Cardiovascular mortality risk prediction using objectively measured physical activity phenotypes in NHANES 2003–2006

Abstract

Increased physical activity (PA) has been associated with a decreased risk of cardiovascular disease (CVD) and mortality. However, most previous studies use self-reported PA instead of objectively measured PA assessed by wearable accelerometers. To the best of our knowledge, there have not been studies that quantified the univariate and multivariate ability of objectively measured PA summaries to predict the risk of CVD mortality. We investigate the ability of objectively measured PA summary variables to predict CVD mortality: as individual predictors, as part of the best multivariate model incorporating traditional predictors, and as additions to the best multivariate model using only traditional CVD predictors. Data were collected in the National Health and Nutrition Examination Survey 2003–2006 waves for US participants aged 50–85. The predictive ability was measured using Concordance, sometimes referred to as the C-statistic. Specifically, we calculated 10-fold cross-validated concordance (CVC) in survey-weighted Cox proportional hazard models. The best univariate predictor of CVD mortality was total activity count (outperformed age). In multivariate models, two of the eight predictors identified using the improvement in CVC threshold of 0.001 were PA measures (CVC = 0.844). The best model without physical activity (7 predictors) had CVC of 0.830. The addition of PA measures to the best traditional model was significantly better at predicting CVD mortality (P < 0.001). Accelerometer-derived PA measures have excellent cardiovascular mortality prediction performance. Wearable accelerometers have a potential for assessment of individuals’ CVD mortality risks.

1. Introduction

Cardiovascular disease (CVD) is among the leading causes of death in the US adult population (Go et al., 2014; Heron and Anderson, 2016; Kochanek et al., 2019; Kochanek et al., 2020). Physical activity (PA) plays an important role in decreasing the risk of CVD (Joseph et al., 2017; Li and Siegrist, 2012; Martinez-Gomez et al., 2019). There is extensive research linking the association between self-reported PA and the risk of developing CVD (Fang et al., 2019; McGuire et al., 2009; Georgousopoulou et al., 2016; Young et al., 2014), CVD mortality (Evenson et al., 2016; Kraus et al., 2019), as well as all-cause mortality (Evenson et al., 2016; Jeong et al., 2019). An increase in PA that corresponds to 500 metabolic equivalent (MET) min/week reduced the risk of mortality by 14% for those who had a history of CVD and 7% for those without a history of CVD (Jeong et al., 2019). According to Evenson et al. (2016) (Evenson et al., 2016), adults over 40 years old who self-reported meeting the PA guidelines for Americans (Piercy et al., 2018) had a significant reduction of CVD mortality risk. However, these results rely on self-reported PA, which has been shown to have recall and social desirability bias (Luke et al., 2011; Wanner et al., 2017). To address this limitation, methods that measure a person’s fitness level can be used to predict CVD mortality risk (Wickramasinghe et al., 2014). However, the test requires the use of a treadmill and speed/incline modifications every minute and indirect calorimetry for reliable oxygen uptake (VO₂) peak testing (Keteyian et al., 2008), which makes it difficult to perform for persons without access to the equipment and is less reliable due to self-administration.

Accelerometers provide objective measures of PA and provide a viable, objective alternative to self-reported PA. A major advantage of accelerometers is easy self-administration, which leads to reliable use in both personal and large populations. Recent studies explored the association between CVD and accelerometer-based measures of PA (Luke et al., 2011; Dempsey et al., 2020; Andersson et al., 2015), death attributed to CVD (Evenson et al., 2016) and to all causes (Saint-Maurice et al., 2018). Evenson et al. (2016) (Evenson et al., 2016) and Luke et al. (2011) (Luke et al., 2011) based their studies on the National Health and Nutrition Examination Study (NHANES), which collects data on the US population. Dempsey et al. (2020) (Dempsey et al., 2020) analyzed the European Prospective Investigation into Cancer and Nutrition-Norfolk study (EPICNN), a population-based data set of adults in Norfolk, UK. Andersson et al. (2015) (Andersson et al., 2015) analyzed the Framingham Heart Study, a US multigenerational data set spanning from 1948 to the present. They demonstrated that individuals with higher levels of PA had a lower risk of incident CVD (Dempsey et al., 2020), a lower risk of mortality attributed to CVD (Evenson et al., 2016), and a reduced risk of all-cause mortality (Saint-Maurice et al., 2018). These studies concentrated on the association between CVD mortality and a small number of common accelerometer-derived PA measures, such as total activity count (TAC), moderate to vigorous physical activity (MVPA), light intensity physical activity (LIPA), sedentary/sleep time (ST), and wear time (WT). However, these summaries may not be sufficient to capture the full complexity of daily minute level PA (Di et al., 2017). To the best of our knowledge, there have not been any studies which focused on quantification of individual and combined CVD mortality risk prediction using a large number of objective PA summaries and traditional risk factors.

