Validity of the iPhone M7 Motion Coprocessor to Estimate Physical Activity during Structured and Free-Living Activities in Healthy Adults

23 Modern smartphones such as the iPhone contain an integrated accelerometer which can be used to 24 measure body movement and estimate the volume and intensity of physical activity. 25 Objectives: The primary objective was to assess the validity of the iPhone to measure step count and 26 energy expenditure during laboratory-based physical activities. A further objective was to compare 27 free-living estimates of physical activity between the iPhone and the Actigraph GT3X+ accelerometer. 28 Methods : Twenty healthy adults wore the iPhone 5S and GT3X+ in a waist-mounted pouch during 29 bouts of treadmill walking, jogging, and other physical activities in the laboratory. Step counts were 30 manually counted and energy expenditure was measured using indirect calorimetry. During two weeks 31 of free-living, participants (n=17) continuously wore a GT3X+ attached to their waist and were 32 provided with an iPhone 5S to use as they would their own phone. 33 Results: During treadmill walking, iPhone (703 ± 97 steps) and GT3X+ (675 ± 133 steps) provided 34 accurate measurements of step count compared to the criterion method (700 ± 98 steps). Compared to 35 indirect calorimetry (8 ± 3 kcal · min −1 ), the iPhone (5 ± 1 kcal · min −1 ) underestimated energy 36 expenditure with poor agreement. During free-living, the iPhone (7990 ± 4673 steps·day -1 ) recorded a 37 significantly lower (P < 0.05) daily step count compared to the GT3X+ (9085 ± 4647 steps·day -1 ). 38 Conclusions: The iPhone accurately estimated step count during controlled laboratory walking but 39 records a significantly lower volume of physical activity compared to the GT3X+ during free living. 40


Introduction
continuous remote access to the measurements via data sharing platforms. However, the validity of the 72 smartphone technology to measure parameters of PA needs to be further established. A recent study 73 compared the iPhone 5S to manually counted steps (Major & Alford, 2016) and found good correlation 74 between methods at fast walking speeds (4.68 and 6.48 kmh -1 ) but not at the slowest walking speed 75 of 3.6 kmh -1 . This study did not, however, explore the accuracy of iPhone mobile applications to 76 estimate EE. Of further interest is the impact of user behaviour with mobile devices in a free-living 77 environment and how this influences the accuracy of PA measurements. 78

79
To the authors' knowledge, no study has compared estimations of both step count and EE from the 80 iPhone 5S with laboratory-based gold standards during a variety of different physical activities. 81 Therefore, the primary purpose of this study was to assess the validity of the iPhone M7 motion 82 coprocessor, for estimating step count and subsequently estimating EE (with the use of the 83 'ActivityTracker' app) during treadmill walking, jogging, running, stationary cycling, and an aerobics 84 session. A further aim was to compare measurements of step count between the iPhone and GT3X+ 85 during a two-week period of free-living. 86 87

Study Design 89
The current study consisted of two distinct phases. The first phase comprised a single experimental 90 trial conducted in the laboratory to determine the validity of iPhone estimates of step count and EE in 91 comparison to gold standard measures (step count = manually counted, EE = indirect calorimetry. The 92 second phase was a two-week observational period during which free-living PA data was concurrently 93 monitored using the iPhone and the GT3X+ accelerometer. The study was approved by the School of 94 reserve); Moderate (40-59% heart rate reserve); and Vigorous (60-89% heart rate reserve) (ACSM, 122 2017). The second component consisted of cycling on a stationary ergometer at a fixed power of 50 123 W for 5 min. The last component required the participant to complete a 5 min aerobics session by 124 following a YouTube video. Throughout all activities, measurements of accelerometry, heart rate, 125 iPhone step count, and EE from the 'ActivityTracker' app were continuously recorded. The treadmill 126 component was also video recorded in order to manually count steps, video footage was reviewed and 127 manual steps were counted for each intensity with the use of a hand tally counter. Two members of 128 the research team separately watched and counted each video 3 times and recorded the values. These 129 values were then compared and when discrepancies were noted, the researchers reanalysed the videos 130 until agreement was reached. 131

