Impact of Reduced Sampling Rate on Accelerometer-based Physical Activity Monitoring and Machine Learning Activity Classification

Purpose: Lowering the sampling rate of accelerometer devices can dramatically increase study monitoring periods through longer battery life, however the validity of its output is poorly documented. We therefore aimed to assess the effect of reduced sampling rate on measuring physical activity both overall and by specific behaviour types. Methods: Healthy adults wore two Axivity AX3 accelerometers on the dominant wrist and two on the hip for 24 hours. At each location one accelerometer recorded at 25 Hz and the other at 100 Hz. Overall acceleration magnitude, time in moderate-to-vigorous activity, and behavioural activities were calculated using standard methods. Correlation between acceleration magnitude and activity classifications at both sampling rates was calculated and linear regression was performed. Results: 54 participants wore both hip and wrist monitors, with 45 of the participants contributing >20 hours of wear time at the hip and 51 contributing >20 hours of wear time at the wrist. Strong correlation was observed between 25 Hz and 100 Hz sampling rates in overall activity measurement (r = 0.962 to 0.991), yet consistently lower overall acceleration was observed in data collected at 25 Hz (12.3% to 12.8%). Excellent agreement between sampling rates was observed in all machine learning classified activities (r = 0.850 to 0.952). Wrist-worn vector magnitude measured at 25 Hz (Acc25) can be compared to 100 Hz (Acc100) data using the transformation, Acc100 = 1.038*Acc25 + 3.310. Conclusions: 25 Hz and 100 Hz accelerometer data are highly correlated with predictable differences which can be accounted for in inter-study comparisons. Sampling rate should be consistently reported in physical activity studies, carefully considered in study design, and tailored to the outcome of interest.


Introduction 22
Accelerometers are increasingly used to obtain exposure measures in large observational physical 23 activity healthcare studies [1][2][3] and to obtain outcome measures in randomised controlled trials [4]. 24 Seven days of physical activity measurement is a common measurement window across a variety of 25 accelerometer studies and provides a suitable representation of individual activity in most scenarios 26 [5][6][7][8]. However, in the case of activity monitoring in clinical populations recovering from trauma or a 27 surgical intervention, extended monitoring well beyond a standard seven day protocol may be desired 28 [9]. Extended longitudinal protocols incorporating measurement windows beyond a few weeks 29 requires frequent participant interaction for battery charging, which is burdensome and undesirable 30 [10]. In these cases, a reduction of sampling rate would offer the potential for longer continuous 31 physical activity measurement, while reducing patient and caregiver interactions with the monitoring 32 device. 33 Currently, Brønd and Ardivsson offer the only side-by-side comparison of accelerometers recording at 34 different sampling rates, recording laboratory-based activities with hip-mounted Actigraph GTX3+ 35 devices [11]. Other studies generate downsampled data from a single accelerometer to assess the 36 effects of varying sampling rates, while relying on the untested assumption that this postprocessing 37 accurately reproduces the recording behaviour of the underlying hardware [12,13]. In addition, 38 studies of this nature have been limited to the ActiGraph GTX3+ and GENEA triaxial accelerometers 39 [11][12][13][14], and not the Axivity AX3 which has been used in the UK Biobank study of 100,000 participants 40 [1]. This dataset is a valuable resource and can serve as a reference for physical activity within a wide 41 range of clinical populations. Therefore, the ability to compare physical activity data between studies 42 integrating non-standard sampling rates remains unverified. 43 The aim of this study was to validate data collected at a reduced sampling rate (25 Hz) in comparison 44 to that collected at the commonly used rate of 100 Hz within the AX3 accelerometer. The objectives 45 were to: 1) identify any effect of sampling rate on vector magnitude both overall and for specific free-46 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 27, 2020. ;https://doi.org/10.1101https://doi.org/10. /2020 living activities; 2) characterise the effect of reduced sampling rate on machine learning activity 47 classification; 3) develop a transformation so that data collected at standard and reduced sampling 48 rates can be directly compared. 49

