NHANES 2011–2014: Objective Physical Activity is the Strongest Predictor of All-Cause Mortality

Abstract

Introduction:

Objectively measured physical activity (PA) is a modifiable risk factor for mortality. Understanding the predictive performance of PA is essential to establish potential targets for early intervention to reduce mortality among older adults.

Methods:

The study used a subset of the National Health and Nutrition Examination Survey (NHANES) 2011–2014 data consisting of participants aged 50 to 80 years old (n = 3653, 24297.5 person-years of follow-up, 416 deaths). Eight accelerometry derived features and 14 traditional predictors of all-cause mortality were compared and ranked in terms of their individual and combined predictive performance using the 10-fold cross-validated Concordance (C) from Cox regression.

Results:

The top three predictors of mortality in univariate analysis were PA related: average MIMS in the 10 most active hours (C=0.697), total MIMS per day (C=0.686), and average log transformed MIMS in the most 10 active hours of the day (C=0.684), outperforming age (C=0.676) and other traditional predictors of mortality. In multivariate regression, adding objectively measured PA to the top performing model without PA variables increased concordance from C=0.776 to C=0.790 (p<0.001).

Conclusions:

These findings highlight the importance of PA as a risk marker of mortality and are consistent with prior studies, confirming the importance of accelerometer-derived activity measures beyond total volume.

INTRODUCTION

Physical activity (PA) is an important modifiable risk factor for human morbidity and mortality. The emergence of wearable accelerometers led to a paradigm shift for PA measurement, allowing objective, continuous, and non-invasive assessment of PA in an individual’s environment, which may be more informative to health outcomes than self-reported PA or physical function measures in laboratory or clinical settings (1). Indeed, objective physical activity (PA) measurements are strong predictors of mortality (2–10). Yet, despite the extraordinary recent successes of objective PA research, there is an ever increasing need to reproduce, validate, and improve the translation of these findings.

The results presented here reproduce key findings in similar studies using data from NHANES 2003–2006 and the UK Biobank and expand on these findings by adding assessment of wrist-measured PA in the NHANES population.

The National Health and Nutrition Examination Survey (NHANES) is a large, ongoing, cross-sectional study of the non-institutionalized US population conducted by the Centers for Disease Control (CDC) in two-year waves using a multi-stage stratified sampling scheme. Participants in the NHANES 2011–2014 waves are a nationally representative sample from the non-institutionalized US population. NHANES collects a vast array of demographic, socioeconomic, lifestyle and medical data, though the exact data collected and cohort varies from year to year. The wrist-worn accelerometry data collected in the NHANES 2011–2012 and 2013–2014 waves was released in December 2020. This data is of particular interest because: 1) it is publicly available and linked to the National Death Index (NDI) through the National Center for Health Statistics; 2) the sample is representative of the non-institutionalized US population after accounting for the complex survey design; 3) it was collected using wrist accelerometers and processed into “monitor-independent movement summary” (MIMS) units using an open source, reproducible algorithm (https://github.com/mHealthGroup/MIMSunit); 4) unlike the previous hip-worn NHANES accelerometry data (NHANES 2003–2006), the protocol required 24-hour continuous wear of the wrist accelerometers; and 5) the protocol for data collection, including wear location and sub-second data sampling, was closer to the UK Biobank accelerometry study, a very large study that included accelerometry on more than 100,000 individuals in the UK (9). Thus, NHANES 2011–2014 accelerometry data is a crucial new resource that can be used to reproduce, validate, and improve previous results.

Two previous studies assessed the predictive value of accelerometry derived PA features relative to traditional risk factors in predicting all-cause mortality (8,9). In NHANES 2003–2006, (8) showed that the volume of high intensity activities and PA fragmentation were the most highly predictive objective measures of PA for 5-year all-cause mortality and one measure of volume out-performed age as measured by the Receiver Operating Area Under the Curve (ROC-AUC or AUC for short). A similar study in the UK Biobank (9) using a survival analysis approach found that objective measures of volume and circadian rhythmicity were most highly associated with all-cause mortality, though age outperformed PA measures. Using a time to event analysis in the UK Biobank, (9) showed that PA measures were the strongest predictors of all-cause mortality in single-predictor models and added substantial predictive performance above and beyond that of traditional predictors of mortality (e.g., age, sex, cigarette smoking). While the results of the two studies agreed qualitatively, some differences may be due to device placement (hip in NHANES 2003–2006 versus wrist in UK Biobank), PA measures summarization (activity counts in NHANES versus ENMO in UK Biobank), wear time protocol (waking hours versus 24-hours), study population (nationally representative US, convenience sample UK), sample characteristics (age at baseline, cohort effect, etc.), or analytic approach (binary versus time-to-event outcomes). The new NHANES 2011–2014 study provides the opportunity to study the association between objectively measured PA using wrist-worn accelerometers and time to all-cause mortality in a US representative population. This is important as more studies use wrist accelerometry due its acceptance by study participants, high rates of compliance, and acceptability during sleep and wake periods. Our paper investigates whether the association between PA and time to all-cause mortality is maintained when switching from hip to wrist actigraphy and quantifies the strength of this association as measured by 10-fold cross-validated Concordance (C). We note that our work here focuses strictly on the predictive power of objectively measured PA without making any causal claims due to the cross-sectional nature of the data and the potential for reverse causality. That is, PA may have a causal effect on mortality risk, reflect existing disease severity, or contribute to risk through some combination of the two.

