Individual Physical Activity
MEASURES REGISTRY USER GUIDE
Supplemental Considerations for Monitor-based Assessments (New Technology and Data Processing Techniques)
The rapid development and evaluation of new monitoring techniques and technologies makes it challenging for researchers and practitioners to determine the best way to collect, process, and interpret data with these methods. Recent advances in data processing methodologies available for activity monitors also may necessitate advanced data management and analytic skills, which can further complicate assessment decisions. This supplemental section introduces some of the complexities and refers to other sources for additional information.
Background on Accelerometry-based Activity Monitors
Considerations for Collecting and Processing Physical Activity Data
Accelerometry-based activity monitors are relatively easy to use, but a number of decisions must be made when collecting, processing, and interpreting data. One of the first decisions is the epoch or sampling timeframe for the assessments. Most monitors allow for the selection of different time interval windows (e.g., 1-second epochs, 5-second epochs, 30-second epochs) and the choice can have implications for the outcomes. Researchers routinely use a 60-second epoch because a minute is a reasonable unit for comparison and evaluation. However, the intermittent nature of children’s activity patterns requires that lower epochs be used to capture more sporadic activity patterns (i.e., 5 seconds).
A number of other decisions must then be made when processing activity monitor data. Some of the more critical decisions include selecting thresholds or equations to interpret or scale the data, selecting a method to determine non-wear time (e.g., 60 minutes vs. 20 minutes of consecutive zeros), defining minimum wear time to consider a day to be representative (e.g., 60% vs. 80% of total day), identifying spurious data (e.g., ≥20,000 counts vs. ≥16,000 counts in Actigraph data), and selecting the number of valid days needed to characterize “habitual physical activity” (e.g., 4 days vs. 7 days; week vs. weekend). The availability of product-specific software can facilitate this process and allow the user to create customized settings for the data. For example, Actigraph data can be processed using Actilife data analysis software platform. The software can easily allow the user to set specific non-wear time algorithms, activity cutpoints, and a minimum wear requirement, among others. The software also allows data from the monitors to be converted to a variety of formats (e.g., csv, dat) and analyzed using various statistical packages (e.g., SPSS, SAS, R).
The different decisions regarding the data generated from activity monitors require that researchers clearly identify the data reduction procedures used when using activity monitor data. Depending on the decisions about data reduction protocols, researchers and practitioners can expect significant differences in wear time, activity counts per minute, average activity per day (in counts/day), average MVPA levels (in minutes/day), and average MVPA bouts per day. The number of participants meeting physical activity guidelines also will differ depending on the different data reduction protocols. The more conservative the data reduction protocol is, the lower the number of participants with valid data, the lower the number of minutes of inactivity, and the higher the number of minutes spent in light and MVPA.74 Many papers in the literature have alluded to this problem, and there have been calls for standards to help facilitate comparisons among studies.74-76 Some examples of excellent detailed procedures used for data reduction protocols include the National Health and Nutritional Examination Survey (NHANES),12 the International Children’s Accelerometry Database (ICAD),77 the Canadian Health Measures Survey,78 and the ENERGY-project.
Considerations of Hip versus Wrist Placement
The majority of work on accelerometry-based monitors has been conducted using devices worn at the waist or hip. However, investigators have begun to transition toward the use of wrist-worn monitors. This transition has been fueled by the progression in consumer-based monitors as well as by evidence that compliance is enhanced when participants are asked to wear monitors on the wrist (more like a watch). The wrist placement may offer some advantages, but it is important to note that equations and methods developed for hip-worn devices cannot be directly applied to data collected with wrist-worn monitors. Acceleration at the wrist is generally higher than that at the hip and therefore requires new calibration studies to determine physical activity intensity cutpoints so that wrist data can be interpretable. For example, only a few wrist cutpoints have been proposed for youth, and the evidence supporting them is still limited. The uncertainties about the utility of existing cutpoints for the wrist limit the ability to make direct comparisons with previous research when monitors were worn at the hip. It is important to carefully consider the relative advantages and disadvantages of monitor position when planning a study. Even so, several large epidemiological studies including, NHANES,t have elected to use the wrist position; this will likely drive additional development and innovation. The reports in the Measures Registry may or may not specify the location used in the various validation studies because this distinction is a relatively new consideration with the use of monitor-based approaches.
