Accelerometers are instruments which measure acceleration, the change in velocity of an object over time (SI unit: m.s-2). Acceleration is directly proportional to the force acting on the object to move it (as is the mass of the object).
Physical activity can be regarded as the change of position of body segments resulting from skeletal muscle contractions. A measure of acceleration of body segments can therefore be used to infer intensity of physical activity over time, allowing derivation of activity dimensions such as duration, frequency and overall volume. If acceleration and mass (including any external loading) of all major body segments are measured, whole-body energy expenditure due to physical activity can be estimated using the work-energy theorem. In practice, we seldom measure acceleration of the whole body or have any knowledge of external loading; this imperfect knowledge is therefore a source of error when inferring energy expenditure from accelerometry data. Like most other objective methods, accelerometers do not provide contextual information. Methods to identify activity type from raw accelerometry have been proposed (Janidarmian et al., 2016; Preece et al., 2009; Willetts et al., 2018).
Table P.3.5 Physical activity dimensions which can be assessed by accelerometer.
|Dimension||Possible to assess?|
|Physical activity energy expenditure||✔|
|Timing of bouts of activity (i.e. pattern of activity)||✔|
|Contextual information (e.g. location)|
Technological advances have resulted in instruments that can measure acceleration accurately, over extended time periods (a week or longer), and that are sufficiently compact and discrete for people to wear. All accelerometers have two essential parts: 1) a transducer or sensor which senses acceleration; and 2) a data acquisition system which processes and stores the data. The options for the sensor component fall into two basic types:
Accelerometers come in various models and specifications, and from many different manufacturers (brands) (Welk, 2000). It is important to emphasise that validity is tied to the overall method, rather than solely to the accelerometer make or brand. The properties of the captured data depends on number of acceleration axes, piezo-electric or MEMS-based, resolution and dynamic range, anatomical attachment site, and degree of filtering / data processing (if any) before data storage. The method includes all these components but also subsequent data processing decisions of feature extraction and inference.
It is not possible to recommend one device over another without knowing the specific aims of a particular study; however, the more information that is captured, the greater the options are for developing or applying more sophisticated methods to estimate specific variables of interest. Links to specific instruments are provided in the instrument library. The following should be considered when choosing between accelerometers:
Accelerometers can measure acceleration in one, two or three directions, correspondingly denoted uni-axial, bi-axial, and triaxial accelerometers (Chen et al., 2012). Naturally, it is preferable to measure all three dimensions of the physical world (otherwise the device is “blind” to movement in one or two directions), and the majority of contemporary instruments are capable of this. Inference based on uni- and bi-axial acceleration implicitly relies on pre-dominance of importance of one or two of the axes, or cross-axes correlations to capture the complete physical movement of the body segment to which the instrument is attached. Tri-axial accelerometers are therefore more sensitive to certain types of activity where movements are more variable in three-dimensional space, such as climbing, jumping or spontaneous play (Ott et al., 2000).
Sampling frequency and data storage
Modern accelerometers have sufficient storage capacity to store acceleration signals at frequencies sufficiently high to reproduce the acceleration waveform over multiple days (thus, sometimes referred to as waveform accelerometers). For example, the instrument used in the UK Biobank study is capable of storing tri-axial acceleration at 100 Hz for 14 days (Axivity, 2013). Battery life and storage capacity permitting, recording sampling frequency should be as high as possible to enable a wider range of methods.
Due to memory and battery limitations, some older accelerometer instruments initially sample acceleration at frequencies between 10-32 Hz before conducting on-board data processing and feature extraction, summarising signals ‘on-the-fly’ as counts stored at a user-defined epoch. As above, shorter epoch durations are desirable to optimise the resolution of the captured information as data can be down-sampled if required; however in some instruments only minute-by-minute resolution is available which still provides over 10,000 measurements in a week. Regardless of the resolution of raw stored data, a separate decision can be made on the analysis epoch duration; the lower limit of this is constrained by the resolution of the raw data but it can otherwise be tailored to the research question. A common analysis epoch for raw waveform acceleration data is 5 seconds, as for example used for the summarised movement intensity data in the UK Biobank data showcase (Doherty et al. 2017) or WHO surveys (Westgate et al., 2019).
The body segment that the accelerometer is attached to is a strong determinant of what information is captured and recorded. The interpretation of the resulting signal has to take the anatomical placement into account, because the biomechanical profile of different activity types determines the relationship between the acceleration of one body segment and the acceleration of the rest of the body segments.
The most appropriate position of an accelerometer depends on the study question and feasibility considerations. The hip or lower back allows for tracking the movement of the largest and most central part of the body, the trunk. A seismic accelerometer attached to the thigh may be used to estimate posture from its orientation with respect to gravity. The wrist is generally considered to be the most acceptable wear location to participants during free-living monitoring, and is therefore increasingly used in surveillance systems where representativeness of the sample is a particular concern.
