The main findings of this study are; 1) significant differences exist between the older piezoelectric-based monitors and newer micro electro-mechanical system (MEMS) based monitors, throughout a wide range of frequencies and different radii of oscillation in a mechanical setup and; 2) the direction of the differences observed in the mechanical setup were confirmed in a free living scenario resulting in differences in mean PA level and across PA intensities and 3) enabling the LFE did not increase similarities in the mechanical setup but attenuated the differences in mean PA between the AM7164 and GT3X in the free-living setup due to increased similarities in sedentary time and light PA.
As observed in the mechanical setup, the differences in free living were “non-systematic” throughout the intensity interval. This resulted in an eight percent lower mean PA of the GT3X compared to the AM7164 in free living. Enabling the LFE option attenuated the difference between generations in mean CPM due to an attenuation of differences in time spent in sedentary and light PA. Enabling the LFE option did not attenuate the differences observed in time spent on vigorous PA, but induced a larger difference in time spent on moderate PA.
We observed that the differences between monitors affected how persons were classified according the current PA guidelines. This could introduce biased estimates on population level when estimating the number of persons complying with guidelines. This post hoc analysis should be interpreted with caution as the sample was small and the calculation was based on a single day of measurements. There are currently no recommendations on non-specific sedentary time.
Differences between generations
Response curves of AM7164 and GT1M are broadly comparable to previous observations from studies using sinusoidal oscillations [6, 18]. Rothney et al. observed that the AM7164 yielded a higher output and a steeper slope (higher gain with increasing acceleration) compared to the GT1M monitor in the low frequency range (<1000 CPM) indicating a lower sensitivity of the GT1M [6]. Our observations support this notion for all newer MEMS-based Actigraph activity monitors at the shorter radius. At the longer radius this was only the case for the GT1M, indicating that either the MEMS accelerometer and/or the filtering approach differs between the newer generations. The difference between the GT1M and the GT3X was confirmed post hoc, by rerunning six GT3X randomly chosen from a new batch of monitors. We did not rerun the GT1Ms as our observations were comparable to previous observations.
Only a few studies have compared different generation of the Actigraph activity monitors in humans, showing varying results. John et al. did not find any differences between the AM7164 and the GT1M [7], whereas Fudge et al., observed higher output of the GT1M compared to the AM7164 [7, 9], using a similar sample and protocol. During self-paced locomotion at three self-selected speeds (mean (SD)); slow 0.7 (0.22) m·s-1, medium 1.3 (0.17) m·s-1 and fast 2.1 (0.61) m·s-1. Kozey et al. found a 2.7% higher output of the GT1M, compared to the AM7164 across all speeds (p < 0.05). Analyzing the result by speed, the GT1M only provided significantly higher outputs at the medium speed (5.3%, p < 0.05) [8]. The two latter observations are in accordance to our observations in the moderate and higher intensities. Only one study has so far compared the GT1M to the GT3X; 50 healthy participants performed a treadmill test at four medium to high intensity speeds (4.8, 6.4, 9.7 and 12 km·h-1) [11]. As in the mechanical setup, they did not observe any differences between the outputs on the vertical axis between the two generations in the mid-range intensities.
To our knowledge only one study has compared the AM7164 with the GT1M monitors during free-living conditions. Corder et al. found that the GT1M yielded a 9% lower output (p < 0.05) compared to the AM7164, (LOA; -36% to 23%) and the difference between the monitors increased with activity level in Indian adolescents. This did not translate into observable differences in time spent at moderate and high intensities [10]. The difference in mean CPM was similar compared to the difference in our free living study. In contrast we also observed significant differences in time spent in moderate and vigorous PA. In the mechanical setup the differences between generations seemed to depend on the radii and thus the magnitude of displacement in free living. As the size and the direction of the inter-generation bias is intensity dependent, the absolute differences during free-living conditions would depend on the time spent at certain intensities and the magnitude of displacement during locomotion. The differences between monitor generations may therefore be population-specific. Therefore, the difference could be ascribed the population (adult compared to child sample) or differences in activity types.
More free-living studies on different subpopulations are needed to interpret the impact of the differences observed in the mechanical setup in free-living. These studies would be most informative if waveform acceleration data were collected alongside epoch-level data with analogue Actigraphs; raw data may then be exposed to the same mathematical operations currently imbedded in firmware to produce epoch-level count data.
Low frequency extension option
Enabling the LFE option did alter the response curve of the GT3X and GT3X + at all frequencies at both radii. This increased the count output in the low frequency range and did in fact decrease the difference between the AM7164 and the GT3X + in the low frequency range at the short radius as described by Actigraph. This decreased the difference between the monitor outputs at the lower frequencies. However, the upward shift of the curve increased the difference between the AM7164 and the GT3X + at the higher frequencies. In free living enabling the LFE option attenuated the differences of the mean PA level and time spend sedentary in our free living study. Therefore, it could be advisable to use this option in future studies. As the direction of the bias introduced by the LFE option observed in the mechanical setup was confirmed in free living, thus increased the differences at the moderate intensities, using the LFE option to the decrease the inter-generation differences, should be done with caution. This observation needs to be confirmed in other samples.
