The UKB is a prospective cohort study of 502,507 recruited between 2007 and 2010 [18]. Participants were asked to participate in follow-up studies (at random). Between 2013 and 2015, a total of 103,661 participants consented to wear a wrist accelerometer for 7 consecutive days [19]. The study participants all took part in subsequent visits between 2013 and 2015, following their participation in the accelerometer measurement.

Of the 502,507 participants included in the UKB, 103,661 had accelerometer records available. Of those, 16,467 had valid accelerometric data and had anthropometric and biomarker data available following the accelerometer measurement visit (see Figure 1 for the flowchart of the study). The study cohort was comparable to the whole UKB population, in that it was primarily white (97% vs 95%), had similar measured anthropometrics such as BMI (26.29 [4.39] kgm2 vs 27.43 [4.8] kgm2) and similar proportions of males and females (53% females vs 54%). However, the study population was more educated (56% in the highest education category vs 33%), had higher income (31.5% in the top 2 income categories vs 25.4%), and healthier lifestyle habits in that the majority had never smoked (63% vs 55%). The studied cohort was also on average 4 years older than the whole UKB population at baseline, even after accounting for the time elapsed between initial recruitment and the time of measured PA. Refer to Table 1 for all evaluated characteristics.

Measured or objective PA was assessed at both the individual level (Figure 2A). We observed concordance between 3 random participants’ hourly classified activities and the averaged triaxial acceleration, and at the population level in Figure 2B. At the population level, and even at the individuals’ level, we found that a little proportion of time was spent in MVPA. Throughout the day, individuals spent at most 12.5% of an hour engaged in MVPA, whereas 50% of their time was spent in sedentary status outside of their sleep. The weekly median number of minutes spent in sedentary time was 3200 minutes (7.62 daily hours), 2000 minutes (4.76 daily hours) in walking, and 500 minutes (1.19 daily hours) in MVPA (Figure 2C). Participants averaged 24-hour period habits during the weekday did not differ from habits during the weekend.

The self-reported levels of sedentary time were not concordant with the objectively measured sedentary time. Participants reported 2000 weekly minutes (4.76 daily hours) of sedentary time, or about 1200 weekly minutes (2.85 daily hours) less than the measured level. The population mean of self-reported walking was similar to the self-reported MVPA population mean, each averaging 250 weekly minutes (0.6 daily hours). However, these self-reported values deviated from their measured counterparts by 1750 minutes for walking (4.17 daily hours) and 250 minutes for MVPA (0.60 daily hours; Figure 2C).

Objective PA was modestly associated with subjective PA in Figure 3A-D (r=0.12; P<.001 for sedentary; r=0.18; P<.001 for walking; r=0.16; P<.001 for MVPA; and r=0.13; P<.001 for sleeping). Interestingly, the correlation was higher for walking and MVPA activities than for sedentary time and sleeping. We found the same correlations for the subpopulation that had noninterrupted accelerometer wear time during the study period (figures not included).

Figure 4A depicts the standardized effect size of measured and self-reported MVPA and sedentary time on selected cardiometabolic-related biomarker or “response” levels (see Figure S1 in Multimedia Appendix 1). The subjective and objective PA are modeled jointly, such that each measurement’s effect estimate assumes the other is held constant; in other words, the coefficients measure the effect of one PA measure independent of the other. The objective PA association sizes were greatest for anthropometric markers (BMI, waist circumference, body fat mass), certain physiological markers such as pulse rate, and certain plasma biomarkers such as high-density lipoprotein and triglycerides. As a negative control, objective sedentary time did not predict other biomarkers presumably unrelated to physical fitness, such as sunscreen usage (β=−8.42e-6; 95% CI −2.89e to 5,1.20e0; P=.42); however, objective MVPA was nominally significant (β=−1.13e−4; 95% CI −1.79e−4 to 4.72e−5; P=.009). Also, BMI and pulse rate levels had an inverse association with MVPA (eg, lower BMI and pulse rate for higher levels of MVPA) and a positive association with sedentary time. Further, subjective MVPA estimates were null across all biomarkers, but subjective sedentary time estimates were similar to the estimates predicted by objective sedentary time.