During some activities, the acceleration of one body segment may not be representative of other body segments or the body as a whole; one such example of this is upper and central body movement whilst cycling, where measurement of acceleration at the wrist or trunk would likely result in under-estimating energy expenditure.
Number of monitors
As indicated above, the data that can be captured by a single accelerometer only yields incomplete information about an individual’s physical activity across a range of activity types. More complete information can be collected by measuring acceleration at two or more body locations simultaneously, which may enhance inferences about physical activity. However, it is only recently that accelerometer hardware and price have reached the point where multi-monitor measurement might be considered a feasible option for population research, so it has not yet been implemented in any large scale studies; accordingly, development of the methodology for interpreting multiple acceleration signals in a complementary manner has attracted considerably less attention but there are examples (Staudenmayer et al., 2009, Straczkiewicz et al., 2016, White et al., 2019).
Using labelled data collected in a laboratory, classification (e.g. machine learning) models utilising raw acceleration signals recorded at multiple body sites are better able to discriminate between different activity types (Bao & Intille, 2004, Preece et al., 2009a, Preece et al., 2009b). Another study showed that models based on acceleration measured at wrist and thigh locations had slightly lower error in predicting activity energy expenditure during free-living than either signal alone (White et al., 2019). Further methodological work is required to understand how best to implement and interpret different monitor configurations.
Naturally, the increase in information data capture has to be weighed against the increased burden upon the participant, which can cause issues with reactivity and reduce protocol adherence.
Duration of monitoring period
If inferences are to be made about habitual physical activity (whether this is in terms of volume, type, intensity etc.), it is necessary to consider the duration of monitoring, ie number of days of measurement required to capture a stable average estimate of habitual activity, given day-to-day within-person variation in physical activity. Minimum criteria can be applied based on the within-and between-individual variation of the variable of interest, as described in the section on physical activity variability.
Accelerometry is the most common objective method used to measure physical activity in population studies; it has been used extensively in field settings for:
Accelerometers were initially used as an outcome measure in small studies and a criterion measure to compare with self-report data. As accelerometers have become cheaper they have increasingly been used in large studies (Wijndaele et al., 2015), including thousands of participants in the Pelotas birth cohorts (da Silva et al., 2014) and National Health and Nutrition Examination Survey (NHANES) (Troiano et al., 2014), and 100 thousand in UK Biobank (Doherty et al., 2017). In addition, commercial wearables, smartphones, and other "smart" everyday devices such as cameras and toothbrushes include accelerometers.
When raw acceleration data is provided in SI units (m.s-2), processing data and estimation of dimensions of physical activity may follow a series of user-defined steps, all of which affect the validity of the final inference.
The measured value of a raw acceleration signal contains three components:
In order to assess physical activity, one may wish to isolate the human movement component of the signal (van Hees et al., 2012). Interpreting this signal in a reliable way, and the methodology for doing so is still under active development. While there is no consensus on standard procedures for making such inferences, an example pipeline is summarised in Figure P.3.3, which is described in detail below.
Figure P.3.3 Commonly applied processing steps for deriving physical activity summary variables from raw acceleration data. Adapted from: Doherty et al., 2017.
Preparatory steps involved in processing raw accelerometry data include:
Vector magnitude processing
These steps summarise the tri-axial acceleration movement component as a single-value average over an analytical epoch duration defined by the researcher.
Figure P.3.4 Calculation of the vector magnitude (VM) using the Euclidean norm of the x, y and z axes.
Figure P.3.5 Use of high-pass filter (HPFVM) and Euclidean norm minus one (ENMO) techniques to remove gravity component of raw acceleration signal.
It is important to recognise that regardless of the wear protocol dictated by the study design, there are likely to be periods where study participants have not worn their device. These periods can be identified in a number of different ways.
Figure P.3.6 Difference between true non-wear (left) and wrist acceleration during sleep (right). Panels are x, y, z, and high-pass filtered vector magnitude.
Brand-specific software packages are available which enable users to process data, as described in the section below. In some cases, it is also possible to export data and conduct all or some of the above steps in freely available software.
Physical activity summary variables
The steps in Figure P.3.3 result in a summary of physical activity as mean acceleration (g or mg) per user-defined epoch. For count-based accelerometers the unit will be in counts per user-defined epoch. Additional variables such as average acceleration by day and hour, or time spent at different acceleration levels at participant level (see Figure P.3.7) can be derived to investigate patterns of activity and compare individuals.
Figure P.3.7 Cumulative time spent in various movement intensities by sex for 45-54 year old women (white) and men (grey) in the UK Biobank Study. Source: Doherty et al., 2017.