Study design 51
Ethical approval for participant recruitment was obtained from the Central University Research Ethics 52 Committee of the University of Oxford (Ref: R63137/RE001). Written informed consent was obtained 53 from adult volunteers (aged 18 and above) with no lower limb injury within the previous 6 months 54 and who were able to walk without an assistive device. Participants were recruited through advertising 55 within Oxford University, the local community and in senior citizen groups. 56 Participants were instructed to wear four triaxial accelerometers (AX3, Axivity, Newcastle, UK) for 57 24 hours, except during bathing or swimming activities. Participants could remove sensors at night if 58 they disrupted sleep. Two accelerometers were placed side-by-side on the dominant wrist using a 59 wristband, and two on the dominant-side hip, waist level at the anterior-posterior midline via a belt 60 clip ( Figure 1). 61 Accelerometers were synchronised and programmed via the Open Movement software (v1.0.0.42) 62 (https://openmovement.dev). At each body location, one sensor was programmed to collect data at 63 a sampling rate of 25 Hz while the other collected data at 100 Hz, both with a dynamic range of ±8 g. 64 The same four unique sensors were used in all participants. Accelerometer orientation and axis 65 alignment within the hip and wrist straps were consistent with manufacturer guidance and verified 66 prior to the start of each recording session. The assignment of hip or wrist body location and sampling 67 rate of each specific accelerometer was randomized by serial number so that the four accelerometers 68 contributed to measurements at both body locations and sampling rates. Participants additionally 69 filled out an activity diary (Supplement Note 1) to log the times during which they slept, cycled, walked 70 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 27, 2020. ;https://doi.org/10.1101https://doi.org/10. /2020 more than 100 meters, ate a meal, participated in self-defined exercise, or removed the 71 accelerometer. 72

Acceleration Magnitude 73
A schematic of our study design is presented in Figure 2 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint with concurrent ground truth images recorded every 20 seconds for a total of 160,000 minutes in 134 104 participants. 105

Statistical Analysis 106
A minimum sample size of 30 participants was calculated to detect a conservative correlation 107 coefficient estimate of 0.6 between sampling rate and body location measurements with 80% power 108 and alpha set at 0.05. This minimum sample size allowed for 30% data loss for dropout, protocol non-109 compliance, or poor data quality. Participants with less than 20 hours of wear were excluded from 110 analysis. Summary and descriptive statistics were calculated for participant demographics, overall 111 activity levels, and activity levels performed during self-reported activity designations. Two-sided 112 Spearman's rank correlation was calculated to compare activity levels, time in MVPA, and time in 113 classified activities between accelerometer sampling rates and body locations. Mean difference and 114 relative percentage difference in activity level between sampling rates was calculated with reference 115 to 100Hz measurements. Bland-Altman plots were used to assess fixed or proportional bias and limits 116 of agreement between sampling rates across the full range of participant activity levels. Mean 117 participant activity was plotted in 1-hour increments over a 24-hour day to visualise differences in 118 activity patterns recorded at either sampling rate. Linear regression was used to determine the 119 association between sampling rates and to estimate a conversion factor between acceleration vector 120 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 27, 2020. ; https://doi.org/10.1101/2020.10.22.20217927 doi: medRxiv preprint magnitude and activity classification recorded at the two sampling rates, with 25 Hz measurements 121 modelled as the predictor, 100 Hz measurements as the regression outcome. Leave-one-subject-out 122 cross validation (n-1) was repeated n times across each dataset to assess linear model generalisability, 123 with the corresponding root mean square error (RMSE) and R 2 reported. Statistical analysis was 124 performed in R (v.4.0.0) and RStudio (v1.2.5042). 125

