Edinburgh Research Explorer Effects and moderators of exercise on quality of life and physical function in patients with cancer

This individual patient data meta-analysis aimed to evaluate the effects of exercise on quality of life (QoL) and physical function (PF) in patients with cancer, and to identify moderator effects of demographic (age, sex, marital status, education), clinical (body mass index, cancer type, presence of metastasis), intervention-related (intervention timing, delivery mode and duration, and type of control group), and exercise-related (exercise frequency, intensity, type, time) characteristics. Relevant published and unpublished studies were identiﬁed in September 2012 via PubMed, EMBASE, PsycINFO, and CINAHL, reference checking and personal communications. Principle investigators of all 69 eligible trials were requested to share IPD from their study. IPD from 34 randomised controlled trials ( n = 4519 patients) that evaluated the effects of exercise compared to a usual care, wait-list or attention control group on QoL and PF in adult patients with cancer were retrieved and pooled. Linear mixed-effect models were used to evaluate the effects of the exercise on post-intervention outcome values ( z -score) adjusting for baseline values. Moderator effects were studies by testing interactions. Exercise signiﬁcantly improved QoL ( b = 0.15, 95%CI = 0.10;0.20) and PF ( b = 0.18, 95%CI = 0.13;0.23). The effects were not moderated by demographic, clinical or exercise characteristics. Effects on QoL ( b difference_in_effect = 0.13, 95%CI = 0.03;0.22) and PF ( b difference_in_effect = 0.10, 95%CI = 0.01;0.20) were signif- icantly larger for supervised than unsupervised interventions. In conclusion, exercise, and particularly supervised exercise, effectively improves QoL and PF in patients with cancer with different demographic and clinical characteristics during and following treatment. Although effect sizes are small, there is consistent empirical evidence to support implementation of exercise as part of cancer care.


Introduction
As a consequence of the increased number of cancer diagnoses, and concomitant mortality reductions for most types of cancer [1][2][3], many patients live with physical and psychosocial problems associated with the disease and its treatment that may compromise their quality of life (QoL). Exercise has been recommended as part of standard care for patients with cancer to help prevent and manage physical and psychosocial problems, and improve QoL [4,5].
Previous meta-analyses of randomised controlled trials (RCT) reported benefits of exercise during and following cancer treatment [6]. Benefits include improved physical fitness, function, and quality of life (QoL), and reduced fatigue, and depression [6][7][8][9]. However, average reported effect sizes on these outcomes were small to moderate.
To maximize benefits of exercise, it is important to target subgroups of patients that respond best to a particular intervention [10]. A number of RCTs showed that demographic, clinical, and personal factors, such as age, marital status, disease stage and type of treatment, moderate the effects of exercise in patients with cancer [11][12][13][14][15]. However, these single studies are generally underpowered to analyse moderators of intervention effects and conduct subsequent stratified analysis. Meta-analyses based on aggregate data are limited to using summary data, such as the mean age of the patients or the proportion of men in a study, and they are unable to investigate intervention-covariate interactions at the level of the patient [16,17].
Optimizing benefits of exercise also requires a better understanding of important intervention-related characteristics, including the timing and mode of intervention delivery, intervention duration, and exercise dimensions, in terms of frequency, intensity, type and time (FITT factors).
Meta-analyses of raw individual patient data (IPD) are suggested as the preferred method to evaluate moderators of intervention effects, since the large number of raw data points facilitates testing of interactions at the patient level, conducting subsequent stratified analyses, and standardizing analytic techniques across the included studies [18,19]. In the current IPD meta-analysis we used data collected in the Predicting OptimaL Cancer RehabIlitation and Supportive care (POLARIS) study [20]. The aims were to evaluate the effects of exercise on QoL and physical function (PF) in patients with cancer, and to identify demographic, clinical, intervention-, and exercise-related moderators of intervention effects.

Methods
The conduct and reporting of this IPD meta-analysis is based on the Preferred Reporting Items for Systematic Review and Meta-Analyses of Individual Participant Data (PRISMA-IPD) statement [21].

