Key words
individual response - exercise physiology - physical activity - non-responder
Introduction
Health organizations provide exercise recommendations aimed at reducing chronic
disease and premature mortality [1]
[2]
[3]. While regular exercise is generally associated
with health benefits, studies dating back to the 1980s report the magnitude of
change in a given outcome varies widely between individuals, to the degree that some
are seemingly unable to garner the expected benefits [4]
[5]. The seminal HEalth, RIsk factors, exercise
Training And Genetics (HERITAGE) Family Study was among the first to demonstrate
heterogeneity in the observed responses following a standardized aerobic exercise
training intervention [6]
[7]. A recent review by Williamson and colleagues
[8] critiqued these findings mainly because the
HERITAGE Family Study lacked a control group and referred to short-term reliability
– a limitation discussed in more detail in the next section. Although many studies
following HERITAGE also lacked control groups [8], the
number of reports investigating heterogeneity in observed responses has constantly
grown [5]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17].
A by-product of investigating exercise treatment response heterogeneity has been a
more clinically applicable focus on the individuals who struggle to experience the
anticipated benefits, typically referred to as ‘non-responders’. Accurate
interpretation of an individual’s ability to experience the targeted benefits – or
not – from a prescribed exercise program would allow clinicians to make an informed
decision and adapt the exercise program parameters or implementing a different
treatment regimen. Moreover, identifying the mechanisms contributing to an
individual’s ability to benefit from an exercise program may provide insight to the
true effectiveness of exercise and allow for personalized exercise prescriptions,
moving beyond the standard recommendation of achieving current physical activity
guidelines in a clinical setting [18]
[19].
It is important to recognize that investigating the influence of exercise training
on
treatment response heterogeneity, and determining if a participant responded to an
intervention (or not) are two distinct analytical approaches, designed to answer
different questions. Conflating the two may lead to inappropriate interpretation of
results. Common errors include interpreting improvements in the group mean as an
exercise intervention increasing the ‘response rate’, or associating higher response
rates with reduced treatment response heterogeneity [20]
[21]. Moreover, as the desire to understand exercise
treatment response heterogeneity and individual response has grown, the number of
proposed methods for conducting such research has significantly inflated.
Accordingly, there is an ongoing debate on the most appropriate, rigorous and
feasible techniques [8]
[19]
[20]
[22]
[23]
[24]
[25]
[26]
[27]
[28]. Various questions and concerns have emerged,
including the ability to truly isolate the impact of exercise training, the nature
of individual responses in relation to mean group changes, and the utility of
individual response research [20]
[29]
[30]
[31]
[32].
The purpose of this review is threefold. First, the strengths and limitations
associated with current methods for quantifying the contribution of exercise to
observed treatment response heterogeneity will be discussed. Second, current methods
used to categorize participants based on their response to exercise will be
outlined, as well as proposed mechanisms to identify factors that contribute to
response variation. Finally, this review will provide an overview of some current
issues at the forefront of individual response research.
Sources of Variation Contributing to an Individual’s Observed Response
Sources of Variation Contributing to an Individual’s Observed Response
Prior to discussing exercise treatment response heterogeneity or individual response
categorizations, it is important to understand that individual observed changes
following an exercise intervention are the product of some combination of random
variation, within-subject variation, and the subject-by-training interaction. As
these sources of variation have been explained in depth previously, below we provide
a succinct overview [19]
[22]
[23]
[25]
[28]
[33]
[34]. It should be noted the terminology used in this
review – as defined below – is inconsistently applied across the literature, with
alternative terminology outlined by Bonafiglia et al. [23].
Random variation
Random variation is comprised of the error introduced by the measurement
instrument (technical error) and day-to-day biological variation. When multiple
measures are taken over a period of time within which the true value is not
expected to change, the noise introduced by random variation will result in a
cluster of observed scores that are normally distributed around the true value.
Using the average value of multiple measurements for analysis can reduce the
influence of random variation [33]. Likewise,
random variation can be estimated by collecting repeated measurements on a group
of individuals and calculating the typical error of the measurement
(TEM) [26]
[28]
[33].
Within-subject variation
Within-subject variation is inconsistent variance introduced by changes in the
environment or behaviour unrelated to the intervention (e. g., short-term
changes in eating patterns, seasonal changes influencing behaviour or mood)
[8]. Theoretically, if the same individual
were provided the same intervention at a different time, within-subject
variation would be responsible for much of the variance in the observed
difference in the change scores, if random variation was removed. The impact of
within-subject variation on a participant change score may be dependent on the
duration of the intervention, with longer duration trials expected to increase
its influence [23].
True responses to exercise training
The true response to exercise represents the ‘trainability’ of an individual, or
the consistent, repeatable changes experienced in relation to the provided
intervention. Although genetic endowment may contribute to an individual’s
trainability, stable characteristics or traits such as lived experiences,
lifestyle habits (e.g., exercise, diet), and epigenetic modifications may also
influence the ability of an individual to respond to the training stimulus [11]
[25]
[26]. The subject-by-training interaction refers
to the degree to which true responses to training (i.e. trainability) differs
across a group of participants.