132
In order to assess the validity of the iPhone during the laboratory trials, iPhone estimates of step count 133 and EE were compared to the gold standards (manually counted and indirect calorimetry, respectively). 134 For all activities, step count is reported as the number of measured steps for that component. Participants were instructed to step onto the motorised treadmill (PPS Med, Woodway, Waukesha, 138 WI) on which the incline was increased to 1% to mimic the metabolic cost of outdoor walking (Jones 139 & Doust, 1996). After finding the speed which elicited the desired heart rate range, participants were 140 asked to straddle the treadmill in order to record iPhone step count and EE from the 'ActivityTracker' 141 app before recommencing walking. This was repeated for light (5.4  1.0 kmh -1 ; 9  2 RPE), moderate 142 (6.5  0.9 kmh -1 ; 11  2 RPE) and vigorous (8.0  1.1 kmh -1 ; 13  2 RPE) intensities. Between each 143 stage, participants were again asked to straddle the treadmill and to stand motionless so the iPhone 144 step count and EE could be recorded from the 'ActivityTracker' app.

Components 2 and 3: Cycling and aerobics 147
Participants carried out 5 min of cycling on an electronically-braked ergometer (Lode Excalibur Sport; 148 Lode Medical Technology, Groningen, The Netherlands) with a constant external power output of 50 149 W. Participants were instructed to cycle at a comfortable cadence as they would on a leisurely cycle. 150 Following this, participants followed a 5 min segment of a YouTube aerobics-style cardiovascular 151 workout video (https://www.youtube.com/watch?v=istOU9nxhm8). The iPhone step count and EE 152 from the 'ActivityTracker' app were recorded at the beginning and end of each activity. 153 154 iPhone 5S 155 The iPhone application 'ActivityTracker' was downloaded onto the iPhone 5S from the Apple store. 156 This app was selected as it provided a live reading of daily total steps and estimates of EE. The 157 'ActivityTracker' app is reported by the developer to collect step count directly from the iPhones' 158 Health Kit and uses a bespoke algorithm based on step count, gender, stature and body mass to estimate 159

EE. 160 161
Accelerometers 162 Before each trial, a triaxial GT3X+ accelerometer (Actigraph, FL, USA) was initialised to record data 163 at a sampling frequency of 30 Hz in three axes: vertical, mediolateral and anteroposterior, using 164 ActiLife software (V6.13.3 Lite Edition, Actigraph, FL, USA). The Actigraph GT3X+ accelerometer 165 was selected for use in the current study as it has been previously shown to accurately assess step count  Reproducibility was evaluated using the concordance correlation coefficient of Lin (CCC) (Lin, 1989) 181 with the thresholds: almost perfect > 0.90; substantial > 0.8 -0.9; moderate 0.65 -0.8; poor < 0.65. 182 Bland and Altman (1986) analysis was used to express agreement between methods of measuring step 183 count and EE. The 95% limits of agreement (LOA) were calculated as mean bias  (1.96  standard 184 deviation). Log-transformation of EE data was attempted as the difference between measurement 185 methods increased as EE increased. However, this did not reduce the linear change of the data, so the 186 original, non-log scaled data were used. The mean percentage error (MPE) was computed as (steps 187 detected -observed steps (manually counted))/ observed steps (manually counted)  100, for step 188 count and (estimated EE -measured EE (indirect calorimetry))/ measured EE (indirect calorimetry)  189 100, for EE. The mean absolute percentage error (MAPE) was also computed using the same formulas, was used to assess differences between measurement methods (iPhone, GT3X+, and criterion 194 methods). Statistical significance was set at P < 0.05. All statistical procedures were carried out using Twenty adults volunteered to take part in the second phase of the study. Two participants withdrew 200 (reasons undisclosed), one participant did not have sufficient wear time of the GT3X+ accelerometer, 201 and the iPhone application malfunctioned for another participant. Therefore, sixteen participants, ten 202 female and six males (mean ± SD: 42 ± 17 years old), completed phase 2 of the study. 203 204