Identification and inclusion of studies
Detailed descriptions of the design and procedures of the POLARIS study were published previously [20]. In short, relevant published and unpublished studies (e.g. study protocol papers) were identified in September 2012 via systematic searches in four electronic databases (PubMed, EMBASE, PsycINFO, and CINAHL), reference checking of systematic reviews, meta-analyses, and personal communication with collaborators, colleagues, and other experts in the field [20]. POLARIS included RCTs that evaluated the effects of exercise interventions and/or psychosocial interventions on QoL compared to a wait-list, usual care or attention control group in adult patients with cancer. We excluded studies focusing on spiritual or existential therapy, yoga, and diet or multimodal lifestyle interventions. The study protocol was registered in PROSPERO in February 2013 (CRD42013003805) [20].
A letter of invitation to join the POLARIS consortium and share data was sent to the principal investigator (PI) of eligible RCTs. In case of no response, we sent reminders or contacted another PI. In case the study was not yet published, we maintained contact about the study completion date, to allow inclusion at a later stage during the data collection process of approximately 3 years. After PI's expressed interest in data sharing, they were requested to sign a data sharing agreement stating that they agreed with the POLARIS policy document, and were willing to share and transform anonymised data of study participants who were randomised. Data could be sent in various formats, were re-coded according to standardised protocols, and were checked for completeness, improbable values, consistency with published articles, and missing items. Subsequently, datasets were imported into the POLARIS database where they were harmonized [20].

Data extraction and quality assessment
Two independent researchers (LB and MS) extracted study characteristics and rated the quality of included studies from published papers, using the 'risk-of-bias' assessment tool of the Cochrane Collaboration. The quality of following aspects was graded as high ('+'), low ('À') or unclear ('?') quality: random sequence generation (high quality if random component was used), allocation concealment (high quality if central, computerized allocation or sequentially numbered sealed envelopes were used), incomplete outcome (high quality if intention-to-treat analyses were performed and missing outcome data were <10% or adequate imputation techniques were used), and incomplete reporting (high quality if QoL or PF was reported such that data could be entered in an aggregate data meta-analysis). We also included ratings of adherence (high quality if P80% of patients had high attendance, defined as P80% of sessions attended [22,23]) and contamination (high quality if no or limited exercise was present in the control group, i.e. moderate to vigorous exercise was present in <25% of patients or increased less than 60 min [24]). Items related to blinding were omitted because blinding of patients and personnel is difficult in the case of exercise interventions, and QoL and PF were assessed using patient-reported outcomes. Quality assessments of both reviewers were compared and disagreements in the scores were resolved by discussion.

Representativeness of included studies
To examine whether the included RCTs were a representative sample of all eligible RCTs, we compared pooled effect sizes of RCTs included versus those not included. Effect sizes per RCT were calculated by subtracting the published average post-intervention value of QoL or PF of the control group from that of the intervention group, and dividing the result by the pooled standard deviation. We corrected effect sizes for small samples as suggested by Hedges and Olkin. Effect sizes (Hedges' g) were pooled with a random effects model and differences in effects between studies providing data and those that did not were examined using Comprehensive Meta-analysis software (version 2.2.064).
We evaluated publication bias for all eligible studies and for studies providing data by inspecting the funnel plot and by the Duval and Tweedie's trim and fill procedure [25,26]. The procedure provides estimates of the number of missing studies and the effect size after the publication bias has been taken into account. The Egger's test was used to test whether the bias captured by the funnel plot was significant.

Outcome variables
QoL and PF were assessed with patient reported outcomes (PRO, Table 1). In the present paper, we used baseline (pre-intervention) and post-intervention values. To allow pooling of the different PROs, we recoded the individual scores into z-scores by subtracting the individual score from the mean score at baseline, and dividing the result by the mean standard deviation at baseline. Subsequently, the pooled z-scores were used for further analyses. If studies used both a cancer-specific and a generic QoL PRO, data from the cancer-specific PRO were used.