Note on processing proprietary “count” accelerometer data
Older accelerometer store data in proprietary formats (i.e. counts), and several processing steps occur on-board the instrument. The details of these on-board processing steps are often known only to the manufacturer and are not decided by the user. It is important to acknowledge that these data processing ‘decisions’ made by any given monitor makes the stored information fundamentally different to the original acceleration signal. It is also important to acknowledge that some older count-based accelerometers capture movement in only one or two directions.
Physical activity energy expenditure
Accelerometers do not directly measure activity energy expenditure but there is a natural relationship between bodily movement and energy expenditure which can be exploited by predictive models. In practice, the characterisation of this relationship is complex because it varies by the body segment being measured and the activity type being performed.
The relationship between accelerometer recordings and energy expenditure is often studied by collecting data in a laboratory, where participants can be measured contemporaneously with a gold-standard measure of energy expenditure, such as respiratory gas analysis using facemask/mouthpiece or inside a calorimeter. In a typical study design, participants are often asked to perform a set routine of activities of varying intensity (Freedson et al., 1998; Swartz et al., 2000; van Hees et al., 2011) .
The overall relationship between activity energy expenditure and uniaxial acceleration measured at the waist during rest, walking, and jogging is fairly linear; however, deviations from linearity occur for high-intensity running (Brage et al., 2003a; Cavagna et al., 1976), for which movements are better captured with additional measurement of the antero-posterior axis of acceleration (Brandes et al., 2012). Linear relationships derived for rest and ambulation show much poorer validity in biomechanically diverse activities, such as cycling or lifting weights. Non-linear statistical models have been proposed to improve prediction equations. For example, a 2-segment regression model has been shown to improve accuracy of energy expenditure estimates for some activities, compared with simple regression (Crouter et al., 2006).
Those intending to use models derived by laboratory studies should critically evaluate the study population, and judge how appropriate it is to generalise from the activities performed in the lab setting. Laboratory studies may not reflect relationships between acceleration and energy expenditure in free living, and laboratory-derived prediction equations have been found to substantially under- or overestimate free-living energy expenditure if implemented non-critically (Corder et al., 2008; Ellis et al., 2016).
The validity of different accelerometer methods have been reviewed, demonstrating large variability in mean bias and correlation with estimates of energy expenditure using the doubly labelled water (DLW) method (Plasqui & Westerterp, 2007). Studies using DLW have shown that, in children, accelerometers explained 13% of DLW-measured PAEE variance and 31% of TEE variance. In adults, explained variance was higher, 29 and 44% for PAEE and TEE (Sardinha & Judice, 2016). Studies have also shown differences in values both within- and between-models (Brage et al., 2003b; Freedson et al., 2005; Ried-Larsen et al., 2012).
Using “cut-points” to estimate time spent in various intensity categories
A model of the relationship between acceleration magnitude and rate of energy expenditure (intensity) can be used to derive “cut-points” that correspond to a given energy expenditure value. For example, a researcher interested in quantifying time spent at or above “moderate” intensity may want to find acceleration intensity cut-point that best discriminates between < 4 METs and > 4 METs.
However, cut-points for defining different intensity levels in relative metabolic terms are somewhat arbitrary and the use of different cut-points can have a profound impact on estimates of physical activity (Freedson et al., 2005, Matthew, 2005). A researcher using accelerometry must understand the derivation of prediction equations from calibration studies and the rationale and implications of choosing a particular set of cut-points (Matthew, 2005; Rowlands et al., 2013 ). For example, published cut-points for sedentary behaviour from one accelerometer vary from 100 cpm to 1100 cpm (Atkin et al., 2013). Similarly, the range of cut points for moderate-intensity activity varies between 200 cpm to 3000 cpm.
Overall, cut-points make limited use of the wealth of detail available from raw acceleration data, often sufficient only for a crude approximation of the intended outcome (Matthews, 2005). Given the arbitrary nature of the count-based cut-offs, and the difference in unit expression across accelerometer models, reporting accelerometry data in standardised units of acceleration (m.s-2) is recommended (Brage et al., 2003a; Corder et al., 2008; Freedson et al., 2005). Extraction of signal features and patterns from raw acceleration data may improve PAEE estimation, and offers the ability to make inferences about posture from limb angles (Rowlands et al., 2015).
Using raw accelerometer data, it is possible to make inferences about an individual’s posture. When the instrument is stationary, it should measure 1g of acceleration in total over all 3 axes. The ratio of X:Y:Z indicates the direction gravity is acting on the device. From this, pitch and roll can be calculated (assuming the accelerometer X and Y axes align with a body segment's transverse and longitudinal axes, respectively):
One way to think about this is to liken the accelerometer to a spirit level. For example, placing an accelerometer on the thigh will allow us to measure thigh pitch, i.e. angle with horizontal; this feature has been used to discriminate between sitting and standing (Edwardson et al., 2016). Others have used the orientation patterns of the wrist to infer sleep episodes (Van Hees et al., 2015) or describe sedentary behaviour (Rowlands et al., 2013; Rowlands et al., 2015).