Experimental Design and Procedures 205
Participants were monitored for a total period of 14 days. Participants were given an iPhone 5S and 206 were asked to carry the iPhone with them as they would their own mobile phone.
Step count data from 207 the iPhone were automatically uploaded to a bespoke online digital platform (Lenus, StormID, 208 Edinburgh, UK) which enabled continuous data exchange between the user and the researcher. The 209 user experience of the Lenus health platform was evaluated as a separate component of this study and 210 will be reported elsewhere. Participants were also given a GT3X+ accelerometer which was attached 211 to an elastic waistband. Participants were instructed to wear the GT3X+ all day, every day on the right 212 anterior-superior iliac spine, removing only for sleep and showering/swimming. The accelerometers 213 (Actigraph, FL, USA) were initialised to record data at a sampling frequency of 30 Hz in three axes of 214 motion and data was downloaded as previously described. with the thresholds: almost perfect > 0.90; substantial > 0.8 -0.9; moderate 0.65 -0.8; poor < 0.65. 222 Bland and Altman analysis (Martin Bland & Altman, 1986) was used to assess agreement between 223 step count estimates from the iPhone and the GT3X+ as previously described. Paired T-tests were used 224 to determine whether there was a difference in step count between measurement methods (Jamovi Step count data from the iPhone, GT3X+, and criterion method during the treadmill trial are presented 231 in Table 1. The agreement in measurements of step count between iPhone and manually counted was 232 almost perfect (CCC = 0.993; 95% CI 0.988 to 0.996 steps) throughout the treadmill trial with a mean 233 difference of 3 steps (95% LOA -19 to 25 steps) ( Fig. 1.a) and a MAPE of 1.1%. When comparing the 234 intensities separately, there was almost perfect agreement between the iPhone and criterion methods 235 with a MAPE of < 2 % at each intensity (Table 2). 236

243
Estimates of EE from the iPhone, GT3X+, and criterion method (indirect calorimetry) during the 244 treadmill trial can be viewed in Table 3. The agreement in measurements of EE between iPhone and 245 indirect calorimetry was poor (CCC = 0.48; 95% CI 0.36 to 0.58 kcal·min -1 ) throughout the treadmill trial with a mean difference of -1.9 kcal·min -1 (95% LOA -5.6 to 1.8 kcal·min -1 ) (Fig. 3) and a MAPE 247 of 23.7%. When comparing the intensities separately there was moderate agreement at the light 248 intensity, while the iPhone estimates of EE were significantly lower than indirect calorimetry with 249 poor agreement at moderate and vigorous intensities (Table 4). 250

286
The primary purpose of this study was to assess the validity of the iPhone 5S for estimating step count 287 and EE during laboratory-based physical activities. We compared step counts from the iPhone 5S and 288 a research-grade accelerometer (GT3X+) to the criterion method. Both devices were found to provide 289 valid estimates of step count during walking and jogging on a treadmill. The GT3X+ was also found 290 to provide accurate estimates of EE during treadmill walking and jogging but the iPhone significantly 291 underestimated EE compared to indirect calorimetry. A further objective was to compare estimates of 292 step count between the iPhone 5S and GT3X+ during a two-week period of free-living. We found that 293 the iPhone recorded significantly fewer daily steps compared to the GT3X+, suggesting the iPhone 294 may not be a suitable method of estimating daily physical activity.
In the treadmill component of the laboratory trial, the iPhone provided near perfect estimates of step 297 count compared to manually counted, whereas the GT3X+ provided moderate estimates of step count 298 compared to the criterion method. Both the iPhone and GT3X+ were most accurate at the vigorous 299 (8.0  1.1 kmh -1 ) intensity and least accurate at the moderate intensity (6.5  0.9 kmh -1 ). This suggests 300 that the relationship between the accuracy of the GT3X+/iPhone and speed is not linear, as previously 301 reported (Lee et al., 2014;Major & Alford, 2016). However, the difference in accuracy of the iPhone 302 between intensities was very minimal, with MAPEs ranging from 0.6% to 1.5%, whereas the GT3X+ 303 ranged from 1.2% to 6.9%. Contrastingly, the iPhone was least accurate at estimating EE at the 304 vigorous intensity and performed best at light intensity (5.4  1.0 kmh -1 ), when compared to the 305 criterion method of indirect calorimetry. 306