To address this gap in the literature, we explored the prediction performance of 12 accelerometer-derived PA and 16 traditional measures for CVD linked mortality prediction in NHANES 2003–2006. We hypothesized that PA measures will be highly predictive of short to medium term CVD mortality and adding PA to the model with traditional predictors will increase predictive ability. The main goals of the current study were: (Go et al., 2014) to compare and rank the performance of individual PA and traditional measures to predict CVD mortality using Cox proportional hazards; (Heron and Anderson, 2016) to identify the best combination of traditional and PA predictors for quantifying the CVD mortality risk; and (Kochanek et al., 2019) to quantify the improvement in CVD mortality prediction by adding the PA measures to the best traditional CVD mortality measures. A major advantage of our study is the NHANES data used a large and representative sample of the US non-institutionalized civilian population. Thus, this study is crucial for CVD risk assessment in free-living conditions and generalizable to the US population.

2. Methods

2.1. Study population

The National Health and Nutrition Examination Survey (NHANES) publicly available data collected by Centers for Disease Control and Prevention (CDC) to measure the vital health statistics of the United States population was used in this study. There was a total of 14,631 study participants with accelerometry data in the 2003–2004 and 2005–2006 cohorts of NHANES. We limited our analysis to participants who were at least 50 and not >85 years of age (3772 participants). Study participants were further excluded for missing socio-demographic information, dietary indicators, health indicators, and PA measures as follows: (Go et al., 2014) missing level of education (6 participants); (Heron and Anderson, 2016) missing body mass index (BMI) values (35 subjects); (Kochanek et al., 2019) missing mortality data (7 participants) and any combination of systolic blood pressure, total cholesterol, or HDL cholesterol measurements (293 participants); and (Kochanek et al., 2020) less than three days of at least 10 h of wear time for the accelerometer (517 participants). The final sample size contained 2987 study participants. For these participants, there were 161 deaths attributed to cardiovascular causes with an average of 9.9 years of follow-up. Study participants with mortality events for reasons other than cardiovascular (N = 631) were treated as having right-censored events.

2.2. Traditional mortality predictors

The following 16 traditional mortality predictors have been included in the analysis. The socio-demographic variables were age at the time of the physical exam, sex (male/female), education level (less than high school, high school, and more than high school), and race (black, white, Mexican American, other Hispanic, and other). Dietary indicators were body mass index category (underweight, normal, overweight, and obese) and alcohol consumption status (non-drinker, moderate drinker, heavy drinker, and missing alcohol status). Health indicators were smoking status (never, former, and current), diabetes (yes/no), coronary heart disease (yes/no), cancer (yes/no), congestive heart failure (yes/no), stroke (yes/no), mobility problems (yes/no), total cholesterol level (mg/dL), HDL cholesterol level (mg/dL), and systolic blood pressure (mm Hg).

2.3. Accelerometry derived predictors

PA was measured by a hip-worn accelerometer (ActiGraph AM-7164). The following 12 physical activity derived variables were used in the analysis. Total wear time (WT), total activity count (TAC), total log activity count (TLAC) or log(1 + activity count), minutes of sedentary/sleep time (ST), minutes of light intensity physical activity (LIPA), and minutes of moderate to vigorous physical activity (MVPA). The ratio of bouts to the total time was used to estimate the following: sedentary to active transition probability (SATP) and active to sedentary transition probability (ASTP). Traditional measures associated with circadian rhythms were defined as the average log activity during the 10 most active hours of the day (M10), the average log activity during the least five active hours of the day (L5), and the relative amplitude (M10-L5)/(M10 + L5). The daily measures above were first calculated within valid wear days (see Supplementary Materials) and then averaged across days.

