Study participants

The current study was conducted as part of two clinical trials examining the effects of unsaturated fatty acids on cardiorespiratory fitness (UFA-Preserved2; NCT03966755) and the role of mitochondrial dysfunction on exercise capacity in patients with HFpEF (MitoP; NCT03960073). All participants provided written informed consent. Ethical approval was provided by the Virginia Commonwealth University Institutional Review Board for both trials (UFA-Preserved2: HM20016253; MitoP: HM20015440). Both studies were carried out at the same institution during overlapping periods and utilized the same measurement protocols for cardiopulmonary exercise testing. The inclusion criteria for both trials were identical, except for the additional criterion of obesity in the UFA-Preserved2 trial. Sixteen adults aged 18 and over with a confirmed clinical diagnosis of stable HF (i.e., New York Heart Association (NYHA) class II–III) and LVEF > 50% documented in the prior 12 months were included in the current study. Of those, 15 participants came from UFA-Preserved2 trial, which has an additional criterion of Body Mass Index (BMI) ≥ 30 kg/m² or total body fat percentage >25% in men and >35% in women. The exclusion criteria differed slightly depending on the risks associated with experimental conditions for each trial. Overall, those who had absolute contraindications to exercise testing according to the American College of Sports Medicine guidelines [32], significant chronic diseases (i.e., angina, uncontrolled arterial hypertension, chronic pulmonary disease, current cancer, etc.), electrocardiography changes (i.e., ischemia or arrhythmias), comorbidity limiting survival, stage V kidney disease, unstable fluid overload, pregnancy, or inability to give informed consent were excluded.

Procedures and measures

Statistical analysis

We first conducted scatter plots to examine the linear relationship between AG-derived and criterion-derived measurements. Total step counts and mean METs were calculated and compared between the measurements (i.e., AG vs. criterion) using a mixed model, with a random intercept accounting for multiple observations within each participant. The correlation coefficients between the AG-derived and criterion-derived measurements were also estimated using a mixed model with a random intercept, according to Hanlett et al [38]. We used the bootstrapping method to calculate the bias-corrected 95% confidence intervals, with 200 bootstrap resamples generated from the observations at the individual level. We then calculated a mean CPET duration (minutes), VO2 ml/kg/min, total steps (counts), and EE estimations (METs) across different PA intensity levels and summarized them for comparison between AG-derived and criterion-derived measurements. PA intensity levels were categorized into light-intensity PA (LPA; 1.50–2.99 METs) and moderate-to-vigorous intensity PA (MVPA; ≥3 METs) based on PA guidelines for Americans [39], using EE estimation data obtained from indirect calorimetry during CPET.

We also performed the equivalences test using the two one-sided t-tests (TOST) to determine if AG-derived SC and EE estimations were significantly equivalent to the criterion measures. Specifically, we compared the mean ratio of SC and EE estimates between AG and criterion measures with the upper and lower limits of 10% equivalence zones (i.e., Ha₁: 0.9< mean ratio and Ha₂: mean ratio<1.11) [37, 40]. The manually counted steps and indirect calorimetry-derived EE were used as criterion measures when creating 10% equivalence zones. Considering the multiple observations within each individual, a mixed model with a random intercept was used to test the mean of AG-derived measures against zero. We determined that PA estimates from AG were significantly equivalent to the criterion measures if the one-sided p-values from both tests were less than 0.05 (i.e., two-sided p-value estimation was divided by 2 to obtain the one-sided p-value when the hypothesized direction has met) [37, 40].

The accuracy of AG-derived SC and EE estimates compared to criterion measures was examined using mean absolute percentage error (MAPE; %), percentage bias (%), mean bias, and the Bland-Altman method. We have determined that MAPE <10% is an acceptable level of accuracy in the current study based on the accuracy standard of wearable PA monitors established by the Consumer Technology Association [41]. In addition, we conducted the modified Bland-Altman plots with a mean bias and 95% limit of agreement (LOA) to assess the agreement between AG-derived and criterion-derived measurements. The mean bias and 95% LOA were calculated using a one-way random effects model, taking into account repeated measures nested within a random factor of the subject. LOA is defined by the true difference between measurements (i.e., AG–criterion) accounting for random variability between- and within-subjects. Additionally, we calculated the 95% confidence intervals of LOA using the method of variance estimate recovery (MOVER), which provided estimates for the maximum limits of lower and upper limits of LOA. The MOVER approach is considered more effective than the conventional delta method with the data involving multiple observations per individual [42]. A detailed description with a SAS macro employing the Bland-Altman method using repeated measures within subjects is followed by Zou [42]. The modified Bland-Altman plots differ from the original method by using the criterion measures on the X-axis against the differences between the two measurements on the Y-axis, which allows for examining the accuracy of AG compared to the criterion measures [43]. The proportional bias was also assessed by examining the linear association between AG-derived and criterion-derived measurements in modified Bland-Altman plots. All statistical analyses were performed using SAS v9.4 (SAS Institute, Cary, NC, USA), and the statistical significance was set at p<0.05.