307
The GT3X+ on average overestimated EE at all speeds and was least accurate at the light intensity, 308 while performing best at moderate intensity when compared to indirect calorimetry. Previous studies 309 comparing the GT3X+ to indirect calorimetry have also found the device to overestimate EE at speeds 310 comparable to those in the current study but to underestimate at faster running speeds, higher intensity 311 activities, and at much slower walking speeds (2.6 kmh -1 ) ( During stationary cycling, the iPhone and GT3X+ significantly underestimated EE and had poor 318 agreement with indirect calorimetry. The likely reason for the consistent underestimation of EE during 319 stationary cycling is due to the stable position of the trunk where both the iPhone and GT3X+ were 320 located. The adoption of a lower-limb accelerometer placement has previously been shown to improve the accuracy of pedal-revolution count during cycling when compared to waist-placement (Gatti,322 Stratford, Brenneman, & Maly, 2016). During aerobics, the iPhone significantly overestimated steps 323 compared to the GT3X+ and underestimated EE compared to indirect calorimetry, with poor 324 agreement for both comparisons. There was poor agreement between the GT3X+ estimations of EE 325 compared to indirect calorimetry, however methods were not significantly different. The poor 326 agreement between the iPhone and GT3X+ compared to indirect calorimetry during the aerobics trial 327 suggests that both methods are unsuitable for monitoring EE during exercise that is not steady-state. 328 The iPhone's overestimation of steps compared to the GT3X+ during aerobics suggests that it may not 329 be suitable for monitoring exercise that requires non-uniform movement patterns. 330

331
In the free-living component of the study, the agreement in daily step count between the iPhone and 332 GT3X+ devices was substantial although the iPhone, recorded significantly fewer steps (1095 333 steps·day -1 ). It is not possible to ascertain the precise reason for this discord although user behaviour 334 with the iPhone devices seems a likely explanation. While participants wore the GT3X+ attached to 335 their waist, they were instructed to carry the iPhone as they would their own personal phone to ensure 336 an ecologically valid measurement method. Depending on the individual, the iPhone may have been 337 regularly left on a surface or carried in a bag. Participants may have been less likely to carry the iPhone 338 on their person as it was additional to their own phone. Unfortunately, there was no way to monitor 339 "wear-time" of the iPhone so this hypothesis remains speculative. In conclusion, the iPhone 5S is a suitable method of measuring step count but not EE during walking 366 and jogging. In the free-living phase of the study, the iPhone significantly underestimated daily step 367 count compared to an accelerometer worn continuously around the waist. This is likely because the 368 phone was not carried on the person as frequently as the accelerometer. Further optimisation of the 369 prediction algorithms in the mobile apps to incorporate measurements of heart rate and/or GPS data data set on PA and sedentary behaviour patterns. Finally, when using smartphones such as the iPhone 372 5S to measure step-count, users should be cognisant that there may be a significant underestimation of 373

384
The authors have no conflicts of interest to report. 385 386 Authors 387 All co-authors made a substantial contribution to the concept of the work, or acquisition, analysis, or 388 interpretation of the data. All authors helped to draft and revised the article and approve the version to 389 be submitted. CE is the guarantor for the work.