In summery we observed differences between the old and newer generations of Actigraph monitors throughout a wide range of frequencies at different radii and that the absolute differences depend on both frequency and amplitude. The pattern observed in the mechanical setup was confirmed in free living. When enabling the LFE option, and thereby extending the sensitivity in the low-frequency range, the output of GT3X + increased to a level, comparable to the AM7164. This attenuated the differences in the sedentary and lighter intensities in free-living.
This indicates that the lower output observed at the lower frequencies might not be due to lower sensitivity of the hardware in the newer generations of monitors per se but due to non-optimal matching of filtering and processing algorithms imbedded in GT firmware for mimicking the analogue AM7164 frequency and magnitude response. In addition, the old and newer generations differ in two important parameters that could influence the monitor output at a given frequency and amplitude; resolution of the A/D signal converters (8 bits vs. 12 bits) and sampling rate (10 Hz vs. 30 Hz).
Strengths and limitations
The strength of the study is the combined use of mechanical setup with free living confirmation of the comparison. Furthermore, we used different batches of the same Actigraph generation and the newer activity monitors were randomly picked among more than 400 monitors compiled by three batches (bought at three different time points in the period 2009–2010). We also used multiple amplitudes of displacement within the range of displacement observed in human movement [27]. Taken together this increases the generalizability of our findings to other batches of Actigraphs and the performance within the displacement amplitude in humans.
The study also had some limitations. The activity monitors were acquired from different study centers and have been used by several different field investigators. Therefore we do not have complete knowledge of service history and frequency of use. To rule out batch differences we reran the protocol using six GT3X monitors randomly picked from an independent batch and six of the AM7164 used in the study. The results did not differ from the main results (data not shown). Previous observations have shown that the AM7164 output increases over time when used repeatedly in the field [16]. Thus, the difference between the newer and older generations could be a product of changes in the mechanical properties of the older accelerometer. This would probably not influence the shape of the response curves, as the observed drifting seems to be an offset drift (parallel shift of the curve) as Moeller et al. observed a constant increase at all settings in their setup (+2.5% over three months) [16]. This would consequently increase the bias observed in the low frequency range and decrease the bias in the medium frequency range. We only used the GT1M version 4 (bought in 2008), which might limit the generalizability of our finding to previous studies, as the signal processing is different in the newer generation (8-bit A/D signal converters in the older version vs. 12-bit A/D signal converters in the newer versions). However, our results are comparable to the findings by Rothney et al. [6], and smaller differences could be explained by differences in the mechanical setup.
Further, the free living study was based on a small sample. However, the study had time represented at all intensities of PA as would a background sample. Furthermore, we observed a significant difference between generations despite a small sample. Increasing the sample size would probably decrease the variance in the sample but not the mean bias. Post hoc inspections of Bland Altman plots revealed that the AM7164 output was higher in sixteen of twenty subjects whereas the bias was opposite (by ~10-20 CPM) for the remaining four subjects. There was no consistency in the position order of the activity monitors in the belt nor did we observe a different trend in the output of these monitors in the Actigraph calibrator compared to the remaining AM7164. It has been suggested that batch differences exist within the same brand of monitors [17]. The four monitors with the non-uniform direction of the bias were from the two different batches used in this study. Therefore we do not suspect this to explain the observation. As a mean CPM can be compiled through different intensity distributions, we investigated whether the four participants differed from the remaining sample herein. Interestingly, they spent ~40 min (p > 0.1) more time in the sedentary interval when assessed by the GT3X (LFE disabled), ~60 min more (p > 0.1) when assessed by the GT3X (LFE enabled) and ~80 min more (p < 0.05) when assessed by the AM7164 compared to the remaining sample. No substantial differences were observed in minutes spend at moderate and high intensities. This indicates that the four AM7164 monitors used for the participants with the non-uniform direction of the bias are less sensitive in the sedentary interval compared to the remaining sample. We therefore visually compared the trajectories of the AM7164 four monitors obtained in the mechanical setup to the remaining sample. When comparing the AM7164 on the short radius, the output from the four monitors was slightly lower than the remaining sample. This was not observed using the long radius, thus confirming that these four AM7164 monitors might in fact be less sensitive at the low intensities. No differences were observed when comparing the GT generations (data not shown). Larger inter-monitor variation has been observed between the AM7164 validated on the Actigraph calibrator compared to the newer GT-generations [6]. The non-uniform direction of the bias could thus be explained by the inter-monitor variation due to a lower sensitivity of the four monitors during sedentary activities compared the remaining batch. This problem would not be captured by the standard use of the Actigraph calibrator as the recommended gain limits are based on the peak value at 0.75 Hz. Even though we manually inspected the plots during calibration, we did not see a marked different output at the lower intensities when calibrating the four AM7164 monitors with specific monitors.
Finally, we scaled the one-minute cut-points to fit the ten-second epochs. This could compromise the precision of the cut-points related to the energy consumption. However, increasing the epoch time could blunt the response to especially higher PA intensities [28]. It would therefore an advantage to employ short epochs when evaluating the differences between generations at higher intensities.