Possible moderators
Potential demographic and clinical moderators were identified from single studies that reported on the moderating effects with some inconsistent findings [11][12][13][14]27].
Potential demographic moderators included baseline age, sex, marital status, and education level. Marital status was dichotomised into single versus married or living with partner. As a consequence of different coding schemes of the original RCTs, education level was dichotomised into low-medium (elementary, primary, or secondary school, lower or secondary vocational education) or high (higher vocational, college, or university education). Potential clinical moderators included body mass index (BMI), type of cancer, the presence of distant metastases, and type of treatment. BMI was categorised into underweight (<18.5 kg/m 2 ), normal weight (18.5-<25 kg/m 2 ), overweight (25-<30 kg/m 2 ) and obese (P30 kg/m 2 ) according to the World Health Organization. The type of cancer was categorised into breast, male genitourinary, gastrointestinal, haematological, gynaecological, respiratory tract, and other types. Treatment with surgery, chemotherapy, radiotherapy, hormone therapy or stem cell transplantation were each dichotomised into previous or current treatment versus no such treatment. As the majority of men diagnosed with prostate cancer received androgen deprivation therapy, we were unable to study the moderating effects of hormone therapy in prostate cancer.
Timing of intervention delivery in relation to primary cancer treatment was categorised into pre-treatment, during treatment, post-treatment and end-of-life, according to the Physical Activity and Cancer Control (PACC) framework [28]. Because interventions pre-treatment and during end-of-life were not available, we tested differences in intervention effects between those delivered during treatment versus post-treatment. As hormone therapy for breast cancer may continue for five years post-treatment, we considered Table 1 Descriptives of studies evaluating the effects of exercise on quality of life and physical function included in the database (n = 34), in alphabetical order of first author. (continued on next page)  women on hormone therapy who completed other primary cancer treatments as being post-treatment. Men receiving androgen deprivation therapy for prostate cancer were considered as being during treatment. Delivery mode of intervention was dichotomized into supervised (in case (part of) the weekly exercise sessions were conducted under supervision) versus unsupervised (in case exercise sessions were performed unsupervised from or at home). Intervention duration was categorised based on tertiles (612 weeks; >12-24 weeks; >24 weeks). Exercise frequency was dichotomised based on the median, into 62 versus >2 supervised sessions per week for supervised exercise and into <5 versus P5 sessions per week for unsupervised exercise. Exercise intensity was categorised from low to high intensity using the definitions of the American College of Sports Medicine [29]. Exercise type was categorised into aerobic, resistance, combined aerobic and resistance and combined resistance and impact loading (e.g. skipping, jumping) exercise. Exercise time per session was categorised into 630 min, >30-60 min and >60 min.

Statistical analysis
We conducted one-step IPD meta-analyses to study the effects and moderators of exercise on QoL and PF. The effects were evaluated by regressing the intervention on the post-intervention value (z-score) of the outcome adjusted for the baseline value (z-score) using linear mixed model analyses with a two-level structure (1: patient; 2: study) to take into account the clustering of patients within studies by using a random intercept on study level. Moderators of exercise effects were examined by adding the moderator and its interaction term with the intervention into the regression model, for each moderator separately. To reduce ecological bias for patient-level interactions, we separated within-trial interaction from between-trial interaction by centring the individual value of the covariate around the mean study value of that covariate [19]. If interaction terms were significant (p < 0.05), stratified analyses were performed. In case a RCT had three study arms with different study-level moderators across study arms, interaction testing for a study-level moderator was not possible. Therefore, in those situations, we tested differences between subgroups using dummy variables. Regression coefficients and 95% confidence intervals (CI) were reported, which represent the between group difference in z-scores of QoL and PF, and correspond to a Cohen's d effect size. Effects of 0.2 were considered small, 0.50 as moderate and at or above 0.8 as large.
Since the majority of patients were women with breast cancer, we performed a sensitivity analysis to check robustness of findings in the subgroup of patients that were not women with breast cancer, despite non-significant overall interaction effects for women with breast cancer vs other (b = 0.09, 95%CI = À0.12; 0.29 for QoL; b = À0.06, 95%CI = À0.27;0.14 for PF). Statistical analyses were performed using SPSS 22.0 and R Studio.

Characteristics of studies and patients
Of the 136 RCTs that met the inclusion criteria (Fig. 1), 66 evaluated the effects of exercise and three [30][31][32] evaluated the effects of a combined exercise and psychosocial intervention and also included a third arm with exercise only. Principal investigators of 34 of these 69 RCTs (response 49%) shared IPD. In total, 27 RCTs reported adequate random sequence generation, 26 studies reported adequate allocation concealment, 26 RCTs had adequate completeness of outcome data, and 26 RCTs had complete outcome reporting (Table 1). Intervention adherence was reported in 26 RCTs, and was of high quality in 13 RCTs, and 7 of the 13 RCTs that  provided information on contamination met the criteria for high quality.
The sample included 4519 patients with cancer, of whom 2514 were randomized to the intervention group and 2005 to the control group. The mean age was 54.6 (SD 11.3) years, 78% were women, 70% were diagnosed with breast cancer, 2% had metastatic disease, 51% exercised following cancer treatment, and 65% received supervised exercise ( Table 2).