Methods for Quantifying Exercise Treatment Response Heterogeneity
Methods for Quantifying Exercise Treatment Response Heterogeneity
Exercise treatment response heterogeneity exists when the true training-induced
changes experienced across a sample of participants differ to a degree that can be
considered meaningful [20]
[23]. Quantifying exercise treatment response
heterogeneity requires the subject-by-training interactions to be isolated from
within-subject and random variation [25]
[34]
[35], necessitating a crossover trial with multiple
intervention and control periods, each separated by an adequate washout [35]. The true value of the subject-by-training
interaction can then be calculated using a linear mixed model approach [25]. While this study design may be possible in certain
research areas [36], conducting such a trial using an
exercise intervention presents several limitations: high operating costs, heavy
resource requirements, significant time investments, and challenges with participant
recruitment and compliance. Moreover, the carry-over effects of an initial training
intervention are not well enough understood to accurately suggest an adequate
washout period between phases [8].
Alternatives for estimating exercise treatment response heterogeneity have been
proposed, each of which must complete two steps to increase accuracy. First, the
observed heterogeneity among participant change scores must be shown to be
attributable to exercise training per se, and therefore not primarily a
product of random or within-subject variation. Second, the magnitude of
heterogeneity throughout the sample should be contextualized in relation to the
chosen outcome measure [27]. Contextualizing the
magnitude of the heterogeneity helps to determine if the variance introduced by
exercise is meaningful in relation to the anticipated variation associated with the
measurement technique. [Table 1] outlines the various
statistical methods proposed for estimating treatment response heterogeneity, the
subject-by-training interaction, random variation, and within-subject variation. The
following sub-sections will provide additional detail for some of the notable
methods.
Table 1 Proposed methods for assessing variation†.
|
Method
|
Statistical Approach
|
Outcome
|
Potential Limitations
|
|
Repeated measurements at individual timepoint
|
TEM (TEM = SDdiff / √2)
|
Estimate of random variation
|
Does not account for within-subject variation
|
|
Replication of a cross-over trial
|
Linear mixed model Fixed Effect: Training Group Random Effect:
Subject Identity
|
Determines individual subject-by-training interaction
|
Feasibility Unknown wash-out duration
|
|
Comparator Group: Control Group
|
SDIR = √(SDEXP
2 − SDCON
2)
|
Standard deviation of individual responses; Group-based estimate
of the variance across participants’ subject-by-training
interaction
|
Potential differences in variation between the control and
training groups
|
|
TEΔ (TEΔ = SDdiff / √2)
|
Estimate of within-subject variation.
|
May also contain random variation
|
|
Comparator Group: Reliability Data
|
SDIR = √(SDEXP
2 – [√2 * TEM]) or SDIR =
√(SDEXP
2 – [√2 * CV])
|
Standard deviation of individual responses; Group-based estimate
of the variance across participants’ subject-by-training
interaction
|
Transferability of sample used to calculate the reliability data
Duration of previous trial
|
|
Repeated measurements throughout the intervention
|
Linear mixed model Fixed effect: measurement number Random
effect: subject ID-by-measurement number interaction Dependent
variable: measured value
|
Individual estimates of subject-by-training interaction
|
Increased demand on participants and resources Potential
influence of accumulating tests
|
†Interpreted based on Table 3 from Hecksteden et al., 2015 and
Table 2 from Ross et al., 2019. TEM = typical error of a
measurement; SDdiff = standard deviation of the difference score;
SDIR = standard deviation of individual responses;
SDEXP = standard deviation of the change scores from the
experimental group; SDCON = standard deviation of the change
scores from the experimental group; TEΔ = typical error of a
change score; CV = coefficient of variation.
Control group designs: SDIR method
If exercise training meaningfully contributes to treatment response
heterogeneity, exercising participants will display greater variation in their
pre-post difference scores than a non-exercising control group [27]. As the control group did not participate in
the intervention, it is assumed variance among the difference scores is a
product of random and within-subject variation. Conversely, the intervention
change scores also include variance introduced by individual responses to
training (i.e. subject-by-training interactions). Therefore, the magnitude of
variation introduced by participating in the exercise program can be estimated
by subtracting the variance of the control group change scores, from the
variance of the intervention group changes [23]
[27]
[28]. The resulting value is referred to as the
standard deviation of individual responses (SDIR), which estimates
the influence of individual responses to exercise training on overall response
heterogeneity, accounting for the influence of random and within-subject
variation [8]
[22]
[23]
[27]. If the standard deviation of the changes in
the intervention group are not substantially larger than the control group,
theoretically the exercise training has not contributed to the observed variance
[8]
[22]
[27].
The SDIR method relies on the assumption that larger variance among
the change scores in the experimental group, when compared to controls, is
sufficient to estimate the magnitude of treatment response heterogeneity
attributable to exercise training. To improve the accuracy of the
SDIR, steps should be taken to improve estimates of
within-subject and random variation. Ensuring the control group follows the same
time interval as the intervention group [8], and
taking multiple measurements at every timepoint (for both the control and
intervention groups) can reduce these influences [33]. Simply stated, researchers should follow proper randomized
control trial practices to ascertain the impacts of participating in the
provided intervention [20].
A number of additional assumptions and limitations require consideration when
using the SDIR. Most notably, it is important to consider that the
SDIR relies on the assumption that the combined effect of random
and within-subject variation is equal between the intervention and control
groups [23]. Even with randomized allocation to
control and intervention groups, an inability to calculate within-subject
variation across each group leaves the potential for its influence to differ.
Accordingly, this assumption should be reported as a limitation whenever the
SDIR is used.
