levels.

Figure 1.

Histograms of AAI per 15-seconds by activity type in OPACH calibration study (N=199). The six types of activities are: watching DVD while sitting quietly, DVD; assembling puzzle while sitting, PUZZLE; washing dishes while standing, DISHES; doing laundry while standing, LAUNDRY; dust mopping while standing, MOPPING; 400-meter walking, WALK. The x and y axis across the figures are different. Abbreviations: AAI = Accelerometry Activity Index; OPACH = Objective Physical Activity and Cardiovascular Health.

Linear regression models were used to build calibration equations for continuous METs using AAIs. After exploring three transformations of AAI (original, logarithm, and square root), we found that the square root transformation yielded the best fit (Figure 2) and thus used it in subsequent analysis. Table 2 shows model fitting results from two models: univariate (AAI only, model 1) and interaction with age (model 2). Note that the models had no intercept or main effects of age, as it is desirable to force MET=1 when AAI=0, regardless of age. Based on model 1, AAI was highly predictive for METs with an R² of 0.74. This model provided a calibration equation, MET = 1 + 0.08*sqrt root(AAI), which implied a strong positive relationship between AAI and METs. Allowing age-specific relationships slightly improved predictive performance (R²=0.77 in Model 2).

Figure 2.

Scatterplots of METs versus AAI per 15-seconds (original, square root, and logarithm scales) in OPACH calibration study (N=199), with linear fitted lines superimposed. Abbreviations: MET = Metabolic Equivalents; AAI = Accelerometry Activity Index; OPACH = Objective Physical Activity and Cardiovascular Health.

Table 2.

Calibration equations for METs based on AAI per 15-seconds epochs from hip-worn accelerometers; WHI OPACH Calibration Study (N=199).

	Variable	Calibration Equation	RMSE	R-square
Model 1: Overall			0.503	0.74
	AAI	$MET=1+0.070\sqrt{AAI}$
Model 2: Age specific			0.482	0.77
	60 – 69 years old	$MET=1+0.070\sqrt{AAI}$
	70 – 74 years old	$MET=1+0.073\sqrt{AAI}$
	75 – 81 years old	$MET=1+0.080\sqrt{AAI}$
	82+ years old	$MET=1+0.091\sqrt{AAI}$

Abbreviations: AAI=Accelerometer activity index; MET = metabolic equivalents; RMSE: root mean square errors; All models have p-value < 0.001.

Calibration: Derivation of AAI Cutpoints

Table 3 displays ROC-derived cutpoints for classifying AAI per 15-seconds epochs into one of the four intensity categories: SB, LLPA, HLPA and MVPA. The estimated AAI cutpoints were 101, 270, and 573 for SB vs. LLPA, LLPA vs. HLPA, and HLPA vs. MVPA, respectively, with corresponding area under the ROC curve (AUC) values of 0.93, 0.94, and 0.97, respectively.

Table 3.

Hip-worn accelerometer cutpoints for AAI per 15-seconds derived from ROC-based approach (N=199); WHI OPACH Calibration Study (N=199).

Activity Intensity	Cutpoints (AAI/15-sec)	Sensitivity	Specificity	Sensitivity+Specificity	AUC
SB vs. LLPA	101	0.79	0.88	1.67	0.92
LLPA vs. HLPA	270	0.79	0.90	1.70	0.94
HLPA vs. MVPA	587	0.82	0.97	1.79	0.97

Abbreviations: AUC = Area under the ROC curve; SB = sedentary behavior; LLPA = low light physical activity; HLPA = high light physical activity; MVPA = moderate and vigorous physical activity

Table 4 contains descriptive statistics of PA-related metrics derived using both AAI- and count-based cutpoints for the OPACH main study. Notably, AAI-based metrics implied an average of 9.3 min/d more SB, 8.5 min/d less LLPA, 11.5 min/d more HLPA and 12.3 min/d less MVPA for these women compared to those by count-based metrics.

Table 4.

Summary statistics of metrics derived from AAI and count cutpoints for the OPACH cohort (N=4,655).