Physical behaviour type classification (human activity recognition)
An acceleration trace recorded by a modern raw accelerometer is of sufficiently high resolution and fine detail that individual motions and gestures leave discernible patterns and signatures; activity classification (or recognition) is the process of using these data to automatically identify types of activities, such as sitting, standing, walking or running (Preece et al., 2009b; Janidarmian et al., 2016).
Identifying activities from acceleration traces is a challenging task, and is commonly done using supervised machine learning techniques such as simple neural networks (Staudenmayer et al., 2009), random forests (Bao & Intille, 2004; Hammerla et al., 2016; Willetts et al., 2017) or deep learning (Avilés-Cruz et al., 2019; Nawaratne et al., 2020), whereas others adopt a prescriptive approach and design classifiers from first principles (van Hees et al., 2013; Urbanek et al., 2018). Supervised learning requires models to be trained with labelled data, which is typically acquired by direct observation in a laboratory setting and is expensive in terms of time, labour and equipment; this has prompted the exploration of alternative data collection methods such as wearable cameras to capture free-living behaviours (Doherty et al., 2013).
Activity recognition is often formulated as a short-term decision problem, where the acceleration data is chopped up into many frames of a fixed length (usually less than a minute), and a model is used to classify that short sequence (Janidarmian et al., 2016). A feature extraction process is applied to describe the data in that time window, and this feature vector is used as the input to the model; for example, features describing the frequency domain are a common choice because they are naturally suited to capturing repetitive motions typical of ubiquitous human activities such as walking (Preece et al. 2009a; Preece et al., 2009b). However, recent advances indicate that the techniques of deep learning, which do away with the explicit signal feature engineering step, have the potential to supersede traditional machine learning approaches (Lane et al., 2015; LeCun et al., 2015).
It is widely accepted that sensor data collected at multiple body sites is more informative for activity classification (Bao & Intille et al., 2004; Preece et al., 2009b); however, prediction models reliant on multiple sensor input signals cannot be utilised by the majority of studies that typically only administer one device per participant.
Practitioners intending to rely upon the output of an activity classifier should critically evaluate its reported accuracy, and carefully consider the consequences of misclassification. It should be noted that while many lab studies are reporting high classification accuracies (Saez et al., 2016), validation in free-living is relatively scarce and notably less impressive (Plasqui et al., 2012), which is perhaps why activity classification estimates from single devices has not yet reached mainstream adoption but the approach provides a complimentary interpretation of accelerometry data to that of directly measured movement intensity-based metrics.
Characteristics of accelerometers are described in Table P.3.6.
Missing data / non-wear time / non-compliance:
Cost and resources:
Table P.3.6 Characteristics of accelerometers.
|Number of participants||Small to very large|
|Risk of reactivity bias||Low|
|Risk of recall bias||Low to high|
|Risk of social desirability bias||Yes|
|Risk of observer bias||No|
|Risk of social desirability bias||No|
|Risk of observer bias||No|
|Participant literacy required||No|
Considerations relating to the use of accelerometers for assessing physical activity are summarised by population in Table P.3.7.
It has been suggested that establishing the relationship between acceleration data and energy expenditure is especially problematic in children due to their growth and development which affects estimates of resting metabolic rate and energy expended (i.e. movement economy) during activity. Children’s resting metabolic rate expressed relative to body weight decreases with age and maturation, and similarly the energy expended relative to body mass during walking and running also decreases (movement economy improves) with age (Krahenbuhl & Williams, 1992).
Table P.3.7 Physical activity assessment by accelerometers in different populations.
|Pregnancy||Waist-worn devices may be problematic. Depending on term, consider placement of monitors to avoid discomfort. Skin can also be more sensitive during pregnancy which could increase risk of irritation.|
|Infancy and lactation||Consider safety of attachment (should not be removable by infant to limit choking hazard).
Depending on the age, posture/orientation may also be very different, especially when still crawling.
When carried, activity of parent/carer is measured rather than of infant.
|Toddlers and young children||High sampling frequency recommended to capture intermittent patterns of activity. Consider safety of attachment (should not be able to be removed by infant).|
|Adolescents||Size and design of devices may negatively affect adherence.|
|Adults||There may be occupational complications (e.g. food preparation, nursing).|
|Older Adults||If memory impairment is a concern, low-maintenance devices are preferred. For example, a device that can be worn comfortably overnight, so it will not be forgotten in the morning.|
Administration of accelerometers
Achieving adequate wear adherence may be difficult in some populations, such as adolescents. The following points may enhance compliance:
A list of specific accelerometer instruments is being developed for this section. In the meantime, please refer to the overall instrument library page by clicking here to open in a new page.