Despite a necessary reliance on the assumptions associated with implementing
randomized control trials in exercise science [23], the SDIR remains the preferred method for estimating the
influence of exercise training on observed treatment response heterogeneity.
This is in large part due to the SDIR being, to our knowledge, the
only method able to estimate the magnitude of confounding sources of variation
in parallel-arm randomized control trials [17]
[24]. Nonetheless, it is important to recognize
there are many statistical methods designed to compare the equality of variance
across groups. For example, Leifer et al. [37]
compared variability in observed responses between exercise and control groups
using Levene’s test. We find the SDIR to be preferred over such tests
as it estimates the magnitude of treatment response heterogeneity in units of
the measured outcomes rather than relying on the interpretations of a
p-value. However, to maximize confidence in the utility of the
SDIR, potential threats to major assumptions need to be avoided,
and if present, reported.
Using reliability data
If a control group is not available, using data from a relevant reliability study
(i.e., test, re-test) has been proposed as an alternative [25]
[26]
[27]. When using reliability data, it is suggested
the SDIR can be estimated by subtracting the variance of test,
re-test change scores from the variance in change scores following exercise
training. However, ‘replacing’ a control group should not be done without
concern, as the limitations when using reliability data to calculate the
SDIR far outweigh the benefits. Data from a relevant reliability
trial will solely allow for an estimate of random variation, meaning
within-subject variation remains unaccounted for. Accordingly, the
estimate of what would have happened to intervention participants had they not
participated in the intervention is no longer valid; meaning the SDIR
cannot distinguish variability in true changes attributable to exercise training
from the changes resulting from behavioural and/or environmental factors.
Additionally, differences between the intervention participants and the
reliability group can significantly impact the results. As such, we recommend
against using reliability data to calculate the SDIR.
Repeated measures design
Hecksteden et al. [25] introduced a longitudinal
approach using the collection of repeated measurements throughout the duration
of an intervention as an alternative to calculating the SDIR. The
concept has been subsequently demonstrated on two occasions [26]
[38]. When using the repeated measurement design,
alternative tools for detecting exercise treatment response heterogeneity become
available. Namely, true response estimates are derived from the slope of each
individual’s regression line of the measured values throughout the duration of
the intervention. As opposed to calculating the SDIR based on the
variance among the change scores from the intervention and control groups or
reliability trials, taking repeated measures allows for the SDIR to
be estimated by calculating the between-subject standard deviation of individual
slopes or by using a linear mixed model. Importantly, and in line with any
single-group design, this method cannot account for the counterfactual, and
therefore cannot discern whether the variance introduced was the product of
exercise training or behavioural/environmental changes that occurred during the
intervention, per se. Moreover, this method assumes that responses
generated over time will be linear, which may not be the case for certain
physiological variables (e.g. cardiorespiratory fitness) [38]. Frequent testing may also introduce learning
effects, carry-over effects or performance biases, masking results. Therefore,
using a control group to calculate the SDIR remains the preferred
method.
Contextualizing the magnitude of exercise treatment response
heterogeneity
Once the SDIR has been calculated, its magnitude should be
contextualized to determine if the variance introduced by exercise is
meaningful. This can be done via standardization, or by comparing the
SDIR to a predetermined threshold value. Standardization consists
of dividing the SDIR by the standard deviation of all subjects at
baseline, and comparing them to threshold values of 0.1, 0.3, 0.6, 1.0, and 2.0
(representing small, moderate, large, very large, and extremely large effects)
[27]
[39]. Alternatively, the SDIR can be
viewed in relation to a predetermined threshold such as the minimal clinically
important difference, or the smallest worthwhile difference [28]
[33].
Summary
Understanding the contribution of exercise to the observed heterogeneity
following an intervention can aid researchers in interpreting the effects of
exercise. Currently, there are a number of limitations restricting the ability
to accurately quantify exercise treatment response heterogeneity. While multiple
methods have been proposed to provide estimates of the influence of exercise
training on the observed variance, the SDIR remains the preferred
metric. It is important to emphasize that a negative SDIR, or a
suggestion that treatment response heterogeneity was not only a product of
exercise training; does not preclude the possibility of responders and
non-responders being identified [16]. A negative
SDIR could suggest the observed heterogeneity was largely
influenced by other factors than participation in the prescribed exercise
training. Alternatively, negative SDIR values may reflect sampling
errors in estimates of observed variability due to small sample sizes, which
remains a major challenge for exercise training studies aiming to investigate
treatment response heterogeneity.
Categorization of Participants as Responders or Non-responders
Categorization of Participants as Responders or Non-responders
Categorizing participants as responders or non-responders speaks to the ability of
each individual to improve beyond a threshold value for a selected outcome, as a
result of participating in a specific exercise intervention. Such a categorization
requires a consideration and comparison to a control condition. It is important to
clearly define what being categorized as a ‘responder’ means for the current
investigation, prior to conducting an analysis. Categorization may occur in a
variety of ways, each requiring a response threshold to be selected, and a method
to
account for the variance limiting the ability to directly quantify an individual’s
true response [26]. The method used to categorize
individuals should be carefully selected to answer the research question and be
constructed in a way that accounts for the desired sources of variance.
Categorizations may be made based on estimates of random variation (see below), and
it can be argued these participants did experience benefit. However, without
comparison to a control group the observed changes cannot be assigned to the
exercise intervention, per se. Accordingly, we suggest participants only be
categorized as responders/non-responders when a control condition is part of the
design and considered when making the categorization. If a control group is not
considered, participants can only – at best – be categorized as experiencing a
benefit, or not experiencing a benefit.