The primary goal of the current study was to quantify the predictive performance of objective PA measures obtained from wrist-worn accelerometers for time to death in the NHANES 2011–2014 among middle-aged and older adults. A total of 8 accelerometry derived features and 14 traditional predictors of all-cause mortality were selected for this analysis. Their predictive performance was quantified and ranked based on individual (one predictor at a time) and combined (multiple predictors) 10-fold cross-validated C-index. We contextualize our findings using previously published results from two other large cohort accelerometry studies: NHANES 2003–2006 and UK Biobank.

METHODS

Study Population

In this study we used data from the NHANES 2011–2014 waves. In the 2011–2014 waves, all participants 6 years and older were asked to wear a wrist-worn accelerometer, the ActiGraph model GT3X+ (ActiGraph, Pensacola, FL) device, consecutively over 9 calendar days. The first and last days were partial days per protocol, resulting in up to 7 full calendar days of data. The NHANES accelerometry data are provided at several resolutions: day, hour, and minute level. In this study, we use the minute level MIMS units for deriving all accelerometry features considered in our models. Broadly, we followed the data processing pipeline proposed by (11) for organizing and analyzing the NHANES 2003–2006 data, which has also been used for analyzing the UK Biobank accelerometry data. Additionally, we applied exclusion criteria based on calendar day data quality and estimated wear-time. Specifically, we considered a day to be “valid” if the day had at least 95% estimated wear time (1368 minutes) with no data quality flags as determined by NHANES. We further required participants to have at least 3 days of valid data. Participants aged 50 to 80 years old at the time of the household interview were eligible for inclusion in this study (n=4609). Individuals aged 80 and over at the time of interview (n=715) were excluded as NHANES 2011–2014 topcodes individuals’ age at 80 for privacy reasons. Among individuals who met our age criteria, a total of n = 3653 study participants (ages 50 to 80 years old, 24297.5 person-years of follow-up, 416 deaths) had both wrist accelerometry data that met specified quality criteria and complete data on a set of traditional mortality risk factors. Survey weights for the combined 2011–2012 and 2013–2014 waves were calculated using the 2-year examination weights divided by 2 as suggested by the NHANES analytic guidelines.

Variables and Measures

Traditional mortality predictors.

We merged the NHANES 2011–2014 data with mortality data obtained through national mortality registries (12) with follow-up through 2019. We considered the following 14 variables as traditional predictors of mortality: age (at interview), gender (male/female), race/ethnicity (non-Hispanic White, non-Hispanic Black, Mexican American, other Hispanic, non-Hispanic Asian, other), self-reported overall health (poor, fair, good, very good, excellent), educational level (less than high school, high school equivalent, more than high school), body mass index (BMI, kg/m², underweight, normal, overweight, obese), diabetes (yes/no with pre-diabetes considered “no”), coronary heart disease (yes/no), congestive heart failure (yes/no), heart attack (yes/no), stroke (yes/no), cancer (yes/no), alcohol consumption (moderate drinker, heavy drinker, former drinker, never drinker, missing alcohol consumption data), cigarette smoking (never, former, current), and mobility problem (difficulty either: 1) walking without special equipment, 2) walking up ten steps, or 3) walking a quarter mile versus no difficulties). These are the same predictors considered by (8) and (9) and correspond to the best-known risk factors of all-cause mortality.

Notice that all comorbid conditions were self-reported via response to questions of the form “Has a doctor or other health professional ever told you that you had (comorbid condition)?”. Any individuals who responded “don’t know” or “refused” to any self-reported variable were considered missing for this analysis and dropped from the analytic sample. The one exception to dropping missing data is the alcohol consumption variable, for which we created a level “missing alcohol”.

Accelerometry-derived predictors.

Data was provided in MIMS units, a relatively new proposed summary of the raw sub-second acceleration data provided by accelerometers. Briefly, MIMS units are obtained by starting with raw acceleration values, interpolating the data to 100Hz, extrapolating the signal if it reaches the device maximum value, applying a bandpass filter (0.2–5.0 Hz), taking the absolute value, taking the area under the curve of each axis using the trapezoidal rule, and then adding these areas under the curve across the three axes to get one value per epoch. In addition to providing MIMS units at different temporal resolutions, the NHANES 2011–2014 accelerometry data were provided separately for each of the three orthogonal axes (X, Y, Z) and combined. Here we used the combined triaxial MIMS units. In addition to MIMS units, flags for wear: wake, wear: sleep, and non-wear were estimated using a procedure described in the help files for the minute level accelerometry data and provided by NHANES (13). We used combined triaxial minute level MIMS units to create day-level summaries of volume and patterns of physical activity which were then averaged across days for each participant.