Handling Missing Physical Activity Data
One of the challenges associated with physical activity measurement protocols using direct measures is the burden placed on the children being assessed. Complete, full-week measures of activity using accelerometers would ideally require that youth use an activity monitor device for 24 hours a day, during 7 consecutive days. Instead, youth often forget to replace the monitor after showering or sleeping, or choose not to wear the accelerometer during some periods of the day. It is important to account for these periods when the activity monitors are not worn.80-81 One strategy is to remove these bouts of non-wear time from further analysis. However, after removing these periods, it is important to check whether the remaining recorded data can still provide a representative picture of activity levels during the week being assessed or whether the children who were ultimately excluded from the data differ from those who were compliant (e.g., less active). It is not surprising to find non-compliance rates of 30% of the total sample, meaning that 1 out of every 3 children does not comply with the physical activity measurement protocol.82-84 Previous research has shown that youth characteristics, such as BMI, age, and screen time, can predict non-compliance.82-83 However, the more important distinction is whether data are missing at random or systematic patterns exist. This influences whether missing data can be imputed or not. Readers interested in this topic should consult with statisticians or the literature for guidance.
Newer Monitoring Technologies and Methods
Considerations with Consumer-based Monitors
The consumer marketplace has been flooded with an array of activity monitors designed to enhance self-monitoring and behavior change, and these features also have led to interest among researchers and health professionals. Products have been released with little or no evidence of reliability and validity, but researchers have started to identify potential strengths and limitations of the various devices. Evidence suggests that the accuracy of some consumer monitors may be comparable to findings from other, established research monitors. However, it cannot be assumed that all monitors have similar utility. A key distinction is the relative utility of step count estimates from devices. Many products and smartphone apps can provide estimates of steps, but the accuracy is questionable in many devices when directly compared with pedometer counts or manual step counts. Consumer monitors may be fashionable and trendy, but they may have limitations when used in research applications. It is up to the researcher to ensure that the selected device has sufficient reliability and validity for the desired application. Efforts are underway to establish benchmarks or standards for accuracy within the wearable monitor industry, which will facilitate comparisons in the future.
Distinctions between Raw and Count-based Accelerometer Data
A major challenge with the use of accelerometry-based activity monitors has been the lack of standardization about the processing and filtering of the raw accelerometer data. The use of different filtering and processing methods by various manufacturers has prevented movement “counts” from one monitor from being compared to those of another monitor. As a result, support for tracking and processing “raw” accelerometer data has increased. In theory, this would enable standardization of output in terms of real acceleration units (i.e., g values) and promote standardization of methods using open source processing techniques. This transition has some advantages but it also has dramatically complicated the data processing methods. The sheer volume of data is one challenge. If data are collected on a minute-by-minute basis, researchers must process 1,440 lines of data per day of assessment. However, this number grows to 8.64 million lines when processing raw data at 100 Hz (100 samples per second). New, open-source macro processing methods are being released to facilitate the processing, but additional expertise and time are needed to process these types of files. The new methods offer considerable promise for standardization in the future, but they present challenges for current researchers and practitioners interested in using them in studies or projects.
Applications of Pattern Recognition Methods for Activity Classification
The availability of raw data at a low resolution, as described previously, creates a variety of opportunities to enhance the accuracy of activity monitors. Traditional cutpoints based on counts are still widely used, although the field is evolving to more advanced methods that can make use of large amounts of data to predict activity type or posture and use this information to estimate energy expenditure. The accuracy of single prediction equations to estimate energy expenditure is influenced by the activity being performed because the relation between energy expenditure and movement counts varies from activity to activity. Machine learning is a popular method in computer sciences but just recently has been used in physical activity research. This method involves selecting and extracting features from movement signals obtained from wearable sensors such as an accelerometer. Examples of actual features from movement signals include for wrist algorithms (from g’s) the mean of vector magnitude (vm or mvm), the standard deviation of the vector magnitude (sdvm), the percentage of the power of the vector magnitude that is in 0.6–2.5 Hz range, the mean angle of acceleration relative to vertical on the device (mangle), the SD of the angle of acceleration relative to the device, and many others. The variability in the raw acceleration signal is then used to detect patterns and create activity classification schemes using advanced methods such as Hidden Markov models or Random Forests models. The resultant models are known to be able to differentiate between a set of postures (e.g., sitting vs. standing vs. walking) and also determine absolute activity intensities. Therefore, they can create implicit prediction equations based on the activity type or posture detected. This method can overcome the limitation associated with existing calibration equations. However, this method is still in the early phases of development. More information is available elsewhere.85
t The National Health Examination Survey (NHANES) is a combined surveillance program led by the National Center for Health Statistics (NCHS) and Centers for Disease Control and Prevention (CDC) that tracks health indicators in the U.S. population. NHANES has collected data every 2 years since 1999 and uses a complex, multistage, probability sampling design of all ages to adequately characterize the U.S. population.