Participant Demographics 127
Fifty-four healthy adults (33 female, 21 male) with a mean age of 43.4 years (SD 17.6; range 19.5 -81.2 128 years) participated in free-living physical activity analysis (Table 1). Acceleration measurement error 129 following self-calibration was less than 2.6 mg across all recordings, with no instances of calibration 130 failure or device malfunction. Median hip and wrist accelerometer wear time was 23.7 hours (IQR 0.6) 131 and 24.0 hours (IQR 0.5), respectively. Three sets of wrist-worn and nine sets of hip-worn sensors were 132 removed from analysis as they were worn for less than 20 hours. Participants self-reported an overall 133 mean of 7.4 hours of sleep (SD 1.0), 46.8 minutes of eating (SD 37.8), 28.4 minutes of general exercise 134 (SD 56.7), 58.2 minutes of extended walking (SD 64.0), and 34.1 minutes of cycling (SD 27.8) over the 135 24-hour assessment period as calculated in participants who logged non-zero time in those activities. 136

24-Hour Physical Activity Assessment 137
Mean physical activity across all valid measurements according to time of day is presented in Figure 3  138 and acceleration data for wrist and hip accelerometers are presented in Table 2. Strong correlation 139 was observed between 25 Hz and 100 Hz recorded data in overall vector magnitude and time spent in 140 MVPA at both the hip (r = 0.988) and the wrist (r = 0.960 to 0.971). At the hip, vector magnitude 141 collected at 25 Hz resulted in consistently lower overall activity compared to the 100 Hz data, with an 142 absolute mean difference in vector magnitude (95% CI) of 2.0 (1.6, 2.4) mg and 9.7 (7.9, 11.5) minutes 143 of MVPA per 24 hours. This represents a -12.4% relative difference in vector acceleration magnitude 144 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 27, 2020. ; https://doi.org/10.1101/2020.10.22.20217927 doi: medRxiv preprint and -11.2% difference in MVPA time when 25 Hz measurements when compared to 100 Hz 145 measurements at the hip. At the wrist, 25 Hz data collection also resulted in consistently lower overall 146 activity compared to the 100 Hz data, with an absolute mean difference (95% CI) of 4.4 (3.8, 5.0) mg 147 in vector magnitude and 25.9 (21.5, 30.3) minutes of MVPA, a -13.3% and -23.1% relative difference 148 in 25 Hz measurements at the wrist, respectively. Bland-Altman plots exhibiting the difference 149 between 25 Hz and 100 Hz 24-hour activity measurements are presented for hip and wrist 150 measurements in Figure 4. 151 At both measurement locations, effectively perfect correlation (r = 1.000) was observed for overall 152 activity between the recorded 100 Hz and the downsampled measurements from the same sensor 153 (Table S1). Comparison of sampling rate via the downsampling of 100 Hz data to 25 Hz in the same 154 sensor resulted in less than 0.1% difference in overall activity and MVPA across both the hip and wrist 155 (Table S1, Figure S1). 156