A brief description of the procedure from minute level data to participant level summaries is as follows: 1) minute level MIMS and flag data were transformed in the 1440+ memory efficient format proposed by (11); 2) days were subset based on sufficient estimated wear time (≥ 95% of the day); 3) MIMS during estimated non-wear minutes were imputed using functional principal component analysis on the natural log scale; 4) participants were excluded for having fewer than three days of data with sufficient estimated wear time; 5) day level features were calculated among days with sufficient estimated wear time; and 6) features were averaged across days for each participant.

Notice that the algorithm used to create estimated sleep, wake, and non-wear flags was, at times, unable to classify a minute as belonging to any of these three categories, resulting in an “unknown” flag for these minutes. For this analysis, we considered “unknown” to be wear time based on visual inspection of the data, as the unknown category was often intermittent and sandwiched between two estimated minutes of wear. Participants included in this analysis had an average of 63 “unknown” minutes per day.

This analysis considered features which were derived from minute level physical activity. We only included features that do not require thresholds for activity intensity separation, since there were no well-established thresholds for MIMS units. Using the NHANES 2003–2006 data, (14) showed that applying a log transformation to the minute level activity count data resulted in features which were more correlated with higher intensity activities (untransformed) or lower intensity activities (log transformed). Accordingly, we calculated features which do not depend on thresholds using the MIMS units as well as the log(1+MIMS) transformation applied at the minute level.

The 8 features included in this analysis were: 1) total MIMS (TMIMS); 2) total log(1+MIMS) (TLogMIMS); 3) average MIMS during the most active 10 hours (M10_MIMS); 4) average log(1+MIMS) during the most active 10 hours (M10_LogMIMS); 5) average MIMS during the least active 5 hours (L5_MIMS); 6) average log(1+MIMS) during the least active 5 hours (L5_LogMIMS); 7) relative amplitude on the MIMS scale (RA_MIMS), defined as (M10_MIMS-L5_MIMS)/(M10_MIMS+L5_MIMS); 8) relative amplitude on the log(1+MIMS) scale (RA_LogMIMS), defined as (M10_LogMIMS-L5_LogMIMS)/(M10_LogMIMS+L5_LogMIMS).

Each of the 8 features described above captures different, but complementary aspects of an individual’s PA. Because little has been published on estimating sleep features using MIMS, we only considered physical activity volume (TMIMS, TLogMIMS) and circadian features (M10, L5, RA). An in-depth discussion of the motivation for choosing corollaries of these specific features derived from wrist accelerometry data using another unit of measure (ENMO: Euclidean Norm Minus One) was provided in (9). In addition, the findings from (9) suggested that different features may be more or less predictive in different populations with different survival follow-up times. We therefore compared similar features in NHANES 2011–2014 with NHANES 2003–2006, which is nationally representative of the US population 2003–2006 but had longer follow-up time and estimated the association between PA and 5-year all-cause mortality (binary outcome). We also compared to the UK Biobank, which is not a UK nationally representative due to selection bias inherent in the data acquisition process (15) and also had notably longer follow-up time (5.4 years average follow-up time, maximum 6.8 years), but used summaries derived from a wrist-worn accelerometer, used a Cox regression approach, and quantified the time-dependent predictive performance of the top predictors of all-cause mortality, allowing for a comparison of short-term predictive performance of key variables of interest.

Because accelerometry features are often highly correlated both with each other and age, we created a correlation plot showing the strength of the pairwise association between PA features and age.

Statistical Analysis

Mortality prediction models.

We assessed individual and combined predictive performance using 10-fold cross-validated survey weighted Concordance (cvC). Concordance was estimated using risk estimates from Cox regression models fit using participants’ examination survey weights. For assessing the predictive performance of individual variables, a series of separate Cox regressions were fit for each predictor. For assessing the combined predictive performance of multiple variables, we employed a forward selection procedure whereby variables were included one at a time based on their improvement in cross-validated Concordance. Variables were included sequentially until the next best candidate variable did not improve Concordance by more than a pre-specified threshold. We considered three thresholds for improvement in cross-validated Concordance: 0.01, 0.005, 0.001. These thresholds were chosen based on the findings of (9) which found the threshold of 0.01 to be very conservative, with between 2 and 4 total variables selected, while the threshold of 0.001 led to the inclusion of some predictors which did not improve out-of-sample performance. A threshold of 0.005 may provide a balance between these two stopping rules. In addition to a standard forward selection, we used a two-stage forward selection procedure for identifying the predictive value of PA features above and beyond non-PA predictors of mortality. The first stage estimated a model based on the Concordance threshold stopping rule with non-PA predictors, then PA predictors were added to this base model and were included based on the same Concordance threshold. Variability in feature rankings and the forward selection procedure were assessed using 100 different test-train splits of the data for performing cross-validation.