In addition to these PA summaries, data driven measures were derived using functional principal component analysis (fPCA) (Leroux et al., 2019). Each of the first six principal components (PCs) was summarized using the means and standard deviations of their individual scores. The 12 resulting PC measures were then added to the socio-demographic, dietary, and health measures and used in a backward selection process. At the final step of the backward selection process, the standard deviation for the sixth PC (SD on PC 6) remained in the model (P < 0.001). A surrogate variable was developed for ease of interpretation (Surrogate for SD on PC 6). For all subsequent analyses, this surrogate (defined in the Supplementary Materials) was included among the candidate predictors.

2.4. Statistical analysis

The univariate complex survey weighted 10-fold cross-validated concordance (CVC) of each of the 28 potential predictor variables from a Cox proportional hazards (PH) regression was used to rank their predictive performance for cardiovascular mortality. The concordance statistic, sometimes referred to as the C-statistic, measures the predictive accuracy of the prognostic test in survival models by counting the fractions of all the ordered time pairs in which the risk score (here, from Cox regression) is concordant with the actual outcomes data (Harrell Jr. et al., 1996; Harrell Jr. et al., 1982). The function svycoxph from the survey package in R was used for the analysis. Mortality attributed to causes other than cardiovascular were treated as right censored events. To support parsimony, all forward selection analyses used two separate selection criteria that required an increase in CVC of at least 0.01 or 0.001 (ΔCVC ≥ 0.01 or ΔCVC ≥ 0.001, respectively) to continue the forward selection process, with a threshold of 0.01 resulting in a more parsimonious model (stricter inclusion criteria). Sensitivity analysis (see Supplemental Information, Tables S2 and S3) was performed on each of the forward selection methods. Development of the two stage models is described in the Supplementary Materials.

3. Results

Table 1 provides the unadjusted number of alive and deceased study participants (as of December 31, 2015) along with complex-survey adjusted participant characteristics by mortality status for the 28 predictors ranked by their univariate cross-validated concordance (CVC) in a Cox PH regression. The total number of deaths due to cardiovascular events was 161, the number of individuals censored due to death for causes other than cardiovascular events was 631, and the number of surviving individuals was 2195. The mean age of decedents was 78.39 years, which was on average 0.48 years younger than those who did not die from cardiovascular causes.

Table 1.

Population characteristics at follow-up of the univariate predictors ordered by their cross-validated concordance (CVC) from survey-weighted Cox regressions.

Rank	Predictor	CVC	Survivors and decedents from non-cardiovascular causes (n = 2826)	Decedents from cardiovascular causes (n = 161)
1	TAC (mean (SD))	0.7513	225,954.09 (114,750.85)	146,618.67 (89,478.69)
2	Age (mean (SD))	0.7476	62.87 (9.47)	71.09 (10.04)
3	MVPA (mean (SD))	0.7336	15.93 (18.04)	7.58 (12.31)
4	TLAC (mean (SD))	0.7265	2843.72 (704.07)	2324.18 (655.03)
5	LIPA (mean (SD))	0.7210	1100.24 (104.69)	1174.58 (98.73)
6	Sedentary/sleep time (mean (SD))	0.7209	1099.16 (104.80)	1173.48 (98.82)
7	M10 (mean (SD))	0.7077	3.64 (0.82)	3.06 (0.87)
8	ASTP (mean (SD))	0.6944	0.29 (0.08)	0.35 (0.10)
9	SATP (mean (SD))	0.6784	0.08 (0.02)	0.07 (0.02)
10	Mobility problem = any difficulty (%)	0.6760	731.8 (25.5)	66.8 (57.5)
11	Surrogate for SD on PC 6 (mean (SD))	0.6670	0.70 (0.27)	0.57 (0.26)
12	Education (%)	0.6352
	Less than high school		525.1 (18.3)	43.0 (37.0)
	High school		777.7 (27.1)	34.9 (30.0)
	More than high school		1567.9 (54.6)	38.3 (33.0)
13	Systolic blood pressure (mean (SD))	0.6004	130.94 (20.16)	138.82 (26.14)
14	Total cholesterol (mean (SD))	0.5977	207.08 (41.33)	194.22 (40.88)
15	Gender (% female)	0.5910	1551.2 (54.0)	40.8 (35.1)
16	Diabetes (% yes)	0.5904	364.0 (12.7)	36.5 (31.4)
17	Coronary heart disease (% yes)	0.5819	195.1 (6.8)	28.3 (24.4)
18	Wear time (mean (SD))	0.5793	880.33 (124.95)	856.80 (135.41)
19	HDL cholesterol (mean (SD))	0.5697	56.21 (16.59)	52.07 (15.18)
20	Congestive heart failure (% yes)	0.5694	122.6 (4.3)	20.1 (17.3)
21	Alcohol consumption (%)	0.5583
	Non-drinker		1056.0 (36.8)	56.8 (48.8)
	Moderate drinker		1524.3 (53.1)	49.0 (42.1)
	Heavy drinker		207.2 (7.2)	7.4 (6.4)
	Missing alcohol		83.3 (2.9)	3.1 (2.6)
22	Stroke (% yes)	0.5485	120.8 (4.2)	16.2 (13.9)
23	Cigarette smoker (%)	0.5410
	Never		1326.9 (46.2)	40.3 (34.7)
	Former		1057.3 (36.8)	50.3 (43.3)
	Current		486.6 (17.0)	25.6 (22.0)
24	Cancer (% yes)	0.5311	455.0 (15.8)	25.0 (21.5)
25	Body mass index (BMI) (%)	0.5085
	Underweight		31.5 (1.1)	1.3 (1.1)
	Normal		768.7 (26.8)	32.6 (28.1)
	Overweight		1096.0 (38.2)	37.1 (32.0)
	Obese		974.6 (33.9)	45.1 (38.8)
26	Relative amplitude (mean (SD))	0.4905	0.98 (0.06)	0.97 (0.08)
27	L5 (mean (SD))	0.4753	0.05 (0.13)	0.04 (0.13)
28	Race (%)	0.4517
	Mexican American		116.1 (4.0)	4.9 (4.2)
	Other Hispanic		63.3 (2.2)	4.0 (3.4)
	Black		275.6 (9.6)	9.7 (8.4)
	White		2301.4 (80.2)	92.3 (79.5)
	Other		114.4 (4.0)	5.3 (4.6)