The association sizes of objective MVPA were directionally consistent and stronger in magnitude than those of subjective MVPA in all groups (above/below 65, males/females; Figures 4B and C, Figure S2 in Multimedia Appendix 1). There was also no substantial difference in effect size between the age groups.

In contrast to MVPA, objective and subjective effect sizes for sedentary time coincided almost perfectly for all anthropometric and physiological measurements (see Figure S3 in Multimedia Appendix 1 for all tested biomarkers). Sedentary time was associated with an increase in BMI and pulse rate (Figures 4D and E). We also replicated these tests in the subpopulation that wore the accelerometer without interruption and found the same results (figures not included).

Finally, we calculated the differences of biomarker effect sizes between each subjective/objective PA pair stratified by sex and age (above/below 65 years old) in original and nonstandardized units (eg, beats per minute for pulse rate and kgm2 for BMI). For BMI estimations from weekly MVPA, we found that subjective MVPA was associated with an increase of 1.5 kgm2 more than objective MVPA (1.59kgm2 [1.41, 1.78] for females <65, 1.47kgm2 [1.27, 1.67] for males <65, 1.5 kgm2[1.25, 1.75] for females >=65, and 1.53kgm2 [1.31,1.74] for males >=65). We also found that subjective MVPA was associated with an increase of approximately 2 bpm more than objective MVPA (1.56 bpm [1.09, 2.04] for females <65, 2.34 bpm [1.71, 2.98] for males <65, 1.95 bpm [1.25, 2.66] for females >=65, and 3.03 bpm [2.25, 3.8] for males >=65). For BMI predicted by sedentary time, we found that subjective estimates were lower than objective measurements by, on average, 1kgm2 for all subsets. See Figure S4 in Multimedia Appendix 1 for all tested biomarkers.

Here, we quantified the relationship between objective and subjective PA in two ways: (1) in a direct, one-to-one comparison and (2) in their association with cardiometabolic biomarkers. The former yielded similar results to previously reported studies assessing the differences between measured PA assessed with wearable devices and questionnaires [7,10,20,21].

First, the reported low Pearson correlations between objective and subjective PA measures were unsurprising given other reported levels below r=0.30 [7,21,22]. A European cohort reported a Pearson correlation of r=0.28 [21] while a Norwegian study reported correlations of r=0.46 for sedentary time and r=0.20 for MVPA [20]. Our results were lower. Further, participants underestimated their sedentary durations by almost 1200 weekly minutes (2.86 hours per day), a larger difference than what Prince et al [10] reported in their systematic meta-study review, where an underestimation of 731 weekly minutes (1.74 hours per day) was reported.

We found great mismatches when matching each individual’s subjective/objective measurements. In particular, we found that the mismatch was greatest in those who self-reported higher values of MVPA. This may have indicated that either participants with high activity levels over-reported their MVPA workout activity, either because (1) the accelerometers were less adapted to validate highly-active participants’ levels, or (2) that other factors like education level, age, or physical ability [20], affected participants’ self-reporting values. A reasonable explanation is probably a combination of these hypotheses. For example, at the population level, measured activity medians of all types were higher than the questionnaire-based ones. In fine, discrepancies in perceived PA durations are sure to vary between individuals given their different mobility abilities, which were not assessed in the cohort.

The way that we assess PA has a large influence on the size and robustness of the association with clinical biomarkers of disease. Our findings are in agreement with other studies that describe the inverse association between increased MVPA and BMI and heart rate (respectively, the regular association between sedentary time and BMI and heart rate) [3,14]. However, we observed that objective measurements were more strongly associated with cardiometabolic biomarkers than has been previously appreciated. In the case of increased MVPA, the negative association of measured PA was of a larger magnitude than that of subjective PA by 4e-3 standard units on anthropometric markers and on the order of 1e-3 standard units in the case of plasma biomarkers and pulse rate. In the case of increased sedentary PA, the magnitudes of objective and subjective associations were largely invariant to age and sex/gender.