RESULTS

In total, there were 4,609 participants who agreed to wear an accelerometer in NHANES 2011–2014 who were between the ages of 50 and 80. Of those 4,609 individuals, 703 were excluded for not having any day with at least 95% wear time with no data quality flags. An additional 166 participants were excluded for having only 1 or 2 days with at least 95% wear time with no data quality flags. Finally, 87 participants were excluded for missing covariate or mortality data. Excluding these participants resulted in an analytic sample of 3653 individuals. The survey weighted summary statistics for the overall 2011–2014 analytic sample and stratified by NHANES wave are presented in Table 1. Characteristics for participants included in the analysis versus those excluded for either missing accelerometry, risk factor, or mortality data are presented in Supplemental Table 1 (Supplemental Digital Content). We find that excluded participants were, on average, younger, more male, less white, had lower educational attainment, higher rates of missing alcohol consumption data, and poorer self-reported health. Excluded participants also had lower rates of obesity, CHD, and cancer.

Table 1:

Survey weighted summary statistics are presented for the overall 2011–2014 analytic sample (column 2) and stratified by NHANES wave (columns 3–4). Continuous variables are presented as mean (standard deviation). Categorical variables are presented as N (%). P-values presented in column 5 correspond to t-tests for continuous variables and chi-squared tests for categorical variables.

	Overall	2011–2012	2013–2014	p-value
N	3653	1842	1811
Age at interview (years)	61.67 (7.95)	61.44 (7.95)	61.91 (7.95)	0.198
Female	1960 (53.7)	984 (53.4)	976 (53.9)	0.829
Race				0.695
Mexican American	171 (4.7)	70 (3.8)	101 (5.6)
Other Hispanic	156 (4.3)	92 (5.0)	64 (3.5)
Non-Hispanic White	2740 (75.0)	1381 (75.0)	1359 (75.1)
Non-Hispanic Black	378 (10.3)	190 (10.3)	188 (10.4)
Non-Hispanic Asian	141 (3.9)	69 (3.8)	72 (4.0)
Other Race	67 (1.8)	40 (2.2)	27 (1.5)
BMI				0.247
Normal	878 (24.0)	469 (25.4)	410 (22.6)
Underweight	51 (1.4)	19 (1.1)	32 (1.7)
Overweight	1231 (33.7)	629 (34.1)	603 (33.3)
Obese	1492 (40.8)	725 (39.4)	767 (42.3)
Education Level				0.799
Less Than High School	592 (16.2)	314 (17.0)	278 (15.4)
High School Equivalent	783 (21.4)	388 (21.1)	395 (21.8)
More Than High School	2278 (62.4)	1140 (61.9)	1138 (62.8)
Overall Health				0.168
Excellent	1393 (38.1)	682 (37.0)	711 (39.2)
Very good	364 (10.0)	215 (11.7)	149 (8.2)
Good	1136 (31.1)	594 (32.2)	542 (29.9)
Fair	627 (17.2)	294 (16.0)	333 (18.4)
Poor	133 (3.6)	57 (3.1)	76 (4.2)
Diabetes	628 (17.2)	308 (16.7)	320 (17.7)	0.564
Heart Attack	208 (5.7)	100 (5.4)	108 (6.0)	0.485
CHF	151 (4.1)	78 (4.3)	73 (4.0)	0.794
CHD	218 (6.0)	100 (5.4)	118 (6.5)	0.336
Stroke	161 (4.4)	77 (4.2)	84 (4.6)	0.625
Cancer	643 (17.6)	277 (15.0)	367 (20.3)	<0.001
Mobility Problem	870 (23.8)	370 (20.1)	500 (27.6)	0.012
Alcohol Consumption				0.032
Moderate Drinker	2081 (57.0)	1021 (55.5)	1060 (58.5)
Never Drinker	409 (11.2)	206 (11.2)	203 (11.2)
Former Drinker	686 (18.8)	318 (17.2)	368 (20.4)
Heavy Drinker	331 (9.1)	203 (11.0)	128 (7.1)
Missing Alcohol	146 (4.0)	95 (5.1)	51 (2.8)
Cigarette Smoking				0.924
Never	1803 (49.4)	901 (48.9)	902 (49.8)
Former	1205 (33.0)	607 (33.0)	598 (33.0)
Current	645 (17.6)	333 (18.1)	312 (17.2)
TMIMS	12678.55 (3549.84)	12796.17 (3585.27)	12558.91 (3510.38)	0.138
RAMIMS	0.87 (0.08)	0.88 (0.08)	0.87 (0.08)	0.181
Deceased	351 (9.6)	199 (10.8)	152 (8.4)	0.163

Figure 1 plots the estimated pairwise Pearson correlations between PA variables and age. Many of the PA variables are highly correlated, with absolute values of correlation as high as 0.96. However, the correlation between L5 and RA with both age and PA measures of volume are consistently low (maximum absolute value 0.36 with many correlations estimated at 0.20 or less). This holds true for L5 and RA calculated using both MIMS and the log-transformed MIMS units and is consistent with similar findings reported in the UK Biobank (9). This may indicate that L5 and RA may be good candidates for improving prediction above and beyond age.