The unadjusted annual number of deaths after the initial physical exam that were attributed to cardiovascular events for 12 years were 10, 13, 16, 12, 15, 14, 18, 17, 14, 18, 7, and 7, respectively. Overall person-time CVD death rate rate: 0.46%. Year 1: 0.34%, year 2: 0.44%, year 3: 0.54%, year 4: 0.41%, year 5: 0.51%, year 6: 0.48%, year 7: 0.62%, year 8: 0.59%, year 9: 0.49%, year 10: 0.63%, year 11: 0.25%, year 12: 0.25%.

Among all variables in the analysis, the most predictive as measured by univariate CVC was the total activity count (TAC). TAC alone had a concordance of CVC = 0.751 between the predicted and the observed outcomes. Age was the second most predictive variable (CVC = 0.748). Of the top 10 univariate CVC (0.751–0.676), eight were PA related. Among the predictive measures ranked from 11 to 20 (CVC between 0.667 and 0.569) the PA measures were the surrogate for SD on PC 6 (rank =11) and wear time (rank =18). The two remaining PA measures were the circadian measures of relative amplitude (rank = 26) and L5 (rank = 27) with a CVC <0.5. Non-PA related measures, such as stroke, cancer, diabetes, congestive heart failure, and coronary heart disease that ranked lower in CVC had a small number of decedents, made up a small percentage of the total sample, and/or had a small difference in percentage between the decedent group and the survivor and decedents from non-cardiovascular causes group.

Table 2 contains the results of a forward selection from the 28 possible predictors. Two models were selected using separate selection criteria for a minimum increase in CVC at each step: 1) ΔCVC ≥ 0.01, and 2) ΔCVC ≥ 0.001. The first model had a CVC of 0.823 and the variables selected were TAC (CVC = 0.7513), age (ΔCVC = 0.0349), gender (ΔCVC = 0.0218) and mobility problem (ΔCVC = 0.0149). The age (0.069 ± 0.0365) and mobility problems (0.863 ± 0.3715) effects were positive and significant, indicating that older adults and adults with any mobility difficulty are more likely to experience mortality due to cardiovascular events. Both TAC (−0.638 ± 0.4035) and gender (−1.136 ± 0.347) effects were negative and significant indicating that increased TAC and being female are associated with reduced likelihood of mortality due to cardiovascular events.

Table 2.