Moreover, we found discrepancies between levels of BMI and pulse rate estimated with subjective and objective MVPA. Subjective weekly MVPA was estimated on average, in the different population subsets, 1.5 kgm2 and 2 bpm more than objective MVPA. Contextualized in healthy adults, these numbers reflect a 6% overestimation of MVPA association on BMI, if relying on subjectively-reported observations rather than measured MVPA. The MVPA impact on pulse rate differences was less, varying between 2% and 3% difference on an adult heart beating 60-100 bpm. Though small, these variations may present key influences on clinical biomarkers.

Due in part to the ease of implementation of IPAQ (long or short form) or equivalent questionnaires such as 7-day Physical Activity Recall [23], and in part by the relatively new widespread use of wearable devices, most cohort studies still rely on only subjective PA measurements to assess physical fitness. For instance, a recent study using the UKB found that participants adhering to frequent MVPA patterns based on subjective PA were associated with reduced risk of dementia [24]. Given the discrepancies between subjective and objective PA quantified in this study, however, it is questionable whether the reported effect is discernible in objectively measured PA in the UKB cohort.

PA investigation on cardiometabolic disease may benefit from increased power and precision from the use of accelerometer data. These findings support the hypothesis that the difference observed between subjective and objective MVPA measurements was more likely due to misjudging MVPA duration according to fitness levels and not age. Indeed, we did not observe any objective MVPA activity difference between the 2 age groups but found diluted pulse rate associations with increased MVPA activity in both groups (similarly, increased pulse rate with increased sedentary time). Given that the objectively reported effect sizes had greater magnitudes (and perhaps less error) than the subjective ones, greater power may be found with accelerometer-based measurements.

The measured PA data values were extracted and categorized from the freely available software [12,13,19]. Although measured PA was shown to strongly correlate with daily observed and reported behavior, the PA categorizations are only approximations of MET, which is defined as using 3.5 mL of oxygen per minute per kilogram of body weight. By definition, a participant is classified as doing MVPA if the activity duration intensity is greater than or equal to 3 METs. While gold-standard methods such as doubly labeled water [25] exist and result in precise, individual-based calibrations, the method used to process accelerometric data did not benefit from an individual-centric calibration. This, therefore, is one source of the discrepancies between PA classifications.

We hypothesize that individuals with different baseline cardiometabolic health may perceive their PA differently, and this will bias associations between reported activity and cardiometabolic biomarkers. We describe a plausible scenario: for instance, if 2 individuals of different baseline metabolic capacities are carrying out the same activity intensity, such as running a certain distance at a certain pace for a certain time, an individual-centric method would result in different MVPA duration classifications for each individual. Equivalently, an individual with a healthier baseline metabolic capacity would have less MVPA duration than an individual with a lower metabolic capacity. Hence, in a scenario where both individuals run the same distance at the same pace, the accelerometer reading will be identical. However, the self-reported perception of PA may differ; the individual with better cardiometabolic fitness may have perceived and reported doing less MVPA than the other individual. This could translate to the reported discrepancy observed in the study participants and therefore influence the magnitude of the coefficients of the biomarkers studied.

Unlike MVPA, we observed concordant associations between sedentary activity (measured by both accelerometer and self-reported) on biomarker levels. We hypothesize that measurement “error” in the aggregate may be lower for activities related to inactivity, which sum over behaviors such as watching television and using a computer versus reporting one MVPA-related activity carried out weekly [9,20].

With the new generation of mHealth devices that have been developed since the study’s conception, which incorporate personalized thresholds to categorize activity intensities [5]. There is evidence that activity intensity is assessed more objectively and therefore reflects more precise metabolic exertion. Leveraging these threshold personalization allows us to distinguish new behavioral patterns. For instance, McConnell et al [5] identified a cluster of adults who were more active during the weekend (called weekend warriors) in comparison to weekdays. Implementing studies with such devices is, however, expensive, but given the discrepancies highlighted in this study between self-reported and measured PA and predicting cardiometabolic biomarker levels, it may be crucial to re-evaluate PA assessments, especially when these are translated to prevent disease outcomes, such as in guidelines [26]. While we acknowledge the development of new technologies for measuring PA, the core contribution of our paper lies in the method we present to quantify the quality of relationships produced by mHealth technologies. This approach is not limited by specific devices or data sources and, importantly, should generalize to newer technologies as they emerge.