Various types of response thresholds have been used throughout the literature,
including: (1) zero change as a fixed value, (2) the upper limit of observed
differences expected as a result of variation, and (3) the lower limit of clinically
or practically meaningful differences. Each threshold type differs in if, and to
what degree, a method of accounting for the variance masking an individual’s true
change score is intrinsically applied. If an estimate of the potential influence of
variance is not built into the threshold, additional steps should be taken to ensure
the accuracy of a response estimate prior to the response categorization. Moreover,
a threshold which is too permissive to answer the research question may result in
inaccurate response categorizations, and a threshold which is too conservative may
overstate the non-response rate [26]. Regardless of
the method used, the threshold will always remain arbitrary and debated [40]. The following section will outline the thresholds
used to categorize participants, and the methods which may be applied to account for
the variance limiting the accuracy of an individual response estimate.
Zero change as a fixed value
Here the response threshold is set at zero change, defining non-response as a
difference from pre- to post-intervention as zero or less ([Fig. 1a]), [11]
[41]. While straightforward, using zero change as
a threshold does not account for the limited accuracy of a response estimate,
allowing random and/or within-subject variation to bias the categorization. To
improve the accuracy of the categorization, an individual confidence interval
for each participant’s change score can be calculated using an
intervention-specific estimate of error [24]
[28]. Individuals can then be categorized as
successful or not based on whether the confidence interval for the true change
lays across or beyond the set threshold. Using a zero-based threshold,
individuals can be categorized as having a positive response, a negative
response, or an uncertain response, dependent on the interval laying entirely
above, below, or crossing zero, respectively.
Fig. 1 Participants categorized based on (a) zero change
as a fixed value, (b) estimates of variation or clinical
relevance, and (c) individual confidence intervals.
The interpretation of said categorization will be highly dependent upon the
estimate of error used to calculate the confidence interval. Error estimates
based on repeat baseline tests (TEM) can be used if the goal is to
categorize participants based on changes beyond an estimate of random variation.
Accordingly, these participants should be referred to as having ‘experienced
benefit’ or not. Conversely, if the goal is to more confidently categorize
responders based on changes beyond both random variation and within-subject
variation – thereby increasing the likelihood that observed changed resulted
from participation in the provided exercise intervention – a time-matched
control group should be used to estimate the combined influence of random and
within-subject variation by calculating the typical error of the change score
(TEΔ) [8]
[26]
[28], and using it to estimate the influence of
error. Subsequently, participants can be categorized as responders or
non-responders to the intervention.
Swinton et al. [28] provides adjusted multiples
that can be used to calculate the width of the confidence interval. The width of
the confidence interval is at the discretion of the user, understanding that
larger widths increase the risk of making type 2 errors (incorrectly
categorizing an individual as uncertain when they are likely to be responders or
non-responders), whereas smaller widths increase the risk of making type 1
errors (incorrectly categorizing as a responder or non-responder).
Thresholds based on estimates of variation
The aim when using an estimate of variation is to set the threshold at the
highest possible change that may occur as a result of extraneous variation.
Similar to using confidence intervals, the interpretation of response
categorizations will depend upon the estimate of error used to set the response
threshold ([Fig. 1b]) [42]
[43]
[44]
[45]
[46].
The first step when using an estimate of variation to set a response threshold is
to estimate the influence of variation on the observed changes. Therefore, a
decision needs to be made regarding what sources of variation will be accounted
for by the threshold value. Again, the TEM can be used to estimate
the influence of random variation on a single measure. However, a change score
is composed of two independent measurements, each subject to random variation.
To account for this, the estimate should be multiplied by a factor greater than
one prior to setting the response threshold, to increase the level of confidence
in subsequent categorizations [26]
[28]. Again. it must be emphasized that
individuals categorized using this method only experienced a change beyond an
estimate of random variation. This change cannot be attributed to participation
in exercise (or any other source) per se; all that is known is a change
greater than what was expected to occur due to random error has occurred, and
categorized as having experienced benefit, or not. Alternatively, a time-matched
control group can be used to estimate the combined influence of random and
within-subject variation by calculating the typical error of the change score
(TEΔ) [8]
[26]
[28]. By replicating the measurement schedule of
the intervention, the TEΔ estimates the variance that may be
introduced due to within-subject variation in the absence of the prescribed
exercise, while also capturing the influence of random variation [8]
[28]. Utilizing a well-structured randomized
control trial to set a threshold using the TEΔ means those
individuals categorized as responders were likely to have experienced a
beneficial change as a result of the provided intervention.
Once the influence of extraneous variation is estimated, the threshold can be
set. Although the TEΔ theoretically estimates the majority of
variance introduced throughout the duration of an intervention, some suggest
this estimate should also be multiplied by a factor greater than one to increase
the level of confidence in response categorizations [26]
[28]. However, multiplying the TEΔ may
inappropriately increase the threshold to a value which inaccurately categorizes
individuals as non-responders. Despite the estimate of variance being
empirically defined, the factor used to set the limits for response introduces a
potential limitation. As the chosen factor can only be legitimized by convention
and/or recommendation across the scientific community, disagreement will lead to
a variety of factors being implemented and inconsistent thresholds throughout
the literature. As a result, comparing results or responder prevalence across
trials can pose a great challenge.