Forward selection results from a group of predictors including PA measures using two separate selection criteria for the improvement in cross-validated concordance (ΔCVC). The variables are listed in the order of their inclusion into the model along with the cumulative 10-fold cross-validated concordance.

Selection criteria: ΔCVC ≥ 0.01
Variable	Cumulative CVC	ΔCVC	$\widehat{\beta }\pm 2SE\left(\widehat{\beta }\right)$
TAC	0.7513	0.7513	−0.638 (−1.041,−0.234)
Age	0.7862	0.0349	0.069 (0.032, 0.105)
Gender (female)	0.8080	0.0218	−1.136 (−1.483,−0.789)
Mobility problem	0.8229	0.0149	0.863 (0.492, 1.235)
Selection criteria: ΔCVC ≥ 0.001
Variable	Cumulative concordance	ΔCVC	$\widehat{\beta }\pm 2SE\left(\widehat{\beta }\right)$
TAC	0.7513	0.7513	−0.413 (−0.789,−0.037)
Age	0.7862	0.0349	0.064 (0.026, 0.103)
Gender (female)	0.8080	0.0218	−1.118 (−1.447,−0.789)
Mobility problem	0.8229	0.0149	0.696 (0.342, 1.049)
Surrogate for SD on PC 6	0.8294	0.0064	−1.435 (−2.482,−0.388)
Education	0.8386	0.0092
Less than high school			0.326 (−0.213, 0.864)
High school			Ref.
More than high school			−0.336 (−0.828, 0.157)
Congestive heart failure (yes)	0.8422	0.0036	0.754 (0.299, 1.209)
Diabetes (yes)	0.8441	0.0019	0.612 (0.263, 0.961)

The second model had a CVC of 0.844 (using the ΔCVC ≥ 0.001 criterion) and, in addition to the predictors from the first model, included the surrogate for SD on PC 6 (ΔCVC = 0.0064), education (ΔCVC = 0.0092), congestive heart failure (ΔCVC = 0.0036), and diabetes (ΔCVC = 0.0019). The first four variables selected in the second model coincide with first model. TAC (−0.413 ± 0.376), gender (−1.118 ± 0.329), surrogate for SD on PC 6 (−1.435 ± 1.047) and more than high school education (−0.366 ± 0.4925) have negative estimates, which indicates that an increase in these variables corresponds to a reduced risk of cardiovascular mortality. The PA measures in the second model are TAC and surrogate SD for PC 6, and both are significant. Age (0.064 ± 0.0385), mobility problem (0.696 ± 0.3535), education less than high school (0.326 ± 0.5385), congestive heart failure (0.754 ± 0.455), and diabetes (0.612 ± 0.349) have positive estimates. Age, gender, mobility problem, congestive heart failure, and diabetes are the significant non-PA measures in the model.

To quantify the increase in predictive performance associated with the inclusion of accelerometer-derived PA measures in a previously established model using only traditional (non-PA) predictors, a two-stage forward selection process was implemented. Table 3 contains the two-stage forward selection results for models using the same two separate selection criteria (ΔCVC ≥ 0.01 and ΔCVC ≥ 0.001). The first stage selected from non-PA variables, and the second stage selected additional PA variables meeting the corresponding selection criteria. For the model selected using ΔCVC ≥ 0.01, age (CVC = 0.7476), mobility problem (ΔCVC = 0.0401), and gender (ΔCVC = 0.0258) were selected during the first stage. The surrogate for SD on PC 6 was selected in the second stage under this criterion. The resulting model has a CVC of 0.827, which is 1.7% higher than the first stage model without a PA variable. The addition of the PA variable significantly improved the model (P < 0.001). All variables selected according to the criterion ΔCVC ≥ 0.01 were statistically significant (P < 0.001).

Table 3.

Forward selection in 2 stages using concordance as selection criteria. The first stage selects the non-PA variables using forward selection and the 2nd stage adds the PA variables to the first stage variables and does forward selection.