PA-related metrics	Q1	Mean	SD	Median	Q3
Awake wear time (min/d)	846.6	893.9	78.0	900.1	947.7
Counts per 15-second	72.5	103.0	42.4	97.4	128.1
AAI per 15-second	107.3	140.7	46.6	135.4	168.5
Count-based intensity
SB time (min/d)	487.5	553.2	99.3	555.5	619.6
LLPA time (min/d)	154.0	189.3	49.8	187.2	220.0
HLPA time (min/d)	74.6	99.5	35.6	97.1	121.9
MVPA time (min/d)	25.5	51.9	35.0	45.0	70.5
AAI-based intensity time
SB time (min/d)	492.4	562.5	104.1	563.9	633.1
LLPA time (min/d)	142.3	180.8	53.5	178.4	216.1
HLPA time (min/d)	82.8	111.0	40.5	107.1	135.1
MVPA time (min/d)	18.2	39.6	28.7	33.8	54.5

Abbreviations: SB = sedentary behavior; LLPA = low light physical activity; HLPA = high light physical activity; MVPA = moderate and vigorous physical activity

We also conducted ROC analysis to derive age-specific cutpoints for each of four age categories, 60–69, 70–74, 75–81, 82+ years old (Supplementary Table S1). The results showed that cut-points for each intensity category decrease with age. These cutpoints were then used to calculate daily summary variables for SB, LLPA, HLPA and MVPA. Descriptive statistics for these variables (Supplementary Table S2) showed generally similar patterns with those based on universal cutpoints (Table 4), with around 5 minutes/day in mean differences of LLPA, HLPA and MVPA and 16 minutes/day in mean difference of SB.

Association with health outcomes

Associations of each PA measure and three cardiometabolic risk factors (BMI, glucose, and HDL) were estimated from linear regression. Table 5 shows model fitting results, where each model contained only one PA measure and was adjusted for average awake wear time (minutes), age, BMI (except when BMI was the outcome), race and ethnicity, and education. Each coefficient represented effects of 30 min/d difference in time spent in the corresponding activity intensity category. For AAI-based measures, all associations were statistically significant. There were generally monotonic dose-response relationships between all three cardiometabolic outcomes and PA with varying intensity levels. For example, each 30 min/d incremental difference in SB was associated with 0.74 mg/dL lower HDL, while each 30 min/d incremental difference in LLPA, HLPA and MVPA were associated with 1.00, 1.46 and 2.15 mg/dL higher HDL, respectively, demonstrating stronger associations with higher intensity PA. Comparing analysis based on two sets of measures, AAI-based analysis provided a clearer dose-response pattern across varying PA intensity levels for each health outcome. For example, count-based analysis estimated similar effects of HLPA and MVPA on HDL (1.68 and 1.61 mg/dL increase per 30 min/d increase in HLPA and MVPA, respectively), while AAI-based analysis estimated stronger association with MVPA than HLPA (2.15 and 1.46, respectively).

Table 5.

Association between time spent in each intensity category (SB, LLPA, HLPA and MVPA) and BMI, blood glucose, and HDL cholesterol for OPACH Study (N=4,655). Each model was adjusted for average awake wear time (minutes), age, BMI (except when BMI was the outcome), race/ethnicity, and education. Separate linear regression analyses were conducted based on PA metrics derived from AAI and count cutpoints.

	AAI-based intensity				Count-based intensity
Outcomes	SB	LLPA	HLPA	MVPA	SB	LLPA	HLPA	MVPA
BMI
Coefficient*	0.62	−0.70	−1.52	−1.85	0.65	−0.78	−1.85	−1.22
Standard Error	0.03	0.05	0.06	0.09	0.03	0.05	0.07	0.08
P-value	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001
Glucose
Coefficient*	0.74	−0.66	−1.71	−2.99	0.80	−0.67	−1.82	−2.42
Standard Error	0.14	0.25	0.34	0.50	0.15	0.27	0.39	0.39
P-value	<0.001	0.001	<0.001	<0.001	<0.001	0.016	<0.001	<0.001
HDL cholesterol
Coefficient*	−0.74	1.01	1.46	2.15	−0.79	1.08	1.68	1.61
Standard Error	0.07	0.13	0.18	0.26	0.08	0.14	0.20	0.21
P-value	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001

Abbreviations: SB = sedentary behavior; LLPA = low light physical activity; HLPA = high light physical activity; MVPA = moderate and vigorous physical activity; BMI = body mass index.