Lower limit of practical relevance
A practically relevant response threshold exists at the border between a trivial
and a meaningful difference. The preferred method for setting a threshold at the
lower limit of practical relevance is to use a well-established minimal
clinically important difference. While using the smallest worthwhile difference
(0.2 multiplied by the standard deviation of baseline values) is an acceptable
alternative, the calculated value will be sample-specific and restrict
generalizability. An example of categorizations made based on the lower limit of
practical relevance can also be seen in [Fig.
1b], [16]
[38].
Similar to using zero change, a threshold based on practical relevance has no
inherent ability to account for the limited accuracy of a single response
estimate. As such, an individual confidence interval for each participant’s
observed change score should be calculated with the aforementioned
considerations and interpretations in mind [28].
Assuming the utility of a control group to construct the confidence intervals,
individuals can then be categorized as having a clinically meaningful positive
response (responder; confidence interval lays completely beyond the
threshold in the positive direction), a meaningfully negative response
(adverse responder; confidence interval lays completely beyond the
threshold in the negative direction), a non-response (non-responder;
confidence interval lays completely below the positive threshold and above the
negative threshold), or uncertain (confidence interval partially overlaps the
threshold) ([Fig. 1c]), [24]
[28].
Repeated measurements to calculate individualized confidence
intervals
As an alternative to using group-based estimates of error to calculate individual
confidence intervals, Hecksteden et al. [26]
outlines how to produce individualized response estimates based on repeated
measurements taken throughout a longitudinal intervention. This method
eliminates the need to assume equal variance throughout the sample, permitting a
more personalized analysis. Response estimates are calculated as the slope of
the individual’s regression line of observed values vs. time, with the scatter
of the observed values around the individual regression line
(TESLOPE) providing a method to calculate the uncertainty. These
values can then be used to calculate individualized confidence intervals.
Summary
Undeniably, there are numerous options available for setting response thresholds
and accounting for degrees of variance. The decision of which threshold to apply
or which sources of variance to take into account will significantly impact both
the categorization of numerous participants and the interpretation of the
results. Importantly, if the goal is to attribute a response categorization
directly to the provided exercise intervention, a high-quality randomized
control trial design must be used. As such, decisions regarding how response
categorizations will be interpreted must be made prior to study initiation, as
post-hoc decision making can lead to interpretation errors.
Methods for Investigating Factors Influencing Response Variation
Methods for Investigating Factors Influencing Response Variation
An important component to investigating exercise treatment response heterogeneity
and
individual response is to determine factors contributing to the differences in
response. Several reviews have discussed factors which may influence response
heterogeneity following exercise [47]
[48]
[49]. Here, we will outline the methods commonly used
to identify the moderators and mediators of exercise treatment response
heterogeneity and individual responses.
Moderators of the SDIR
As described in a series of articles from Atkinson and Batterham [22], Hecksteden et al. [25], and Hopkins [27], any variable
influencing treatment response heterogeneity would logically impact the
magnitude of the SDIR. As such, Atkinson and Batterham [17] proposed a method to determine the
SDIR and identify its moderators using a modelling approach and
adjusting for identified co-variates at baseline. Using a linear mixed model the
study arm (intervention or control) is entered as the fixed effect, an
additional binary ‘dummy’ variable is entered as a random effect (explained as
allowing for extra variance in the change scores in one group versus the other),
and the baseline value of the outcome entered as a covariate. The
SDIR may then be derived from the parameter estimate.
Subsequently, potential moderators can then be tested by including them in the
model and interpreting the resultant changes in the SDIR. The authors
recommend consulting with a statistician to ensure models are properly
applied.
The Atkinson and Batterham [22] method was
subsequently implemented by Hammond et al. [50]
who, despite large effect sizes, were unable to detect any statistically
significant predictors of exercise induced treatment response heterogeneity.
However, the authors provided meaningful commentary on numerous limitations
associated with this method. Most prominently, the utility of a linear mixed
model – while in line with good analytical practice – will likely require large
sample sizes to find statistically significant results. As an example, the
authors referred to their inadequate power with 181 participants, calculating
that 504 participants were required to achieve statistical significance.
Although samples of that size may pose limitations to individual research
groups, it is important to note a requirement of large sample sizes is not
unique to this modelling approach; small sample sizes are a challenge for all
methods estimating SDIR values and/or moderators treatment response
heterogeneity. High-quality analytical approaches are necessary, to accurately
elucidate moderators of exercise treatment response, meaning successful
collaborative efforts, or innovative, practical alternatives should be
pursued.
Categorization of ‘high’ and ‘low’ responders
Numerous authors have instead opted to divide their sample into groups of ‘high’
and ‘low’ responders before identifying moderators and/or mediators contributing
to the observed differences in the magnitude of change [48]. The two key steps to this method are deciding how to divide the
sample, and choosing a statistical method to identify the factors contributing
to the differences in the observed changes.
A common method used to separate participants into categories of high and low
responders is to break the sample into quartiles or quantiles, based on the
magnitude of each individual’s observed change [51]
[52]
[53]
[54]. Selecting a fixed proportion of participants
with the highest (or lowest) differences as the high (and low) responders
provides balanced groups with easily identifiable differences in response.