It may be possible, in the future, to create a calibration between self-reported and objective PA to reanalyze results from studies that have one measure and not the other. Many studies rely on self-reported measures. In this study, we are primarily focused on how the discrepancies between self-reported and objective measures impact associations with clinical outcomes, rather than directly translating between the two measures. As we show in Figure 3, the correlation between self-reported and objectively measured PA is very modest, below 0.1, highlighting the challenges inherent in creating such a mapping. This low correlation, in itself, demonstrates that providing a simple or reliable translation will require a formal study and potentially more data.

Use of self-reported versus accelerometer-based approaches for measuring PA has implications in biobank-scale genetic studies, such as “genome-wide association studies” (GWASs) [27]. GWASs are a discovery-based approach to broadly uncover inherited genetic loci, or places along the genome that predispose individuals to perform PA. However, we hypothesize that the scale of genetic discovery (the number of loci found) and their association sizes are influenced by the type of PA being interrogated. A number of papers have begun to interrogate the genetic contributions, via GWAS, in PA using both self-reported [28] and/or objective/accelerometer-based modalities.

For example, Klimentidis et al [29] conducted a GWAS of habitual PA of both self-reported PA and accelerometry data in the same participants analyzed in our study, individuals from the UKB. Briefly, they associated genetic loci with self-reported MVPA and VPA. They also associated genetic loci with average acceleration and fraction of time with higher accelerations as proxies of vigorous activity. In brief, while the sample sizes from each of the GWASs were different, they found different numbers and genetic loci from GWASs of self-reported and accelerometer-based measurement of PA. Specifically, the self-reported and objective indicators had an overlap in some loci found (particularly near gene Cell Adhesion Molecule 2), but the concordance was not consistent across all assessments of self-reported and accelerometer-based modalities. At a relaxed genome-wide P value threshold, the authors observed 4 genetic loci in common between self-reported MVPA and accelerometer and 3 loci in common between vigorous activity and accelerometer GWASs. Studying the difference in the genetic “architecture,” or the number and biological implications of genetic loci, is an aim for a future study.

Some caution should be used when interpreting these results. First, objective and subjective measurements were not collected at the same time. Participants of the UKB were invited to fill out IPAQ forms during their medical appointments independently of their participation in the accelerometer study. Furthermore, participants in the accelerometer study were not required to schedule a medical appointment during the week of their participation. Thus, self-reported PA was not collected simultaneously with the measured PA. While most participants had self-reported PA measures within weeks of their measured PA recording, some had at most 2 years between the measures. Though we applied a weight correction accounting for time elapsed between the measures (self-reported PA and measured PA) and considered additional lifestyle factors, there were other factors that could not be captured, which may have influenced the observed differences between objective and subjective MVPA.

In sum, by associating both objective and subjective PA with health indicators and clinical risk variables, we quantified the epidemiological implications of the relationship between subjective and objective PA versus just studying the concordance between objective and subjective PA. We found that BMI and pulse rate followed a positive relation with MVPA and a negative association in the case of increased MVPA. Further, we showed that objective MVPA measurements, even in the presence of subjective MVPA, captured more robust signals than subjective MVPA. Importantly, the framework we presented here allows us to quantify the quality of relationships produced by mHealth versus self-reported approaches. We showed that our comparison should be carried out to evaluate the degree of bias of self-reported versus mHealth devices in large cohort studies. This has implications beyond the accelerometer devices used in this study and is generalizable to other mHealth devices. Thematically, these findings demonstrate addressing key themes in mHealth, including mHealth for Data Collection and Research, Evaluation and Research Methodology for mHealth, and Fitness Trackers.