Step counts

The main finding of this study was that the ankle-worn AG was more accurate in measuring SC in HFpEF patients than the waist-worn AG. The total SC obtained from the ankle-worn AG was significantly similar to manually counted steps and showed lower values of MAPE (<10%), percentage bias, and mean bias with narrow 95% LOA than the waist-worn AG, regardless of PA intensity level. These findings are consistent with previous studies, which have demonstrated that the ankle location is the most accurate for SC detection as compared to the waist location [16, 22, 46]. It is generally considered that the accuracy of wearable devices should perform within MAPE values of less than 10%, indicating a high accuracy [41, 47]. For instance, in adults, Mora-Gonzalez et al. reported a lower MAPE for the ankle location than the waist location overall during the treadmill test (Ankle: 3% vs. Waist: 28%) [46]. Similarly, the ankle-worn AG in our study showed a lower MAPE than the waist-worn AG overall (Ankle: 8.5% vs. Waist: 30.1%). Additionally, in older adults, Korpan et al. found that the ankle location had a lower MAPE than the waist location (Ankle: 2.1% vs. Waist: 23.1%) [16], and Webber and St. John also reported that ankle location had a lower median absolute percentage error than waist location in older geriatric rehabilitation patients (Ankle: 2.5% vs. Waist: 18.9%) [22]. These findings are comparable to those of our study, as the mean age of our study population was over 60 years.

However, some previous studies have reported inconsistent results based on varying walking speeds. For instance, Karaca et al. found that the waist-worn AG was more accurate and had better agreement at higher walking speeds (6, 8, and 10 km/hr; MAPE (%): 1.2, 0.8, and 2.9, respectively) as compared to the ankle-worn AG (6, 8, and 10 km/hr; MAPE (%): 4.9, 47.7, and 50.6, respectively) [20]. In contrast, the ankle-worn AG was more accurate and showed better agreement at lower walking speeds (2 and 4 km/hr; MAPE (%): 12.4 and 1.0%) than the waist-worn AG (2 and 4 km/hr; MAPE (%): 80.0 and 8.3%, respectively) [20]. Although absolute speed measurements during CPET were not utilized in our current study, our findings are partially consistent with the results of Karaca et al. with respect to intensity levels. The ankle-worn AG exhibited agreement with manually counted steps at LPA but underestimated SC at MVPA according to Bland-Altman plots. Conversely, the waist-worn AG significantly underestimated SC at LPA but improved as the intensity level increased. These findings are largely consistent with Bezuidenhout et al., who demonstrated that the ankle-worn AG consistently showed a high level of agreement at overall speeds (0.6–1.4m/s; >94%), while waist-worn AG exhibited poor agreement at lower speeds (0.6–1.0 m/s; 69%) but improved at higher speeds (>1.4 m/s; 94%) [21]. Similar studies examining treadmill exercise tests have also indicated that the waist-worn AG underestimates SC at lower speeds but improves as the speed increases [48, 49]. In the present study, the waist-worn AG had a lower MAPE during MVPA than LPA (LPA: 43.7% vs. MVPA: 16.8%), which is generally consistent with these findings. This is supported by the fact that an AG worn close to the body’s center of mass is more sensitive in measuring locomotion resulting from the ambulatory cycle of both legs during active walking or running [17]. However, additional research is necessary to gather more evidence for the discrepancy in the accuracy between LPA and MVPA.

The current study focused on HFpEF patients, a population known to face challenges in performing absolute high-intensity PA due to reduced exercise capacity [50]. HF patients are typically older individuals who experience limitations in walking as a result of aging [51], muscle depletion [52], and frailty, which is independently associated with the risk of HF in older adults [53]. Ozawa et al. found that HF patients experience a significant decrease in walking speed, even after adjusting age and sex-related factors [51]. Therefore, we suggest that ankle-worn AG, which exhibits less variability and bias across different PA intensity levels in SC measurement, would be more appropriate for this population than waist-worn AG.