Figure 1:

Pairwise correlations between age and PA-derived features.

Best Individual Predictors of Mortality

Table 2 presents the top 22 univariate predictors of mortality ranked by 10-fold cross-validated Concordance from separate univariate Cox regressions, ordered from most to least predictive. In addition to presenting cross-validated Concordance, summary statistics are shown stratified by groups with living or deceased individuals at follow-up time. Among the top 10 predictors of mortality all are objective activity measures (C between 0.65 and 0.70) except for age at interview (rank 4, C=0.676), self-reported mobility problems (rank 5, C=0.672), and self-reported overall health (rank 6, C=0.662). The top predictors of mortality contain measures from all domains of PA considered in this study (volume and circadian rhythmicity). The top predictor of mortality, average MIMS in the 10 most active hours calculated on the MIMS scale, achieves a cross-validated Concordance of 0.697. Even though many of the top ranked PA measures are highly correlated, the relative amplitude measures (calculated both on the MIMS and log-transformed MIMS scale) are less correlated with the other top PA predictors and are among the top 8 predictors of mortality.

Table 2:

Accelerometry and non-accelerometry derived predictors of mortality ranked by univariate Concordance for top 25 predictors of mortality (column 1). Survey weighted summary statistics are presented for the overall 2011–2014 analytic sample (column 2) and stratified by mortality status (columns 3–4) at follow-up. Continuous variables are presented as mean (standard deviation). Categorical variables are presented as N (%). P-values presented in column 5 correspond to t-tests for continuous variables and chi-squared tests for categorical variables.

	Overall	Alive	Deceased	p-value	Concordance
N	3653	3302	351
M10MIMS	14.49 (4.08)	14.78 (3.99)	11.81 (3.91)	<0.001	0.697
TMIMS	12678.55 (3549.84)	12911.84 (3485.36)	10520.18 (3415.13)	<0.001	0.686
M10LogMIMS	2.36 (0.36)	2.39 (0.35)	2.13 (0.40)	<0.001	0.684
Age at interview (years)	61.67 (7.95)	61.12 (7.72)	66.82 (8.28)	<0.001	0.676
Mobility Problem	870 (23.8)	671 (20.3)	199 (55.8)	<0.001	0.672
Overall Health				<0.001	0.657
TLogMIMS	2229.77 (366.25)	2249.81 (355.93)	2044.36 (406.64)	<0.001	0.653
RAMIMS	0.87 (0.08)	0.88 (0.08)	0.82 (0.12)	<0.001	0.652
RALogMIMS	0.79 (0.10)	0.80 (0.09)	0.73 (0.13)	<0.001	0.647
L5LogMIMS	0.28 (0.15)	0.27 (0.14)	0.34 (0.20)	<0.001	0.593
Diabetes	628 (17.2)	505 (15.3)	123 (34.5)	<0.001	0.592
Cigarette Smoking				<0.001	0.588
Never	1803 (49.4)	1681 (51.0)	122 (34.2)
Former	1205 (33.0)	1064 (32.3)	141 (39.6)
Current	645 (17.6)	551 (16.7)	94 (26.2)
Education Level				0.009	0.568
Less Than High School	592 (16.2)	507 (15.4)	85 (23.9)
High School Equivalent	783 (21.4)	691 (21.0)	92 (25.8)
More Than High School	2278 (62.4)	2098 (63.7)	180 (50.4)
L5MIMS	0.95 (0.73)	0.93 (0.71)	1.12 (0.87)	0.018	0.565
CHD	218 (6.0)	164 (5.0)	54 (15.1)	<0.001	0.554
Heart Attack	208 (5.7)	155 (4.7)	53 (14.9)	<0.001	0.548
Alcohol Consumption				0.002	0.546
Moderate Drinker	2081 (57.0)	1916 (58.1)	166 (46.5)
Never Drinker	409 (11.2)	372 (11.3)	37 (10.4)
Former Drinker	686 (18.8)	576 (17.5)	110 (30.9)
Heavy Drinker	331 (9.1)	302 (9.2)	29 (8.0)
Missing Alcohol	146 (4.0)	131 (4.0)	15 (4.1)
Female	1961 (53.7)	1805 (54.8)	156 (43.6)	<0.001	0.544
CHF	151 (4.1)	100 (3.0)	51 (14.3)	<0.001	0.543
Cancer	643 (17.6)	558 (16.9)	85 (23.9)	0.015	0.537
Stroke	161 (4.4)	122 (3.7)	39 (10.8)	<0.001	0.536
Race				0.212	0.523
Mexican American	171 (4.7)	157 (4.8)	14 (4.0)
Other Hispanic	156 (4.3)	145 (4.4)	11 (2.9)
Non-Hispanic White	2740 (75.0)	2471 (75.0)	269 (75.5)
Non-Hispanic Black	378 (10.3)	332 (10.1)	46 (12.9)
Non-Hispanic Asian	141 (3.9)	131 (4.0)	10 (2.8)
Other Race	67 (1.8)	60 (1.8)	7 (1.9)
BMI				0.011	0.522
Normal	878 (24.0)	803 (24.4)	76 (21.2)
Underweight	51 (1.4)	37 (1.1)	14 (3.9)
Overweight	1231 (33.7)	1126 (34.2)	105 (29.5)
Obese	1492 (40.8)	1330 (40.4)	162 (45.3)