^*indicate association with 30 min/d incremental difference in physical activity measures

We also fit isotemporal substitution models to estimate the potential effect associated with reallocating 30 min/d between two activity intensity categories (Table 6). In this analysis, LLPA and HLPA were combined into a single category, light PA. For AAI-based measures, theoretically reallocating 30 min/d of SB by MVPA was significantly associated with lower BMI and glucose and an higher HDL. Substitution from SB to light PA was associated with lower BMI and glucose and an higher HDL, but the effect is not as strong as that when reallocating SB to MVPA. Reallocation between light PA and MVPA shows similar patterns. AAI-based analysis had higher power to detect substitutional effects than count-based analysis. For example, effects of reallocating LPA to MVPA were not significantly associated with BMI and HDL in count-based analysis, but the associations were stronger and statistically significant in AAI-based analysis,.

Table 6.

Reallocation of equivalent time spent in two activity intensity categories in relation to BMI, blood glucose, and HDL cholesterol for OPACH Study (N=4,655). Each model was adjusted for average awake wear time (minutes), age, BMI (except when BMI was the outcome), race/ethnicity, and education. Separate linear regression analyses were conducted based on PA metrics derived from AAI and count cutpoints.

	AAI-based intensity			Count-based intensity
Outcome	SB -> LPA	SB -> MVPA	LPA -> MVPA	SB -> LPA	SB -> MVPA	LPA -> MVPA
BMI
Coefficient *	−0.49	−1.29	−0.81	−0.62	−0.72	−0.10
Standard Error	0.03	0.10	0.11	0.03	0.08	0.10
p-value	<0.001	<0.001	<0.001	<0.001	<0.001	0.319
Glucose
Coefficient *	−0.41	−2.59	−2.18	−0.39	−2.16	−1.78
Standard Error	0.17	0.53	0.60	0.19	0.41	0.50
p-value	0.015	<0.001	<0.001	0.043	<0.001	<0.001
HDL cholesterol
Coefficient *	0.60	1.56	0.96	0.68	1.15	0.48
Standard Error	0.09	0.27	0.31	0.10	0.21	0.26
p-value	<0.001	<0.001	0.002	<0.001	<0.001	0.067

Abbreviations: SB = sedentary behavior; LLPA = low light physical activity; HLPA = high light physical activity; MVPA = moderate and vigorous physical activity; BMI = body mass index.

^*indicate effects associated with 30 min/d increase in physical activity measures

As sensitivity analysis, we re-ran regression models using AAI-derived intensity measures based on age-specific cut-points (Supplementary Table S3 and S4). There were quantitative differences between these results and prior results based on universal cut-points. However, the overall patterns on comparisons between AAI- and counts-based analysis remain similar, i.e., AAI-based analysis provided a clearer dose-response pattern between PA intensity measures and outcome variables and had higher power to detect substitutional effects, especially effects of reallocating light PA to MVPA.

Discussion

This paper provides calibration of AAI, a raw accelerometry based metric, where PA intensity cutpoints were derived among older women using OPACH calibration study data wherein hip-worn triaxial Actigraph GT3X output was calibrated to ground truth oxygen uptake measured during several activity tasks of daily living. PA intensity measures derived from these cutpoints were then used to study associations between PA intensity and cardiometabolic risk factors in the OPACH main study. Our study provides a framework for calibrating raw accelerometry based metric and deriving intensity cutpoints from such calibration that can be used in subsequent association analysis. The activity intensity cutpoints and calibration equations derived in this study can be applied to other studies that collected raw accelerometry data in similar populations, for example the Women’s Health Study (Shiroma, Freedson et al. 2013). For different populations, such as different age or gender groups, similar methods can be directly applied to derive population specific intensity cutpoints and calibration equations if calibration study data are available. In particular, this is not limited to AAI, and it applies to any raw accelerometry based metrics.

We applied ROC analysis to derive AAI cutpoints for PA intensity classification. Evaluation metrics such as sensitivity, specificity, and AUC were used to demonstrate that the AAI had great predictive performance for activity intensity. For instance, AUCs for AAI-based classification were 0.93, 0.94, and 0.97 for SB vs. LLPA, LLPA vs. LHPA and LHPA vs. MVPA, respectively. We derived the cutpoints by balancing the number of false positives and false negatives in ROC analysis. As pointed out by Evenson et al. (Evenson, Wen et al. 2015), this approach provides roughly unbiased estimates for variables with low positive predicted values such as MVPA for older adults, for whom the prevalence of such activities tends to be relatively low.