Energy expenditure

The current study found that all EE prediction equations showed higher MAPEs overall (>10%) in the HFpEF population. Specifically, the Freedson, refined Crouter, and Santos-Lozano VT/VM equations tended to overestimate EE at LPA but underestimate EE at MVPA when compared to EE derived from indirect calorimetry. However, the Sasaki equation showed the lowest percentage bias overall (0.7%) and at LPA (-1.2%), while the Santos-Lozano VT equation showed the lowest at MVPA (0.2%).

The Freedson equation (1998) was originally developed using the first generation of AG accelerometers (model 7164). Despite the use of an older AG model in the development of this equation, it remains one of the most widely employed equations in PA-related research. However, Grydeland et al. observed that the older AG model (model 7164) tended to overestimate PA outputs compared to the later generation of AG devices [54]. These findings suggest that the Freedson equation may produce less accurate PA estimates when applied to data obtained from the newer AG model. Considering that the Freedson equation showed high MAPE and bias in our study, the later AG model (GT9X) may not be suitable for estimating EE with this equation, at least in adults with HFpEF.

Moreover, studies have reported that the Freedson equation demonstrates accuracy only for PA levels above 1951 counts in young adults and less sensitivity at lower-intensity PA during treadmill tests [22, 23, 55]. Our findings are consistent with these reports, as we observed high MAPE and bias when applying the Freedson equation in the HFpEF population, characterized by lower activity levels and a mean age of over 60 years. The Freedson Combination equation (1998) was developed to overcome the limitation of the Freedson equation by allowing the assessment of lower-intensity activities (≤ 1951 counts) [12, 28]. In the present study, the Freedson Combination equation was the only one that did not exhibit any proportional bias compared to indirect calorimetry, but it showed a wider 95% LOA range compared to other equations. Additionally, this equation still had the highest MAPE overall (28%) and at LPA (40%), with a more significant percentage bias than other equations, indicating large variations and low accuracy for EE estimation. Similarly, Riberio et al. reported that the Freedson Combination equation performed poorly in accurately estimating EE for women with severe obesity requiring clinical treatment [28]. This finding indicates that using this equation to estimate EE may not be advisable for HFpEF patients.

The Crouter equation (2010) was initially developed as a two-regression model in 2006 to differentiate various locomotion activities (i.e., walking/running and lifestyle behaviors) based on the coefficient variations in activity counts and refined in 2010 to improve EE estimation by reducing misclassification of activity types (i.e., walking or running) in free-living settings [24]. In the current study, the refined Crouter equation tended to underestimate EE at MVPA during treadmill testing in the HFpEF population. However, this equation showed varied results in previous studies depending on the study settings and population characteristics. For example, in free-living settings, Crouter et al. reported that it tended to overestimate EE in younger adults [26], but Aguilar-Farias et al. found that it performed more accurately in older adults, particularly at MVPA [14]. In laboratory settings, Rothney et al. found that this equation significantly overestimated total EE compared with a whole-room indirect calorimeter in younger adults [56]. Since the refined Crouter equation was developed based on free-living lifestyle behaviors [24], more evidence of this equation’s validity is needed based on the treadmill exercise test. Therefore, further studies are required to validate this equation for clinical populations based on laboratory settings.

Sasaki et al. (2011) developed the equation using AG’s VM activity count data to establish cut points and PA intensity classification in young adults. Although few studies have examined the validity of the Sasaki equation, Aguilar-Farias et al. found that this equation showed agreement and better accuracy than other equations when compared with indirect calorimetry in free-living settings among older adults [14]. In our study, the Sasaki equation tended to overestimate EE during MVPA but was the only equation to demonstrate significant equivalence to the criterion EE across all PA intensity levels. The Sasaki equation exhibited better accuracy with the lowest bias overall (mean bias = 0.02; percentage bias = 0.65%) and at LPA (mean bias = -0.07; percentage bias = -1.19%). However, we observed inconsistent outcomes regarding the accuracy of this equation, as the MAPE at LPA was comparatively higher at approximately 20% compared to other equations. MAPE is one of the straightforward methods used to examine accuracy and bias, but it may result in undefined or infinite values when dealing with values close to zero, causing misinterpretation [57, 58]. Compared to MAPE, bias represents the mean error that determines the direction of error, allowing researchers to interpret whether there was an overestimation or underestimation [59]. Moreover, the Bland-Altman plot of the Sasaki equation showed agreement with indirect calorimetry, especially between 2–3 METs, which was comparable with the results of percentage bias at LPA. Therefore, our study concluded that the Sasaki equation generally showed better accuracy, particularly at LPA in patients with HFpEF.