Selection criteria: ΔCVC ≥ 0.01
Variable	Cumulative concordance	ΔCVC	$\widehat{\beta }\pm 2SE\left(\widehat{\beta }\right)$
Age	0.7476	0.7476	0.084 (0.052, 0.116)
Mobility problem	0.7877	0.0401	1.090 (0.624, 1.557)
Gender (female)	0.8134	0.0258	−1.110 (−1.441, −0.778)
Surrogate for SD on PC 6	0.8272	0.0137	−1.856 (−2.805, −0.908)
Selection criteria: ΔCVC ≥ 0.001
Variable	Cumulative concordance	ΔCVC	$\widehat{\beta }\pm 2SE\left(\widehat{\beta }\right)$
Age	0.7476	0.7476	0.064 (0.026, 0.101)
Mobility problem	0.7877	0.0401	0.781 (0.396, 1.166)
Gender (female)	0.8134	0.0258	−1.114 (−1.435,−0.794)
Diabetes (yes)	0.8222	0.0088	0.655 (0.273, 1.038)
Education	0.8286	0.0064
Less than high school			0.330 (−0.219, 0.879)
High school			Ref.
More than high school			−0.328 (−0.824, 0.168)
Congestive heart failure (yes)	0.8303	0.0017	0.791 (0.345, 1.237)
Systolic blood pressure	0.8314	0.0011	0.011 (0.002, 0.019)
Surrogate for SD on PC 6	0.8432	0.0118	−1.437 (−2.467, −0.408)
TLAC	0.8450	0.0018	−0.278 (−0.535, −0.022)

The two-stage model that used the selection criteria ΔCVC ≥ 0.001 has a CVC of 0.845. The first three variables selected are the same as those selected using the ΔCVC ≥ 0.01 criteria (age, mobility problem, and gender). The additional variables selected in stage one included diabetes (ΔCVC = 0.0088), education (ΔCVC = 0.0064), congestive heart failure (ΔCVC = 0.0017), and systolic blood pressure (ΔCVC = 0.0011). The second stage selection included the surrogate for SD on PC 6 (ΔCVC = 0.0118) and TLAC (ΔCVC = 0.0018), which resulted in a significant improvement over the first stage model (P < 0.001).

4. Discussion

Our study of the US non-institutionalized civilian population provides novel evidence that objective accelerometer-derived PA measures are highly predictive of CVD mortality. Specifically, accelerometer-derived PA measures were eight out of the top 10 most predictive variables of CVD mortality in single variable regression (age ranked second and self-reported mobility problems ranked tenth). In multivariate models PA measures were among the strongest predictors of CVD mortality. Moreover, even after selecting the most predictive traditional CVD risk factors, inclusion of PA measures resulted in statistically significant increases in predictive performance. These results provide novel predictors of CVD mortality risks for any individual over the age of 50 based on a few easy-to-measure PA summaries obtained from body-worn accelerometers.

Significant associations between PA and CVD mortality have been reported using both self-reported PA (Katzmarzyk et al., 2009; Hamer et al., 2012; Nocon et al., 2008) and accelerometer data (Evenson et al., 2016; Dempsey et al., 2020; Schmid et al., 2015). The current study confirms these previous findings and further ranks the CVD mortality prediction performance using cross validated concordance for wider set of objective PA day-level summaries derived from body-worn accelerometers. Another innovation of the current paper is the focus on rigorous quantification of CVD mortality prediction performance using cross-validated concordance (CVC) compared to other studies that focused primarily on association testing. Furthermore, in contrast with the most widely used in clinical practice 10-year Framingham score for the total cardiovascular risk (e.g., coronary artery disease, stroke, peripheral vascular disease, coronary heart failure, mortality) in the population of mostly healthy individuals (Lloyd-Jones, 2010; Wilson et al., 1998), our study concentrates on the time-to-death risk estimation of CVD death among both healthy and individuals with prior cardiovascular risk. While many risk models focused on the total fatal and non-fatal cardiovascular events in mostly healthy individuals over 40 years old, we studied only fatal events and thus selected the higher-risk population above 50 years old, for which the majority of CVD events were observed.

Eight of the 12 accelerometer derived PA measures were ranked in the top 10 univariate predictors as measured by 10-fold CVC and outperformed the traditional CVD mortality risk factors (CVC between 0.678 and 0.751). Many of the PA measures have similar predictive performance, which is likely due to their high mutual correlations (Leroux et al., 2019; Smirnova et al., 2020). Similar results were reported for all-cause mortality in NHANES for five-year (Leroux et al., 2019; Smirnova et al., 2020) and for one- to five- year time horizons (Tabacu et al., 2020). In the UK Biobank cohort, Leroux et al., (2021) (Leroux et al., 2021) found that nine out of the top 10 strongest univariate predictors of all-cause mortality were accelerometer-derived PA measures as measured by CVC. These results provide novel insights in the context of CVD mortality prediction.