Best Subset of Predictors of Mortality

Table 3 presents the forward selection results from combining both PA and non-PA predictors as candidates for inclusion in the model at each step of the procedure. For the most conservative stopping rule (δC ≥ 0.01), four variables were included in the model, including two PA predictors: average MIMS in the 10 most active hours calculated on the MIMS scale (M10_MIMS) and relative amplitude calculated on the log MIMS scale (RA_LogMIMS), and two non-PA predictors: mobility problems and age at interview. Increased M10 (more active during the 10 most active hours) and RA (stronger circadian rhythm) were associated with lower risk of mortality. Using the second most conservative stopping rule (δC ≥ 0.005), the cigarette smoking status and self reported overall health were selected in the model. No additional PA derived predictors were selected using this stopping rule. For the least conservative stopping rule (δC ≥ 0.001), diabetes was selected into the model. No additional PA variables were selected for this stopping rule. Cumulative cross-validated Concordance improved from 0.775 to 0.788 to 0.791 from the least to the most conservative stopping rule.

Table 3:

Accelerometry and non-accelerometry derived predictors of mortality obtained from a forward selection procedure using improvement in cross-validated Concordance as a criteria for inclusion in the model. Estimates and 95% confidence intervals for regression coefficients ($\widehat{\beta }$, change in log hazard) are corresponding the the variable listed in column 1 are presented in the column 2. Columns 3 and four present cross-validated Concordance and improvement in Concordance from the prior model, respectively.

Threshold for Improvement in Concordance = 0.01
Variable	Estimate (95% CI)	C	ΔC
M10MIMS	−0.088 (−0.123, −0.052)	0.697	0.697
Mobility Problem		0.738	0.041
No difficulty	ref
Any difficulty	0.915 (0.577, 1.254)
Age at Interview (years)	0.057 (0.044, 0.071)	0.763	0.025
RALogMIMS	−3.836 (−4.978, −2.695)	0.775	0.012
Threshold for Improvement in Concordance = 0.005
Variable	Estimate (95% CI)	C	ΔC
M10MIMS	−0.076 (−0.109, −0.042)	0.697	0.697
Mobility Problem		0.738	0.041
No difficulty	ref
Any difficulty	0.647 (0.306, 0.988)
Age at Interview (years)	0.065 (0.049, 0.080)	0.763	0.025
RALogMIMS	−3.211 (−4.486, −1.937)	0.775	0.012
Cigarette Smoking		0.783	0.008
Never	ref
Former	0.303 (−0.082, 0.688)
Current	0.49205 (0.110054, 0.875756)
Overall Health		0.788	0.005
Excellent	ref
Very good	0.154 (−0.398, 0.706)
Good	−0.294 (−0.745, 0.158)
Fair	0.515 (0.207, 0.822)
Poor	0.741 (0.251, 1.231)
Threshold for Improvement in Concordance = 0.001
Variable	Estimate (95% CI)	C	ΔC
M10MIMS	−0.073 (−0.107, −0.039)	0.697	0.697
Mobility Problem		0.738	0.041
No difficulty	ref
Any difficulty	0.589 (0.242, 0.937)
Age at Interview (years)	0.062 (0.048, 0.077)	0.763	0.025
RALogMIMS	−3.174 (−4.469, −1.878)	0.775	0.012
Cigarette Smoking		0.783	0.008
Never	ref
Former	0.283 (−0.101, 0.667)
Current	0.514 (0.125, 0.903)
Overall Health		0.788	0.005
Excellent	ref
Very good
Good
Fair
Poor
CHD		0.780	0.002
Diabetes		0.791	0.001

Table 4 has the same structure as Table 3 but contains the results from the two-stage forward selection procedure. The first stage selects the best traditional predictors, and the second stage quantifies which and how much predictive performance is improved by objective measures of PA. In the first stage, for the most conservative stopping rule (δC ≥ 0.01), age, mobility problem, and self reported overall health were the non-PA predictors selected into the model. RA calculated on the log MIMS scale was the PA variables selected. For the second most conservative stopping rule (δC ≥ 0.005), cigarette smoking, and diabetes were added to the list of non-PA variables selected, with only RA calculated on the log MIMS scale selected among the PA variables again. For the least conservative stopping rule (δC ≥ 0.001), gender was selected from the non-PA variables, while TLogMIMS was selected from the PA variables in addition to RA on the log MIMS scale. For all stopping rules, the addition of PA variables resulted in a statistically significant improvement in model fit (p < 0.001 for all) with improvements in cross-validated Concordance associated with including PA variables ranging from 0.014 to 0.020.