Santos-Lozano et al. (2013) developed two equations to differentiate cut points and classify PA intensities for different age groups, using mean activity counts from three axes (Santos-Lozano VT) and VM outputs from the AG (Santos-Lozano VM) [25]. Similar to other equations, these equations have yielded accurate EE prediction across various PA intensities in young adults in laboratory settings, but their accuracy is limited for older adults [14, 25]. In the current study, we observed that these equations showed the lowest MAPEs (<10%) and biases only at MVPA in HFpEF patients (%bias: Santos-Lozano VT = 0.16%; Santos-Lozano VM = -1.34%). Indeed, based on the Bland-Altman plots, the Santos-Lozano equations demonstrated agreement with the criterion measure around the 3 METs level, but they significantly underestimated EE as the intensity increased. This result may have been influenced by the current study’s categorization of PA intensity, which combined moderate and vigorous activity into one category of MVPA. Given that our study population consisted of clinical HFpEF patients, who are often limited in their ability to perform activity at the typical absolute vigorous-intensity threshold, we suggest that future studies consider classifying PA intensity levels differently, taking into account the characteristics typically observed in this clinical population.

Strengths and limitations

To the best of our knowledge, this is the first study to examine the criterion validity of the AG accelerometer among patients with HFpEF. Thus, our study may have meaningful implications for the efficient use of AG accelerometers in this clinical population. Additionally, our study used the later generation of the AG, the GT9X model, which is equipped with triaxial sensors and improved algorithms capable of providing more comprehensive and accurate activity measurements than the earlier models (e.g., Model GT1M and 7164), which was limited to a uniaxial feature [12]. Moreover, our study was conducted in a well-controlled laboratory setting using treadmill exercise testing and compared the accuracy of AG-derived measurements to the gold standard criterion-derived measurements across different PA intensity levels. This controlled environment allowed for a more standardized assessment of the validity of the AG-GT9X in estimating PA parameters.

Nevertheless, there are several limitations to be acknowledged in this study. First, the majority of participants were female, with only one male included, which limits the generalizability of our findings to the entire HFpEF population, particularly for male HF patients. Future studies with large sample sizes and more equal distribution of both genders are required to enhance the generalizability of the results in this clinical population. Another limitation is the reliance on standard resting EE (3.5 mL/kg/min) instead of measured resting EE for the calculation of METs values, which is considered a more accurate indicator of resting metabolic rate for the calculation of METs values [60]. A recent study found that the discrepancy between measured and estimated resting EE becomes more pronounced as BMI and age increase among individuals with HFpEF [61]. Considering our study population consists of adults with HFpEF whose mean BMI is greater than 38 and whose mean age is above 60 years, it is advisable to measure resting EE before the test to obtain more reliable results. Therefore, future studies should incorporate the measured resting EE to ensure the validity of the findings. In addition, the present study categorized PA levels using the general PA intensity classification (i.e., LPA: 1.5–2.99 METs; MVPA ≥ 3 METs) [39], which may not accurately reflect the PA intensity relative to the low fitness levels of the HFpEF population. Future studies should consider utilizing relative PA intensity cut-points developed specifically for HFpEF patients to improve the accuracy of PA level classification [62]. Among several different EE estimation equations, we used an adapted version of the refined Crouter equation, particularly developed for walking/running activities, using both 10-sec and 60-sec epoch data; yet this modified cut-point scaling may not adequately account for the magnitude or coefficient variation of the counts during the treadmill test [63]. Further research is warranted to explore the use of coefficient variation in determining the most accurate refined Crouter equation for activities rather than relying on a single equation. Moreover, since this study was designed based on a laboratory setting using treadmill exercise testing, the validity of the AG in free-living conditions may differ. Given that a few EE estimation equations used in this study were originally developed for different types of activities (i.e., free-living activities), future research should incorporate a range of activities in addition to CPET and aim to explore the validity of the AG in real-life free-living conditions. Lastly, we could not consider the impact of absolute walking speed on the accuracy of the AG during exercise testing in this study due to the colinear increases in treadmill grade. Instead, we calculated the PA intensity level using EE estimation obtained during CPET and assessed the accuracy of AG at the corresponding PA intensity levels. Still, we must acknowledge that this approach is not specifically tailored for CPET data. Further investigations should explore the impact of absolute walking speed on the accuracy of the AG to better understand its validity across different exercise intensities during CPET, while considering real-world variability.