Alternatively, K-means clustering has been used [55]
[56]
[57]. While these methods are common, it is
important to consider that breaking participants into predetermined groups will
not adequately account for the limited accuracy of a response estimate and is
only meaningful to the current sample. As a result, it remains possible that
participants in the ‘high responder’ group may have not truly responded to the
intervention, or that participants in the “low responder” group did have a
meaningful response to the intervention.
Once divided, various techniques have been used to identify key characteristics
contributing to the differences in change scores, including regression analyses
[53]
[54] and ANOVA [52]
[55]
[56]
[57]. Alternatively, one group has utilized
Principal Component Analysis [51]. It is important
to reiterate that these approaches do not consider the influence of random or
within subject variability when categorizing individuals, nor do they quantify
the magnitude of observed treatment response heterogeneity, or attempt to
clarify the contribution of exercise, per se. Therefore, while these approaches
have been used previously and can indicate potential moderators or mediators for
the response categorizations, the observed changes may be influenced by
variation, or only be truly representative of the current intervention.
Summary
A primary purpose of individual analyses is to identify factors contributing to
an individual’s response categorization. However, in line with many aspects of
this research field, conducting such analyses requires a great deal of
assumptions. Likewise, researchers who choose to investigate these outcomes
should be prepared to experience a number of challenges. While the prospect of
identifying moderators and mediators of exercise response is intriguing,
identifying rigorous study design and analytical methods for doing so represents
one of the areas for future development within treatment response heterogeneity
and individual response research.
Current Questions Facing Exercise Treatment Response Heterogeneity and Individual
Response Research
Current Questions Facing Exercise Treatment Response Heterogeneity and Individual
Response Research
Can we move beyond a categorical approach when describing individual
responses?
There are advantages to identifying individuals who significantly benefit, and
seemingly struggle to benefit, from an exercise trial and categorizing them
accordingly. Doing so can provide clear cut points from which subsequent
analysis can occur; such as investigating the underlying mechanisms contributing
to these differences, or potentially working towards a future with personalized
exercise prescriptions. However, categorizing individuals as responders or
non-responders following a single trial comes with notable limitations. First,
any categorization only holds true in the context of the provided exercise
intervention, the selected response threshold, and the outcome of interest.
Adapting the exercise protocol, choosing an alternative response threshold, or
focusing on a different outcome can result in a different categorization [10]
[16]
[26]
[44]
[45]
[58]. As such, the generalizability of findings
may be severely limited. Second, categorizing an individual fails to consider
the continuous nature that probabilities of response may provide [24]
[28]. As a result, individuals with a high
likelihood of response may be classified as non-responders simply due to their
inability to improve beyond a subjectively chosen threshold, and an individual
who only slightly surpassed the response threshold is considered equal to an
individual who surpassed the threshold five-fold.
Swinton et al. [28] proposes a method to address
the limitations imposed by the categorical approach and move towards
likelihood-based classification, which were used by Bonafiglia et al. [24]. It should be noted that these methods make use
of the Magnitude-Based Decision Making technique [59], which has received much criticism [60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]. However, these critiques are aimed at the
group-level utility of these procedures, with no current debate regarding the
application of these methods on individual analyses. Nonetheless, research
attempting to transition towards an accepted, reliable mechanism for
likelihood-based decision making is warranted.
Are we truly able to identify exercise-training related changes on an
individual level?
One of the most important assumptions associated with individual analyses is that
the true response to exercise training is an identifiable, consistent,
reproducible trait. The origins of this assumption stem from studies using
selectively bred rats, monozygotic twins, and nuclear families, highlighting a
genetic component to the changes experienced following exercise training,
particularly in reference to measures of cardiorespiratory fitness [7]
[9]
[12]
[13]
[68]
[69]. While these studies suggest some degree of
reproducibility in the true response to exercise, evidence of this
reproducibility remains limited. Lindholm et al. [32] detected poor correlations for individual changes in exercise
performance following two identical training sessions separated by a washout
period, Islam et al. [30] reported poor
reproducibility among acute changes in gene expression following repeated
application of exercise, Islam et al. [31]
outlined non-significant correlations among skeletal muscle adaptations in
individuals who repeatedly completed an identical training regimen, and Del
Giudice et al. [29] found changes in
VO2max and time to fatigue were not reproducible and led to some
participants whose response categorization changed following identical four-week
high-intensity training regimens, separated by a three-month washout period.
There are several potential explanations for these findings, each with different
repercussions for individual response research.
First, it is possible the implemented washout periods were unable to account for
potential carry-over effects from the initial training phase, or that
participant behaviour was different prior to each of the training interventions.
Although some authors report similar performance metrics at the initiation of
each training phase [29]
[31], it remains possible that undetected
physiological alterations influenced the observed adaptation in subsequent
training periods. Improved understanding of the carry-over effects following
exercise training (including both a more holistic physiological outlook across
various outcomes and a better understanding of the duration of these effects)
would allow for more accurate washout lengths. Regardless, participants
experiencing differences in response following subsequent, identical training
periods highlights that a response categorization is highly specific to the
provided exercise intervention.