Table 4:

Accelerometry and non-accelerometry derived predictors of mortality obtained from a two-stage forward selection procedure using improvement in cross-validated Concordance as a criterion for inclusion in the model. In the first stage, only non-accelerometry based predictors were considered for inclusion in the model. In the second stage, accelerometry derived predictors were added to the model until the threshold in improvement in Concordance was not met. Estimates and 95% confidence intervals for regression coefficients ($\widehat{\beta }$, change in log hazard) are corresponding the the variable listed in column 1 are presented in the column 2. Columns 3 and four present cross-validated Concordance and improvement in Concordance from the prior model, respectively.

Threshold for Improvement in Concordance = 0.01
Variable	Estimate (95% CI)	C	ΔC
Age at Interview (years)	0.072 (0.058, 0.086)	0.676	0.676
Mobility Problem		0.744	0.068
No difficulty	ref
Any difficulty	0.809 (0.496, 1.122)
Overall Health		0.759	0.015
Excellent	ref
Very good	0.062 (−0.510, 0.633)
Good	−0.326 (−0.770, 0.117)
Fair	0.534 (0.241, 0.827)
Poor	0.878 (0.369, 1.387)
RALogMIMS	−4.029 (−5.263, −2.794)	0.778	0.019
Threshold for Improvement in Concordance = 0.005
Variable	Estimate (95% CI)	C	ΔC
Age at Interview (years)	0.074 (0.058, 0.089)	0.676	0.676
Mobility Problem		0.744	0.068
No difficulty	ref
Any difficulty	0.729 (0.417, 1.041)
Overall Health		0.759	0.015
Excellent	ref
Very good	0.208 (−0.354, 0.770)
Good	−0.227 (−0.679, 0.225)
Fair	0.514 (0.193, 0.835)
Poor	0.860 (0.355, 1.365)
Cigarette Smoking		0.769	0.010
Never	ref
Former	0.313 (−0.078, 0.705)
Current	0.557 (0.155, 0.958)
Diabetes	0.383 (0.031, 0.735)	0.774	0.005
RALogMIMS	−3.616 (−4.927, −2.305)	0.787	0.013
Threshold for Improvement in Concordance = 0.001
Variable	Estimate (95% CI)	C	ΔC
Age at Interview (years)	0.068 (0.052, 0.084)	0.676	0.676
Mobility Problem		0.744	0.068
No difficulty	ref
Any difficulty	0.717 (0.401, 1.033)
Overall Health		0.759	0.015
Excellent	ref
Very good	0.215 (−0.345, 0.775)
Good	−0.219 (−0.671, 0.233)
Fair	0.524 (0.213, 0.835)
Poor	0.787 (0.348, 1.226)
Cigarette Smoking		0.769	0.010
Never	ref
Former	0.234 (−0.150, 0.617)
Current	0.494 (0.104, 0.885)
Diabetes	0.422 (0.078, 0.766)	0.774	0.005
Female	−0.305 (−0.539, −0.071)	0.776	0.001
RALogMIMS	−3.460 (−4.805, −2.114)	0.788	0.012
TLogMIMS	−0.001 (−0.001, −0.000)	0.790	0.002

Sensitivity Analyses

Repeated cross-validation shows that univariate rankings are relatively stable, particularly for the top predictors (See Supplemental Figure 1, Supplemental Digital Content). For example, M10 calculated using MIMS is the top predictor in all 100 partitions of the data. The next several predictors generally alternate between two ranks, such as TMIMS (either second or third), M10 on the log MIMS scale (either second or third), age (either fourth or fifth), and mobility problem (either fourth or fifth).

Regarding the forward selection procedure, the repeated cross-validation results are remarkably stable, particularly for the threshold δC ≥ 0.01. Supplemental Figure 2 and Supplemental Figure 3 (Supplemental Digital Content) present the results of the forward selection and two-stage forward selection procedure, respectively, for the 100 different partitions of the data. We see that for δC ≥ 0.01, in all but 2 partitions, the model selects the variables presented in the manuscript for the forward selection procedure (Supplemental Figure 2, Supplemental Digital Content). For δC ≥ 0.01 the two-stage forward selection procedure results in RA calculated on the log MIMS being selected in all by 2 partitions, though the traditional risk factors selected are somewhat more variable (Supplemental Figure 3, Supplemental Digital Content). Specifically, either diabetes or overall health is generally selected for inclusion, but not both. Additionally, cigarette smoking is selected for inclusion approximately 50\% of the time. Age and mobility problem are selected in 100\% of cases. Moreover, we see that when forward selection is performed, if a particular accelerometry variable is not selected for inclusion, then a highly correlated variable is often selected in its place (e.g., either RA on the log MIMS scale OR RA on the MIMS scale is selected, but not both.