Representativeness and publication bias
Published summary data for QoL were available for 36 out of 69 RCTs, of which five [27,[33][34][35][36] included two exercise arms. Consequently, 41 exercise arms were included in the analyses of representativeness. For PF, summary data were published for 30 RCTs, with two [27,37] evaluating two exercise arms, resulting in 32 exercise arms. We found no significant differences in effects on QoL (p = 0.25) and PF (p = 0.25) between RCTs of which IPD were shared and those of which were not ( Table 3). The trim and fill procedures showed significant publication bias for all eligible RCTs reporting on QoL, but not between RCTs included and those not included (Table 3).

Effects and moderators of exercise on QoL and PF
Exercise effects on QoL (b = 0.15, 95%CI = 0.10;0.20) and PF (b = 0.18, 95%CI = 0.13;0.23, Table 4, Fig. 2) were significant. Patients' demographic and clinical characteristics, intervention timing and duration, and exercise FITT factors did not significantly moderate the effects on QoL or PF (Table 4) Sensitivity analyses among patients other than women with breast cancer (n = 1360, originating from 17 RCTs) showed slight differences in regression coefficients with larger confidence intervals, but the conclusions on moderator effects were similar.

Discussion
Based on IPD meta-analyses of 34 RCTs including data from 4519 individual patients with cancer, we found that exercise significantly improved their QoL and PF. The IPD meta-analytical approach of the present paper enabled the testing of potential moderators in a large sample. The exercise effects did not differ significantly across subgroups of age, sex, education level, marital status, BMI, cancer type, metastatic stage or treatment. Further, exercise was equally effective during and following cancer treatment. These findings support and strengthen the evidence base   for current exercise recommendations that all patients with cancer should be physically active during and following cancer treatment [4]. However, the effects were stronger for supervised exercise. We found no significant moderating effects of intervention timing, duration, and exercise FITT factors. The exercise effects were significant, but small in general, and comparable across the different subgroups. The lack of demographic and clinical moderators suggests that targeting exercise, based on demographic and clinical characteristics may not be useful for improving QoL and PF.
The moderating effects of sex, age, education, marital status, BMI and cancer type have been explored in previous single studies reporting inconsistent findings [11][12][13][14]27]. It has been hypothesized that patients without a partner have less social support at home [38,39] and may therefore either benefit more from the support associated with supervised or guided exercise [13,14], or may be less likely to adhere to the exercise intervention [23]. We analysed the potential moderating effect of being married/having a partner, although this does not necessarily reflect partner support, and found no moderator effect on QoL and PF.
Additionally, we found no moderator effect of BMI. However, due to the higher likelihood of sarcopenic obesity (i.e. increased fat mass in combination with reduced muscle mass) caused by cancer and its treatment [40], BMI may not adequately reflect adiposity in patients with cancer. Additional studies are needed to investigate the moderator effects of muscle and fat mass.
We found no significant differences in effects on QoL and PF across cancer types or between patients with metastatic and non-metastatic disease. However, sample sizes of some subgroups were small, and due to different coding schemes or lack of information on disease stage we were limited to studying differences in intervention effects between patients with metastatic and  non-metastatic disease, and were unable to further disentangle differences in effects between patients with disease stages I, II and III. Furthermore, the majority of studies evaluating the effects of exercise have been conducted in patients with breast cancer, and prostate cancer who were treated with curative intent [4,7]. Therefore exercise effects on QoL and PF remain unclear in under-studied cancer populations, such as head and neck, lung, and gynaecological cancers, and in patients with metastatic disease, and they may differ from those with breast and prostate cancer due to differences in treatment trajectories. We were unable to confirm previous findings that radiotherapy [12] or chemotherapy [13] moderate exercise effects, which may be related to the heterogeneous study population. As treatment types are related to cancer types, the moderator effects of treatment should perhaps be investigated separately within each cancer type. Intervention goals are likely to differ across phases of the cancer continuum. Exercise during cancer treatment typically seeks to influence treatment effectiveness and coping by managing side effects, maintaining physical fitness, and preventing muscle loss, fat gain, fatigue, and deterioration in QoL [28]. Exercise posttreatment typically aims to speed recovery, improve physical fitness and QoL, reduce fatigue, distress and the risk of developing chronic diseases or secondary cancers [28]. Nevertheless, the exercise effects on QoL and PF were similar, and clearly demonstrate significant benefits both during and post cancer treatment, which is consistent with previous meta-analyses based on aggregate data [6,8,9].