Second, it may hold true that the true response to exercise is a constant,
repeatable trait, and the current methods for separating random and/or
within-subject variation from the true response are not sensitive enough to
fully account for their influence. As such, our ability to isolate the
subject-by-training interaction may be inadequate. Regardless of the threshold
used, current methods for categorizing an individual as a responder or
non-responder neglect to account for what would have happened if that
individual had not participated in exercise, making it impossible to
definitively know if an individual responded to the exercise per se
[20]
[34]. Proposed mechanisms for setting response
thresholds or calculating confidence intervals provide estimates for
extraneous variation, but we currently cannot be certain that these influences
are entirely accounted for. This explanation would suggest a need for more
accurate estimates of the true response to exercise training, or improved
methods for accounting for random and/or within-subject variation. Doing so
would provide a more accurate indication of the influence of exercise training
on individuals. A better understanding of the carry-over effects of an exercise
trial and collaboration across research labs may improve the feasibility of
conducting high quality cross-over trials to address this concern. Until that
time, it may be wise to reconsider the ‘exercise responder’ terminology. We
recommend shifting away from categorizing participants as responders or
non-responders to exercise, and re-phrasing these determinations to who
responded beyond an estimate of random or within-subject variation.
Lastly, these findings may suggest there is true intra-individual variation in
response to exercise training, meaning the true response to exercise training is
not a stable, reproducible trait. This would pose a great challenge for the
future of personalized exercise prescription, and emphasizes the importance of
not categorizing individuals following a single trial. Instead, it may be more
worthwhile for research to focus on identifying the factors contributing to
higher and lower change scores among participants. Moreover, as opposed to
individualizing exercise prescription, practitioners may instead focus their
attention on those who do not benefit from an initial training intervention and
adapt exercise to garner improvements in the future [58].
It is important to discern the genuine nature of the true response to exercise
training, and our ability to accurately identify it. Notably, much of the
research questioning the identifiability of the subject-by-training interaction
has focused on outcomes related to fitness or muscle health in relatively young,
healthy men. Future investigations should aim to include various outcomes and
populations to confirm these findings and move the field forward.
Are response rates reflective of the individual, or truly a group
statistic?
The vast majority of research reporting response rates do so by counting the
total number of responders throughout the sample [41]
[42]
[43]
[45]
[46]
[70]. These numbers are often used to compare
interventions and determine the preferred method for reducing the quantity of
non-responders. Results often show higher volumes of exercise leading to reduced
non-response rates, or suggest increases in training volume reduce treatment
response heterogeneity and eradicate non-response [43]
[45]
[46]. Atkinson et al. [16] challenge this assertion and argue responder counts are highly
sensitive to changes in the group mean; concluding that these metrics are truly
representative of changes throughout the group – rather than individualized
analyses – and should be treated as such. Subsequent analyses conducted by
Bonafiglia et al. [21] supported this argument
highlighting how ‘response rates’ are reflective of differences in mean group
changes, not in treatment response heterogeneity or true individual responses.
Recent studies have hypothesized that, compared with exercising at relative
intensities (e.g. a percentage of VO2max), exercising above
physiological thresholds reduces heterogeneity in metabolic stress and thus
decreases interindividual variability in observed responses to exercise training
[71]
[72]
[73]. Although these studies reported that
threshold-based prescriptions increased response rates compared with
relative-intensity prescription, they did not statistically compare the
variability in observed response between groups. It is therefore unclear whether
larger response rates following threshold-based prescription are explained by
larger mean changes in the absence or presence of reduced interindividual
variability [21]. Future work should adopt a
statistical test (e.g. Levene’s test) to compare variability in observed
responses to exercise training prescribed at a relative and threshold-based
intensity.
Atkinson et al. [20] suggest avoiding response
counts when comparing interventions, instead recommending an approach described
by Swinton et al. [28] to estimate the proportion
of responders in the population of interest. Here, the SDIR is used
as a parameter for the distribution of true responses around the mean treatment
effect. The proportion of individuals predicted to be above or below the
selected response threshold is then estimated using the characteristics of a
normal distribution. Simulations run by Atkinson et al. [20] outline the superior accuracy of this method to reflect the
proportion of responders following an intervention. The authors recommend future
researchers use the Swinton method to estimate the proportion of response and
infer to a population of interest to reduce the influence of bias associated
with response categorizations.
Estimating the proportion of response can help remove the influence of mean
changes when comparing interventions, and may provide a more accurate
representation of how many individuals will benefit from a provided
intervention. However, this method poses practical application limitations.
Specifically, the ability to consider the implications of an intervention at the
individual level have been removed. As such, individual non-responders cannot be
identified, meaning subsequent decision making or exercise prescription
adaptations cannot be completed. Therefore, prior to selecting the analytical
method, the purpose of the analysis and utility of the outcomes must be
considered, as these two streams of research (individual analysis vs. group
based analysis) propose different theoretical approaches.
Can multiple outcome measures be simultaneously considered?
Human physiology is complex. As such, categorizing an individual as a ‘universal’
responder or non-responder based solely on the changes experienced in a single
outcome measure following an exercise intervention does not adequately reflect
the complexity of the physiological response. Current response categorizations
are specific to the selected outcome measure and the provided intervention.
However studies have shown intraindividual variance and inconsistency across
response categorizations when numerous outcomes are considered [10]
[16]
[74]. Moreover, the current method of categorizing
individuals using an ‘outcome by outcome’ basis does not allow for conceptual
outcomes, in which several outcomes are considered simultaneously (such as
physical function, frailty, or metabolic syndrome), to be utilized. Therefore,
methods which allow for the consideration of numerous outcomes simultaneously
should be investigated. This would allow for more global categorizations of
responders and non-responders.
What degree of confidence should we use when categorizing
individuals?
