In this study we determined the validity the two-regression model (5
) to predict EE of PA from body movements recorded by the ActiGraph GTM1 accelerometer. We found good agreement between the average daily EE predicted by the filtered two-regression model and EE measured using both a whole-room indirect calorimeter and free-living DLW in a group of adults with wide range of age (20–67 yrs) and BMI (19–52 kg/m2
). However, the accuracy of the model was shown to have a large inter-individual variability.
To our knowledge this is the first reported study that independently validated the two-regression model proposed by Crouter et al
) using whole-room indirect calorimetry and DLW techniques. The model over-estimated the group average total daily EE by approximately 10.15 ± 11.42% when compared to EE measured by the room calorimeter and by 5.97 ± 16.0% when measured by the DLW. Our data also showed that the model underestimated the time spent in the light (1.5-3.0 MET) PA intensity and overestimated time spent in the moderate (3.0-6.0 MET) PA intensity observed in 29 of 33 our study participants. This suggests that the model may consistently overestimate EE prediction in light PA intensity category triggering misclassification of time intervals (min) spent in light into the moderate intensity PA category. This finding could be important in free-living studies using the two-regression model and carried out in populations spending significant amount of time in light PA intensity.
Using minute-to-minute simultaneous ActiGraph and EE measurements in the room calorimeter, we found that transient increases in EE predicted by the two-regression model do not always correspond to changes in measured EE. We also noticed that during activities characterized by constant intensity, such as running on the treadmill with a steady speed, the predicted minute-to-minute EE changed more than EE measured by the room calorimeter. This discrepancy could be considered a limitation of the model since a small change in the number of activity counts and/or their variability can lead to much larger changes in the predicted EE. To mitigate this effect, we applied the LPF to smooth the rapid minute-to-minute changes in the predicted EE. This approach led to a significant reduction in the mean absolute and mean squared error terms of the EE prediction reflecting physiological response to the movement. The LPF also significantly reduced the average total daily error from approximately 10% to 3% during the room calorimeter and from 6% to −2% during the DLW free-living phases of the study. However, the significant magnitude bias identified in both the filtered and unfiltered models during the DLW study indicate that both methods tend to over-predict EE in individuals with lower daily EE and under-predict EE in those with higher daily EE. The LPF model also improved the agreement between measured and predicted time spent in moderate and vigorous PA categories. However, the LPF technique reduced the ability of the model to accurately determine time spent in sedentary PA. This may be an important limitation of the LPF, since there is growing interest in characterizing the duration and intensity of sedentary PA behaviors (8
). Nevertheless, the LPF method may be useful because it can be applied to existing ActiGraph data sets to recalculate and perhaps correct predicted time spent in moderate and vigorous PA intensities and improve accuracy of the minute-to-minute predictions of EE. Future work should determine if the threshold used for activity count variability (CV) and/or the length of the LPF (5 minutes) used in this study should be modified to minimize the significant interindividual variability between measured and predicted EE in the free-living found in our study.
Our results are similar to previous ActiGraph DLW free-living validation studies that used linear regression equations to predict EE from the ActiGraph data (10
). The confidence intervals (CIs) calculated for the DLW in our study are also comparable to those reported previously. In a similar DLW validation study reported by Leenders, et al (11
) involving 13 healthy females wearing ActiGraph and Tritrac (Reining International, Inc., Madison, WI, USA) accelerometers for 7 days, the authors found that the EE predictions using the ActiGraph equations developed by Hendelman, et al (10
) and Swartz, et al (25
) based on lifestyle activities had the best agreement with EE measured by DLW, with differences (expressed as mean ± SD [range]) of −2 ± 16.6 % [−24 to 23%] and −4 ± 16.6 % [−29 to 24%], respectively. In the current study, we observed similar differences of −1.71 ± 15.3 % [−23.7 to 28.3%] and 5.97 ± 16.0% [−20.0 to 32.3%] for the filtered and unfiltered model, respectively. This may suggest that the two-regression model is not providing additional significant improvements in individual estimation of total EE when compared to equations reported earlier. Potential contributors to the variability between measured and predicted EE include individual variation in sleeping EE and sleep patterns, variability in accelerometer wear-time during waking hours, and variable amounts of activities which are difficult to measure using current accelerometer technology (e.g. biking, weight training, climbing stairs, etc.). These and other factors should be considered in designing and conducting studies in which energy expenditure and activity patterns are major outcome variables. Using similar validation approaches, future studies should also assess the accuracy and variability of more recently published approaches and equations (4
) and accelerometer placement. Similarly, many generalized prediction equation, such as two-regression model tested in this study, have not been rigorously validated for use in specialized populations with more unique PA patterns, including children, morbidly obese, older adults, and subjects experiencing significant weight changes. Furthermore, the limitations and variability of these generalized prediction equations should be taken into consideration in designing research protocols or clinical applications in which the amount and pattern of free-living PA or PA related EE are important outcome measures. For example, while we found the two-regression model with LPF might be sufficient to monitor average daily EE as a group, the original model would be better suited to identify sedentary activities.
Although using both the room calorimeter and free-living DLW assessment as calibration tools is a robust way to validate EE prediction equations from accelerometer data, we recognize that these methods have some limitations. First, the effect of diet-induced thermogenesis, which typically accounts for approximately 10% of the total daily EE, was not specifically measured or taken into account in either validation method. Second, while the standardization of activity bouts performed in the room calorimeter were advantageous for the consistency of the results, they may have directed the activity profile toward certain PA intensity categories which may not accurately represent the subjects’ normal daily PA patterns. Third, the relatively small size of the room calorimeter (20.4 m3) limited the equipment used and types of activities that were performed during the study. Finally, variations in the sleeping EE and sleep patterns during the free-living phase of the study were not accounted for and may have contributed to the variation in differences between measured and predicted EE observed in our study.
In conclusion, the EE predicted by a low-pass filtered version of the two-regression model proposed by Crouter et. al.
) showed good agreement with total EE measured in laboratory condition and in free-living using reference standards. The LPF approach offered improved total EE prediction and minute-to-minute accuracy compared to the original model. Despite the improvements in total EE prediction, the individual variability in assessing time spent in sedentary, low, and moderate PA intensities and related EE remained significant.