To illustrate the usefulness of the derived AAI cutpoints, we then investigated associations of AAI-based intensity measures and three cardiometabolic risk factors: BMI, glucose, and HDL cholesterol in the OPACH main study. Based on prior reviews (2018 PA Guidelines Advisory Committee 2018, Jakicic, Powell et al. 2019, Katzmarzyk, Powell et al. 2019, Barone Gibbs, Hivert et al. 2021), we would expect MVPA to be positively associated with HDL and inversely associated with glucose and BMI, while SB would be inversely associated with HDL and glucose and positively associated with BMI. Linear regression models were used to estimate association between specific activity intensity and outcomes, while isotemporal substitution model estimates an expected effect of reallocating equivalent time units between two different intensity categories. These analyses implied that greater time spent in light PA or MVPA were associated with a more favorable cardiometabolic risk profile, as was reallocating SB to light PA or MVPA. We compared the linear regression results based on AAI to those based on count-derived measures, which were previously reported by (LaMonte, Lewis et al. 2017). Note that we re-analyzed counts data to ensure more meaningful comparison with AAI-based analysis, as our analysis sample (N=4,655) is slightly smaller than their original analysis (N=4,832). The conclusions were similar for most associations, though there were a few noticeable differences. Effects of MVPA and these cardiometabolic outcomes were weaker with counts-based measures. In addition, with count-based measures, the dose-response pattern across PA intensity categories were not as clear. For example, MVPA associations were even weaker than LHPA associations with BMI and HDL. For isotemporal substitution models, analysis using count-based metrics yielded similar conclusions for reallocation of equivalent time from SB to light PA, though results were different for a few associations involving MVPA. Notably, reallocating time from light PA to MVPA were not statistically significant with BMI and HDL outcomes in counts-based measures, but were significant for AAI-based measures. Overall, AAI-based summary measures often demonstrated stronger associations, especially for MVPA, and clearer dose-response relationships across varying intensity levels as compared to count-based measures.

Regarding comparison between AAI and counts, it is worth noting that AAI demonstrated better predictive accuracy for METs (higher R² in calibration equations, e.g., 0.74 with AAI v.s. 0.54 with counts) and intensity categories (higher AUC in ROC analysis, e.g., 0.92, 0.94, 0.97 with AAI v.s. 0.84, 0.88, 0.90 with counts) than counts that were reported in a previous calibration study (Evenson, Wen et al. 2015). Both AAI-based and counts-based intensity measures can be considered noisy measurements of true activity intensity (e.g., oxygen uptake). For linear regression with imperfect measured predictors, it is well known that regression coefficients are generally attenuated compared to the effects of true exposures and that the magnitude of attenuation bias increases with the magnitude of measurement error (Carroll, Ruppert et al. 2006). Our results suggest that AAI-based measures are associated with substantially less measurement errors than count-based measures, and thus it is expected that regression based on using AAI-based variables are closer to effects of true activity intensity measures.

This study is limited by the setting, in that it was conducted only in the laboratory which may not fully reflect activities in the real world. Additional free-living collection could help extend the findings. The most significant strength of this study is the calibration of raw accelerometry based metric that facilitates comparisons between studies collecting raw accelerometry data using different devices. The OPACH study also had a large sample size which enhances measurement precision and we focused on women 60 to 91 years old, an understudied age group.

Conclusion

With the availability of raw accelerometry data, nonproprietary metrics have been proposed based on this raw accelerometry data. Compared to counts data given by manufacture algorithms, these raw accelerometry based metrics make comparisons between studies with different devices possible. However, it is often important to calibrate these new metrics for different populations. In this study, we provide a calibration framework based on AAI, a recently proposed raw accelerometry based metric, for the OPACH study that included women 60–91 years. PA intensity cutpoints were derived from the calibration study within OPACH and an association analysis between PA intensity and cardiometabolic risk factors demonstrated the effectiveness of the proposed method. Future work could include calibrations with data from broader demographic groups, such as younger age groups and men, and data collected from free-living settings to expand on this work. The results of the present study indicate that AAI is a promising metric for future studies that collect raw accelerometry data.

Figure 3.

Scatterplots between METs and AAI per 15-seconds from the OPACH Calibration Study (N=199). The left panel illustrates fitted linear model using the full sample. The right panel added fitted lines between METs and AAI for each age category. Abbreviations: MET = Metabolic Equivalents; AAI = Accelerometry Activity Index; OPACH = Objective Physical Activity and Cardiovascular Health.