An important finding of this study is that we built a combined model that includes TAC, age, gender, and mobility problems and is highly predictive of CVD mortality with a CVC of 0.823. Previous accelerometer-based studies (Evenson et al., 2016; Dempsey et al., 2020) focused on causal associations between PA measures and CVD mortality quantified by hazard ratios, which were adjusted using a pre-established set of traditional predictors. Our study identifies specific risk factors for CVD mortality prediction, along with the direction and strength of their CVD mortality prediction performance. We further demonstrate that physical activity is improving the predictive ability of the CVD prediction model when known biological predictors of CVD death (Gender, Diabetes, Smoking, systolic blood pressure, total cholesterol, and BMI) are forced in the model (Table S1 in the Supplementary Materials). The predictive performance is improved (CVC = 0.844, an increase of 2.6%) by including the surrogate for SD on PC6, education, congestive heart failure, and diabetes. Due to the increased robustness of the more parsimonious model and the likelihood that the less parsimonious model’s improvement of 2.6% is not clinically significant, we recommend using the more parsimonious model that includes TAC, age, gender, and mobility problem. Thus, a simple CVD mortality risk score can be calculated based on PA measures derived from the hip-worn ActiGraph accelerometer coupled with traditional predictors.

The current study provides the first evidence that PA measures provide substantial additional predictive performance to a model based solely on traditional predictors. These findings provide evidence that PA measures obtained using inexpensive wearable technology could be used to improve individuals’ assessment of CVD mortality prediction and improve cardiovascular health assessment. These findings are the first step towards future prospective CVD mortality risk assessment studies to improve accuracy of currently used 10-year total and CVD-specific mortality scores, such as the Framingham, SCORE or JBS3 risk models (Lloyd-Jones, 2010; Piepoli et al., 2016; Conroy et al., 2003; Board, 2014).

Among the limitations of this study is that we have not explored covariate interactions, non-linear effects, or time-dependent effects in multivariate models. Our goal was to provide a an interpretable model for CVD mortality prediction could inform future prospective validation studies to develop CVD mortality risk scores to improve the accuracy of currently used CVD mortality risk scores by including PA measures. Our study included individuals with prior cardiovascular disease and other comorbidities as our goal was to build a mortality prediction model relevant to the general population. This approach, in conjunction with the cross sectional nature of the data, limits casual interpretations related to the prospective association between PA and mortality. In particular there is a risk that observed associations are partly driven by reverse causality. Moreover, the prediction models included self-reported mobility problems, which has been shown to be an effect modifier of the association between physical activity and mortality. However, the goal of our study was not to understand the causal effect, but rather to build a prediction framework for the risk of CVD mortality for development risk scores (similar to Framingham, etc) that could be used in clinical and/or research settings. While the significance of physical activity for mortality prediction presented in this study is likely generalizable to the broader class of wearable accelerometers the model coefficients and Concordance estimates derived for the mortality prediction are specific to the activity counts derived from the ActiGraph hip-worn accelerometer (Leroux et al., 2021). Furthermore, CVD mortality risk study is not the primary goal of the observational NHANES study, and the cause of death is derived from the mortality register. Thus, there is a potential for misclassification of the cause of death. Another limitation is that several PA measures are somewhat correlated with age (0.014<|r|<0.431), which may partially account for their success in predicting CVD mortality. In addition, many PA measures are highly correlated with each other (0.722 ≤ |r| ≤ 0.999), which explains why in multivariate models only one or two PA measures are selected. This may also explain why the specific PA variables selected in Table 2 may differ depending on which data is used for training (see Supplemental Information, Tables S3 and S4). However, in our sensitivity analysis study, every model contained at least one PA measure, which substantially increased the predictive performance of the models.

Despite these limitations, the current study provides novel insights into the CVD mortality prediction performance of objectively measured PA summaries. Simple and easy-to-implement models are provided to estimate CVD mortality risks with a CVC as high as 0.845. Extensive evaluation of these model improvements over existing cardiovascular risk score assessment models and extension to consumer grade wearable devices is the topic for future research. These results support the importance of wearable devices to quantify CVD mortality risk and the significance of including PA assessment measures in currently used CVD mortality risk assessment scores.

Data availability

Data will be made available on request.