The confidence with which response categorizations are made is highly varied,
including categorizations made based solely on estimates of variation [40]
[46]
[71]
[73], using 50% confidence intervals [24]
[40], and 90% confidence intervals [24]
[75]. Generalizing the confidence of response
categorizations will likely be difficult, as it will be highly influenced by the
desired threshold value and the variance the researcher or clinician wants to
account for, the outcome measure being focused on, and the acceptable degree of
risk associated with an error in the categorization. Additional commentary on
the acceptable degree of confidence is necessary, but will always be contended.
Currently, we proposed future research provide a complete justification for the
selected level of confidence, and outline the associated strengths and
limitations.
What are the economic costs of ‘precision medicine’?
Throughout this review a number of recommendations and limitations associated
with assessing variance and analyzing on an individual level have been noted,
including larger sample sizes, taking multiple measures at each time point, and
conducting repeated crossover trials with extensive washout periods. Addressing
these recommendations and limitations will require significant resources and
financial support. While we believe conducting these trials is worthwhile, and
results can be directly transferred into practice, the economic costs are not
trivial. To the best of our knowledge, there is minimal evidence for, nor has a
cost-benefit analysis been conducted, suggesting implementing personalized
exercise prescriptions will provide superior, financially responsible outcomes
compared to the current ‘mean improvement’ based exercise prescription
model.
Understandably, cost-benefit analyses and longitudinal trials comparing the
effects of precision exercise prescription to mean improvement based
prescription cannot be completed until previous questions are answered and
optimal procedures are accepted. However, resource requirements and financial
needs should be taken into consideration while attempting to develop and
implement these models.
What should be done once a response categorization is made?
A major question often left unanswered is what should be done with individuals
following a response categorization. The answer will likely depend on the
setting within which the categorization is made (i.e. clinical or research). A
potential course of action for a clinical setting will be described in the
subsequent section. From a research perspective, a decision on how to progress
following response categorizations should be made prior to the initial analysis.
As individual responses are highly specific to the outcome of interest, selected
response threshold, and provided intervention [26], and reporting ‘response rates’ based on responder counts has been
shown to more accurately reflect mean group changes [20]
[21], it is hard to justify simply reporting
individual responder and non-responder categorizations as generalizable, novel,
or helping advance the field. Therefore, we propose future research should look
beyond categorizing Approaches could include investigating any one of the
questions provided throughout this section, investigating the proportion of
responders produced following an exercise intervention and comparing it to
alternate interventions using the Swinton method [20]
[28], with the goal of improving the likelihood of
participants experiencing beneficial changes, attempting to identify the root
causes of heterogeneity across the sample, or exploring if
subsequent/alternative exercise interventions may ‘rescue’ individuals
categorized as non-responders and generate beneficial adaptations within the
targeted outcome [58]
[75]. These are topics worthy of additional
attention, allow for a more thorough analysis beyond the response
categorization, and may help move us closer to precision exercise
prescriptions.
Translating Treatment Response Heterogeneity and Individual Response Research to
Practice
Translating Treatment Response Heterogeneity and Individual Response Research to
Practice
It is worth outlining how research investigating treatment response heterogeneity
and
individual responses to exercise training could improve the application of exercise.
Currently, exercise is prescribed based on broad guidelines, designed to provide a
mean improvement to various aspects of overall health. The goal of the methods
described throughout this review is to provide clinicians a method for implementing
a targeted, personalized approach when prescribing exercise, while recognizing the
complexity of human physiology implies benefits will extend beyond a single
outcome.
The utility of treatment response heterogeneity and individual response research
differ; however, they may be applied in unison when prescribing exercise.
Understanding treatment response heterogeneity and subsequently estimating the
proportion of response (as directed by Swinton et al. [28] and Atkinson et al. [20]) will allow
clinicians to prescribe exercise based on the likelihood of an individual to
experience a change greater than a selected threshold [20]. This means the clinician may be able to target a specific outcome of
importance for the patient (e.g., cardiorespiratory fitness, glycated hemoglobin
concentration, or systolic blood pressure), a desired degree of improvement, and
prescribed exercise intensity, modality, and mode based on which combination would
provide the greatest likelihood of experiencing the targeted improvement.
Subsequently, individual response research may allow for clinicians to more
accurately interpret the individual’s ability to experience the targeted benefits
–
or not – following the prescribed exercise, meaning the clinician can react
appropriately at follow-up. If the individual was determined to be a non-responder
following the provided intervention, the prescription could be adapted, or an
alternative prescription could be provided based on the estimated likelihood of
success. It is our view that these methods provide an alternative, and potentially
more beneficial, approach to the current model of prescribing exercise.
As research progresses and the factors contributing to an individual response
categorization are better understood, the accuracy of an initial prescription or
subsequent adaptations may improve. Moreover, the aforementioned questions facing
treatment response heterogeneity and individual response research must be addressed
for such a future to become reality. Nonetheless, we believe it is a worthwhile
pursuit to advance the utility of exercise.
Conclusion
Research investigating exercise treatment response heterogeneity and individual
responses will surely continue to proliferate as interest in personalized medicine
grows. While many questions remain unanswered, they represent areas for future
research and growth required to advance the field and ensure rigour. Given the
current limitations, researchers analyzing treatment response heterogeneity and
categorizing participants must consider the context of the research question, how
categorizations may be used, and make methodological decisions prior to conducting
a
trial. Subsequently, results should be interpreted within the selected method’s
capabilities, and applicable limitations clearly outlined.