Effects of supervised exercise were twice as large as those of unsupervised exercise, which is consistent with a previous systematic review [41]. The larger effects of supervised exercise may be explained by the attention of the physiotherapist or exercise physiologist delivering the intervention, access to better equipment, more challenging exercise prescriptions, or by better adherence to the prescribed exercise protocol. Reporting adherence and identifying determinants of adherence to unsupervised interventions is important to identify patients who do not need supervision.
The lack of significant differences in exercise effects across different FITT factors might have resulted from little variation in these factors across studies, or the limited power since FITT factors are moderators at the intervention level instead of the patient level. Previous head-to-head comparisons of exercise FITT factors indicated a dose response effect of aerobic exercise on PF during cancer treatment in patients with breast cancer [42] and larger effects of high intensity compared to moderate intensity exercise post treatment in a population with mixed cancer types [27]. More RCTs that directly compare exercise FITT factors are warranted to define optimal exercise prescriptions. Also, specific intervention components, including goal-setting, social support and exercise instructions and monitoring, may differ across interventions, and explain differences in effects.
The effects on QoL and PF were significant, but smaller than expected. There may be several explanations for the smaller effects. First, exercise interventions generally aim to improve exercise behaviour or health-related physical fitness, and probably not all dimensions of QoL (i.e. physical, emotional and social wellbeing) [43] were affected to the same extent. Second, QoL is susceptible to response shift [44,45], i.e., a change in the meaning of one's self-evaluation of QoL over time as a result of changes in internal standards, values and the conceptualization of QoL [46]. Third, results may have been contaminated by the adoption of exercise by patients in the control group. The limited information on contamination hampered us to evaluate its influence on the effects. Fourth, our analyses were based on patients participating in RCTs. Median (interquartile range) participation rates in exercise trails were found to be 63% (33-80) of eligible patients [47]. Patients who decline participation may be less motivated for exercise and have lower exercise levels, thus we may not reach patients who may benefit the most. However, studies comparing exercise of participants and non-participants found no differences [23,48,49]. Nevertheless, demographics may differ between participants and nonparticipants, with the latter more likely to be older [48] and to have lower education levels [23,49]. Therefore, results may not be fully generalizable to all patients with cancer. Future IPD metaanalyses should also study the moderator effects of baseline QoL, PF and fitness [50], and specific symptoms as fatigue and distress [12] and the moderator effects on other physical, psychosocial and clinical outcomes, as they may differ [13,14].
Study strengths are the large number of included RCTs from multiple countries, the consequent large sample size, and the uniform analytical procedures across all studies. Limitations are the following: first, there was considerable publication bias in studies that met our inclusion criteria, overestimating the intervention effects, particularly for studies reporting on QoL. However, no significant differences in effect sizes were found between studies providing data and those that did not, indicating that the 34 RCTs included in the analyses were a representative sample of the published literature. Second, not all RCTs met all quality criteria. In particular, information on exercise adherence and contamination was limited, hampering the ability to check whether adherence was similar across moderator subgroups. However, a previous review on determinants of exercise adherence in patients with cancer concluded that the majority of studies showed no significant association of demographic and clinical factors with adherence [51]. Finally, we focused on short term intervention effects as very few studies have examined maintenance of intervention effects into the long term.
In conclusion, exercise, and particularly those with a supervised component, effectively improves QoL and PF across subgroups of patients with cancer with different demographic and clinical characteristics, both during and following treatment. Although effect sizes were small, our study provides additional evidence to support the implementation of exercise as part of standard care to improve QoL and PF. Current knowledge on the exercise effects on QoL and PF is primarily based on studies in patients with non-metastasised breast or prostate cancer. Future studies should therefore shift the focus to understanding the exercise effects in understudied and advanced cancer populations; on clinical outcomes including specific symptoms, cancer treatment completion, and survival; and on how to optimize exercise participation, adherence, and prescriptions.