Intensity of Self-Reported Activities 157
Correlation between acceleration vector magnitude measured at 25 Hz and 100 Hz sampling rates was 158 high during all self-reported non-sleeping activities at the wrist (r = 0.956 to 0.997) and hip (r = 0.977 159 to 1.000). When recording at the hip, the 25 Hz sampling rate resulted in a mean relative difference in 160 vector magnitude of -12.9% to -16.7% when compared to the 100 Hz sampling rate across all non-161 sleeping activities. The greatest difference in 25 Hz and 100 Hz data at the wrist was observed during 162 cycling activity, where the 25 Hz accelerometer recorded relative difference of -38.6% when compared 163 to the 100 Hz recording. Other non-sleeping activities resulted in a -7.4% to -15.1% relative difference 164 when recorded at 25 Hz at the wrist. Acceleration during sleep was predictably low, with lower 165 correlation between sampling rates when compared to other activities (r = 0.476 to 0.899). The actual 166 differences are small due to the very activity measured during sleep intervals. 167 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 27, 2020. ; https://doi.org/10. 1101/2020 Machine Learning Activity Classification 168 Results of the machine learning activity classification are presented in Table 3. Periods of walking, 169 sleep, sit/stand activities, and mixed activities were identified in all participants, whereas periods of 170 cycling and vehicle activities were identified in 13 and 41 participants, respectively. In all participants 171 who had cycling or vehicle activities classified within the 100 Hz 24-hour recording, that activity was 172 also identified within the 25 Hz data. Reduced sampling rate did not produce consistent trends in 173 underreporting of activity across machine learning activity classification as was observed in the 174 quantification of vector magnitude. Activity classification between 25 Hz and 100 Hz measured data 175 was highly correlated (r = 0.855 to 0.967) in all activity categories. The reduced sampling rate resulted 176 in classification of 12.4% (95% CI: 7.5%, 17.4%) more of walking time, 12.8% (95% CI: 6.7%, 18.9%) less 177 time in mixed activities, and 19.2% (95% CI: 8.2%, 30.3%) less vehicle time per participant when 178 compared to 100 Hz collected data in this machine learning model. Cycling, sleep, and sit/stand 179 showed close agreement between sampling rates, with percent differences in time spent in these 180 activities ranging between 0.2% to 4.6%. 181 Converting data collected at lower sample rates to enable comparison with other studies 182 Results from the linear regression and cross-validation relating sampling rates based on overall 24-183 hour vector magnitude and activity classification models are presented in Table 4. At the hip and wrist, 184 the linear model of sampling rate accounted for between 97% and 99% of the adjusted variability in 185 the vector magnitude as described by R 2 . Cross-validation across 51 folds at the wrist and 45 folds at 186 the hip resulted in RSME values of 6.5% and 4.7% of the mean wrist and hip vector magnitude, 187 respectively. For each unit of increase in activity of 1 mg at the 25 Hz sampling rate, epoch-based 188 vector magnitude at the 100 Hz sample rate increased by 1.158 mg (95% CI: 1.120, 1.195) in 189 hip-mounted accelerometers and 1.038 mg (95% CI: 0.986, 1.090) in wrist-mounted accelerometers. 190 Acceleration magnitude collected at 25 Hz with a wrist-mounted AX3 accelerometer, Acc25, can be 191 adjusted for comparison to a more standard 100 Hz dataset. Acc100, (Box 1) using the relationship: 192 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 27, 2020. ; https://doi.org/10.1101/2020.10.22.20217927 doi: medRxiv preprint 100 = 1.038( 25 ) + 3.310. Transformations from 25 Hz to 100 Hz data were generated for 193 mixed, sit/stand, vehicle, and walking activities. In wrist-based activity classification, 69.5% to 92.1% 194 of the variability between sampling rates was accounted for in the models, with cross-validation RMSE 195 representing 7.5% of the mean time in the sit/stand classification and up to 50.5% the mean time between differing sampling rates. While both acceleration magnitude and activity classification are 208 highly correlated between sampling rates, subtle, yet real differences in outcome measures were 209 observed. Transformation of 25 Hz data can enable extended activity monitoring protocols while 210 retaining comparability to data collected at 100 Hz. 211 In the current study of healthy adults in the free-living environment, we found a repeatable difference 212 in measured acceleration magnitude and MVPA in a side-by-side comparison of data collection at 25 213 Hz and 100 Hz. When using a reduced sampling rate a 12% to 13% lower 24-hour activity and 11% to 214 23% lower time in MVPA was observed, compared to 100 Hz data collection at the hip and wrist. 215 Consistently lower acceleration magnitude was also observed in the 25 Hz recording across diary-216 logged free-living activities. Similar to the current study, Brønd and Ardivsson [11] found ActiGraph 217 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 27, 2020. ; https://doi.org/10. 1101/2020 GT3x counts demonstrated small differences between sampling rates (30 Hz, 40 Hz, and 100 Hz) during 218 walking, yet large differences between sampling rates during running, with up to 3000 counts per 219 minute greater intensity recorded in the faster sampling rates. Two other prior investigations of 220 accelerometer sampling rate relied on downsampling accelerometer data from a single sensor, with 221 mixed results and frequently little difference between sampling rates when using raw acceleration-222 based metrics [13,14]. In this study we conducted a secondary analysis by downsampling data initially 223 collected at 100 Hz, which also resulted in no difference in acceleration magnitude and MVPA at a 224 lower sampling rate, even while raw data collection at 25 Hz demonstrated a reduction in the activity 225 intensity captured during acceleration analysis. Subsequently, basic downsampling of data from a 226 single accelerometer is probably insufficient to determine the full effect of different sampling rates. 227 When using machine learning models of activity classification, we found high correlation between 228 different sampling rates across all classified activities. No repeatable pattern in under or over reporting 229 was observed in the reduced sampling rate, however, higher walking time and lower vehicle travel 230 time classified in the 25 Hz data. This finding is contrary to a study published by Zhang et al. [12] which 231 reported no significant reduction in activity classification accuracy when comparing downsampled 232 data to the original 80 Hz sequence. It is likely the use of downsampling versus a side-by-side 233 comparison has contributed to the opposing findings. Compared to overall vector magnitude and the 234 sit/stand classification, higher RMSE relative to mean values were observed in transformation model 235 for classified time in mixed activity, vehicle, and walking. This may be the result of fewer participants 236 and less time spent in these activities within the current dataset relative to low vector-magnitude 237 sit/stand activities. 238 Most recent physical activity studies include protocols where acceleration data is collected at a 239 sampling frequency between 20 and 100 Hz [10,18,19]. It is important to note that reporting of 240 sampling rate used in testing protocols remains a problem in the field of physical activity, with recent 241 reviews finding that 16% to 73% studies failed to report the sampling rate used in their data collection 242 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 27, 2020. ;https://doi.org/10.1101https://doi.org/10. /2020. In an assessment of benchmark datasets, Khan et al [20] have argued that typical 243 accelerometer sampling rates in many human motion studies are up to 57% higher than necessary for 244 adequate data analysis, with optimal sampling rates between 12 and 63Hz depending on body location 245 and model of accelerometer. Migueles et al [18], however, argue that future data processing needs 246 are unknown, and thus the highest possible sampling frequency should be used. The world's largest 247 objectively-measured physical activity database contains accelerometer data from 100 Hz data 248 collection using the AX3 device [1]. Based on the results of the current study, if a reduced sampling 249 rate of 25 Hz is selected in a measurement protocol, acceleration magnitude can be adjusted for 250 comparison to a more standard 100 Hz dataset. 251