DISCUSSION

This study provides the first evidence that wrist-worn accelerometer-derived PA measures are more predictive of mortality than traditional risk factors, including age, in a US nationally representative sample. These objective summaries correspond to different domains of PA including volume, fragmentation, and circadian rhythmicity. These findings are consistent with findings in NHANES 2003–2006, which used a hip-worn device, and UK Biobank, which also used wrist accelerometers but in a less diverse, target population. Despite the differences in device placement, target population, and accelerometry data summarization, the main message is reinforced by this new study: objectively measured PA outperforms traditional risk factors in terms of all-cause mortality prediction performance. Collectively these results suggest PA monitoring may provide more sensitive insights into health status and mortality risk than traditional clinical assessments.

Movement and health are intrinsically linked (16). Mobility loss with aging occurs across species and is linked to deterioration of the central and peripheral nervous systems, musculoskeletal systems, and sensory systems (16–18). To this end, health and functional status are often better assessed with physical and cognitive assessment than clinical disease measures (19). The current results support these findings and expand upon them by demonstrating that several measures of free-living physical activity outperform the predictive ability of reported mobility problems and other clinical disease burden. The scientific evidence presented here and in previous work (8,9,20–26) warrants the expansion of wearable monitors in clinical practice, especially in the care of older adults.

Four key strengths of this study are that: 1) the sample is nationally representative of the non-institutionalized US population ages 50 to 80; 2) the study implemented a 24-hour wear time protocol (as compared to a wake-wear protocol used in NHANES 2003–2006); 3) the device was worn at the wrist, a body location that is well tolerated by participants and is associated with relatively high rates of compliance with wear-time protocols, limiting the potential biases associated with differential non-wear time; and 4) our sensitivity analysis of the forward selection procedure presented in the supplementary material (Supplemental Figures 1, 2, and 3, Supplemental Digital Content) shows robustness of key results, particularly the individual and added predictive power of PA variables above and beyond traditional risk factors, as well as the relative stability of variable rankings and the forward selection procedure.

There are several limitations of this study. First, the NHANES study measures participants’ activity cross sectionally. The lack of longitudinal data restricts conclusions to that of associations/correlations rather than causal effects. Indeed, it is possible that the strong association phenotypes measured by wearable accelerometers is explained in whole or in part by the impact of prevalent comorbidities on activity/movement rather than PA directly conferring a protective effect on mortality risk. Other studies have attempted to assess the impact of reverse causality on mortality risk estimates for accelerometer derived PA (27), finding that while associations were attenuated, they remained statistically significant. A recently published study showed that objectively measured physical activity in the NHANES 2003–2006 sample remained statistically significant for at least 12 years. Nevertheless, absent true estimates of causal effects of accelerometer derived measures on mortality, the ability to strongly predict current risk in combination with established risk factors may ultimately yield clinically significant advances in predictive models using a relatively inexpensive and unobtrusive method for monitoring patients. Second, in addition to the lack of a causal interpretation for our findings, the NHANES 2011–2014 participants currently have only short- or mid-term follow up data. This limits the interpretations of our findings to a period of 5–8 years. Third, the study did not consider tobacco use dosage, which may improve the predictive power of smoking in assessing mortality risk. Fourth, although we included many risk factors in our analysis, there may be others that were not included in the list including markers of inflammation, genomics data, and other non-wearable biomarkers such as brain age, to name a few. In addition, the self-reported comorbidity and risk factor data may suffer from measurement error and/or recall bias. Fifth, although we did not find consistent evidence of healthy participant bias (Supplemental Table 1, Supplemental Digital Content), we did find some systematic differences in demographic variables and risk factors. Depending on the sources of these differences, the results presented in this paper may not be generalizable to reflect the health status of the entire US population. Sixth, we excluded participants aged 80 and over due to NHANES 2011–2014 topcoding age. The rankings of predictors and overall predictive accuracy may change in an older sample where age is presumably increasingly predictive of risk of mortality. Seventh, considering an approach to model building which balanced complexity, interpretability, and presentation of results, we did not consider non-linear associations of continuous predictors with mortality aside from BMI, nor did we attempt to model time-varying effects due to potential violations of the proportional hazards assumption. Finally, the current work focuses exclusively on prediction and thus the models omit variables which may be important from a clinical and/or scientific perspective, and may attain statistical significance if included in the model, but do not substantially improve predictive performance. This exclusion of key confounders means that individual effect estimates should be interpreted with caution.

CONCLUSIONS

Despite the limitations of this study, the results presented here add to the literature by highlighting the importance of variables beyond volume of PA, particularly the circadian feature relative amplitude. Moreover, we showed that applying a minute level log transformation of the MIMS units did improve the added predictive value of subject-level summary features, suggesting that summarizing minute level MIMS units using a transformation may be appropriate when creating features for mortality prediction. Most importantly, these results add to the growing body of evidence supporting the use of signals derived from wearable accelerometry in predicting an individual’s current health status and risk of mortality beyond measures which are collected in standard surveys.