HR data were pre-processed using a Bayesian approach as described elsewhere [43 (link)]. Briefly, auxiliary variables (HR signal indicator, fastest and slowest heart beats) were used to assign data points to noise clusters, following which Gaussian Process Robust regression incorporating short- and long-term (circadian) covariance functions (priors) was used to infer the true latent HR time-series along with uncertainty estimates. ACC data were checked for anomalies (baseline bleed, axis freezing) but as none occurred in this dataset, these data were analysed in their raw form. Segments of data with continuous zero acceleration lasting >90 minutes were treated as ‘monitor not worn’ if also accompanied by non-physiological HR data (large and prolonged heart rate uncertainty).
PAEE (in kJ·day-1·kg-1) was calculated by time-integration of the activity intensity (in J·min-1·kg-1) time-series, estimated from HR and ACC separately and combined ACC+HR in a branched equation model [13 (link),18 (link)]. We accounted for any potential diurnal imbalance of wear time by weighting all hours of the day equally in the summation [44 ].
For single-signal HR estimates, the flex-HR method as described elsewhere [27 (link),28 (link)] was used. Briefly, this method translates HR to EE estimates according to an individually established calibration (as described above for each of the five levels) but only for time points where HR is above an individually determined flex point (flex HRaS); below this point activity intensity is estimated to be 0 J·min-1·kg-1). Specifically for this study, we used a flex HRaS of 10bpm +50% of lowest exercise HRaS, defined as lowest HRaS after 2-min of walking at 3.2 km·hr-1 for treadmill and walk-calibrated models and 80% of the 2-min HRaS value while stepping for step calibrated models; the latter is roughly equivalent to the HRaS value after 1-min of stepping but easier to determine reliably, and also comparable to the level of exertion used to define flex HR for the treadmill test. For non-exercise (group) calibrated models, predicted flex HRaS from sleeping HR was used [13 (link)].
The branched equation modelling technique assigns different weightings to the HR-PAEE and ACC-PAEE relationships, depending on the epoch-by-epoch observed values [13 (link),18 (link)]. Weightings for the two most extreme branches vary slightly between previous evaluations [14 (link),16 (link),18 (link),19 (link)]; as this matters most for the lower branch, we consolidated the two weightings, 0% and 10%, in the current study by applying the 0% weighting when average acceleration for the previous 2 min was lower than the movement (“X”) branching point, otherwise the 10% weighting was used.
Estimates of total energy expenditure (TEE, in MJ·day-1) were calculated from each model by multiplying PAEE by body weight, adding resting energy expenditure, and dividing this sum by 0.9 to account for diet-induced thermogenesis [45 (link)]. For the two models which use individual-level indirect calorimetry in the dynamic calibration, measured RMR from indirect calorimetry (ventilated hood) would be a likely method combination and so this was used to calculate daily REE; for all remaining models, predicted RMR [46 (link)] was used to calculate daily REE.
PAEE (in kJ·day-1·kg-1) was calculated by time-integration of the activity intensity (in J·min-1·kg-1) time-series, estimated from HR and ACC separately and combined ACC+HR in a branched equation model [13 (link),18 (link)]. We accounted for any potential diurnal imbalance of wear time by weighting all hours of the day equally in the summation [44 ].
For single-signal HR estimates, the flex-HR method as described elsewhere [27 (link),28 (link)] was used. Briefly, this method translates HR to EE estimates according to an individually established calibration (as described above for each of the five levels) but only for time points where HR is above an individually determined flex point (flex HRaS); below this point activity intensity is estimated to be 0 J·min-1·kg-1). Specifically for this study, we used a flex HRaS of 10bpm +50% of lowest exercise HRaS, defined as lowest HRaS after 2-min of walking at 3.2 km·hr-1 for treadmill and walk-calibrated models and 80% of the 2-min HRaS value while stepping for step calibrated models; the latter is roughly equivalent to the HRaS value after 1-min of stepping but easier to determine reliably, and also comparable to the level of exertion used to define flex HR for the treadmill test. For non-exercise (group) calibrated models, predicted flex HRaS from sleeping HR was used [13 (link)].
The branched equation modelling technique assigns different weightings to the HR-PAEE and ACC-PAEE relationships, depending on the epoch-by-epoch observed values [13 (link),18 (link)]. Weightings for the two most extreme branches vary slightly between previous evaluations [14 (link),16 (link),18 (link),19 (link)]; as this matters most for the lower branch, we consolidated the two weightings, 0% and 10%, in the current study by applying the 0% weighting when average acceleration for the previous 2 min was lower than the movement (“X”) branching point, otherwise the 10% weighting was used.
Estimates of total energy expenditure (TEE, in MJ·day-1) were calculated from each model by multiplying PAEE by body weight, adding resting energy expenditure, and dividing this sum by 0.9 to account for diet-induced thermogenesis [45 (link)]. For the two models which use individual-level indirect calorimetry in the dynamic calibration, measured RMR from indirect calorimetry (ventilated hood) would be a likely method combination and so this was used to calculate daily REE; for all remaining models, predicted RMR [46 (link)] was used to calculate daily REE.
Free full text: Click here