Strengths and limitations 253
A major strength of our study is that it includes a side-by-side comparison of standard and reduced 254 accelerometer sampling rates at multiple body locations, without dependency on downsampling. This 255 methodology includes the comprehensive use of vector magnitude, cut-point, and machine learning 256 variables to facilitate direct comparison of data collected at 25 Hz with existing large physical activity 257 cohorts that have collected data at 100 Hz. However, data was collected in a healthy convenience 258 sample, and there may be unquantified effects of measurement in disease populations. While 259 individual devices were randomised to body location (hip/wrist), a potential limitation is that the exact 260 position of devices at each location was not randomised by design. For example, when sensors were 261 placed side-by-side on participants' wrists, it could be possible that the 100 Hz sensor was more 262 frequently placed nearer to the wrist and thus experience slightly higher accelerations. However, we 263 feel this alternative explanation is unlikely to account for the differences we and others have found 264 [11]. Epoch-based comparisons can have sensitivity to exact epoch boundaries, as all individual sensor 265 clocks drift slightly over time. Comparing two epochs between devices that report the same 266 boundaries may be comparing two slightly different periods. 267 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 27, 2020. ; https://doi.org/10. 1101/2020 Conclusion 268 Reducing accelerometer sampling rate from 100 Hz to 25 Hz can dramatically extend study monitoring 269 periods while still providing valid data, after appropriate transformations. In order to facilitate 270 comparisons between studies, researchers should consider and discuss the effect of sampling rate on 271 data analysis, and fully report all accelerometer parameters in the study methodology. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 27, 2020. ;https://doi.org/10.1101https://doi.org/10. /2020  is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 27, 2020. ; https://doi.org/10. 1101/2020