Digital health behaviour change interventions targeting physical activity and diet in cancer survivors: a systematic review and meta-analysis

Purpose The number of cancer survivors has risen substantially due to improvements in early diagnosis and treatment. Health behaviours such as physical activity (PA) and diet can reduce recurrence and mortality, and alleviate negative consequences of cancer and treatments. Digital behaviour change interventions (DBCIs) have the potential to reach large numbers of cancer survivors. Methods We conducted a systematic review and meta-analyses of relevant studies identified by a search of Medline, EMBASE, PubMed and CINAHL. Studies which assessed a DBCI with measures of PA, diet and/or sedentary behaviour were included. Results Fifteen studies were identified. Random effects meta-analyses showed significant improvements in moderate-vigorous PA (seven studies; mean difference (MD) = 41 min per week; 95% CI 12, 71) and body mass index (BMI)/weight (standardised mean difference (SMD) = −0.23; 95% CI −0.41, −0.05). There was a trend towards significance for reduced fatigue and no significant change in cancer-specific measures of quality of life (QoL). Narrative synthesis revealed mixed evidence for effects on diet, generic QoL measures and self-efficacy and no evidence of an effect on mental health. Two studies suggested improved sleep quality. Conclusions DBCIs may improve PA and BMI among cancer survivors, and there is mixed evidence for diet. The number of included studies is small, and risk of bias and heterogeneity was high. Future research should address these limitations with large, high-quality RCTs, with objective measures of PA and sedentary time. Implications for cancer survivors Digital technologies offer a promising approach to encourage health behaviour change among cancer survivors. Electronic supplementary material The online version of this article (doi:10.1007/s11764-017-0632-1) contains supplementary material, which is available to authorized users.


Introduction
Over 14 million people are diagnosed with cancer worldwide each year, and this is expected to rise to 22 million over the next two decades [1]. Improvements in early diagnosis and treatments mean that cancer survival is increasing. In 2012, globally there were 32 million people living beyond 5 years of diagnosis [2] and in the UK, half of people diagnosed with cancer will now survive for more than 10 years [3].
However, long-term negative consequences of cancer and treatment related side-effects are common and often debilitating. Prevalence of fatigue following a cancer diagnosis ranges from 59 to 100% depending on cancer type [4], and pain [5], sleep problems [6], physical side effects (e.g. lymphoedema) [7], weight gain [8], anxiety and depression [9,10], fear of Electronic supplementary material The online version of this article (doi:10.1007/s11764-017-0632-1) contains supplementary material, which is available to authorized users. cancer recurrence [11] and impaired quality of life (QoL) [12] are all commonly reported. Macmillan Cancer Support, a UK cancer charity, estimates that more than 70% of cancer survivors in the UK (~1.8 million people) are also living with at least one other long-term comorbidity [13]. The most common comorbid conditions are hypertension, obesity, mental health problems and chronic heart disease [13]. The shared risk factors between cancer, obesity and cardiovascular disease (CVD) partially explain comorbidities [14]. However, there is also emerging evidence to suggest that cancer treatment can leave survivors at greater risk for developing these conditions (e.g. due to cardiovascular toxicity of cancer therapy [15]). The greater number and severity of comorbidities is linked to greater risk of death and cancer recurrence among cancer survivors [16]. There is now strong impetus to develop interventions that improve long-term outcomes for cancer survivors.
Health behaviours such as physical activity (PA), sedentary behaviour and diet are important in risk reduction and selfmanagement of cancer, CVD and obesity. For example, a meta-analysis of 22 prospective cohort studies of 123,574 breast cancer survivors found that greater post-diagnosis PA participation reduced all-cause (hazard ratio [HR] = 0.52, 95% CI 0.43, 0.64) and breast cancer-specific mortality (HR = 0.59, 95% CI 0.45, 0.78), and breast cancer recurrence (HR = 0.79, 95% CI 0.63, 0.98) [17]. A meta-analysis of prospective studies of colorectal cancer survivors reported similar conclusions and showed that post-diagnosis PA reduced both all-cause (summary relative risk [RR] = 0.58; 95% CI 0.48, 0.70; n = 6 studies) and colorectal cancerspecific mortality (summary RR = 0.61; 95% CI 0.40, 0.92; n = 5 studies) [18]. The authors estimated that each 10 metabolic equivalent task (MET)-hour per week increase in postdiagnosis PA was associated with 24% (95% CI 11-36%) and 28% (95% CI 20-35%) decreased total mortality risk for breast and colorectal cancer survivors, respectively [18]. Mishra et al.'s meta-analysis of non-digital interventions focused on the effect of PA on health-related QoL (HRQoL) outcomes in cancer survivors (various types) and found that greater PA participation significantly improved overall HRQoL at up to 12 weeks of follow-up (11 studies, n = 826; standardised mean difference [SMD] = 0.48, 95% CI 0.16, 0.81) [19]. Individual meta-analyses of other cancer-relevant outcomes identified in this same Cochrane review also found that PA interventions improved emotional well-being/mental health and social functioning, and reduced anxiety, fatigue, pain and sleep disturbance [19]. Although limited to crosssectional and prospective studies, there is some evidence that higher levels of sedentary time are associated with lower physical and role functioning domains of QoL, and greater reporting of comorbidities, disability and fatigue [20][21][22]. As a result of the growing evidence of the benefits of PA following a cancer diagnosis, cancer survivors are encouraged to avoid inactivity as far as possible and to meet the same PA guidelines as the rest of the adult population of at least 150 min of moderate-vigorous PA (MVPA) and two instances of strength/resistance-based exercises per week [23][24][25].
Diet may also influence outcomes following a cancer diagnosis. A meta-analysis of three studies (n = 9966) suggested that a low-fat diet post diagnosis can reduce breast cancer recurrence by 23% and all-cause mortality by 17% [26]. Another meta-analysis of four prospective cohort studies (n = 3675) found that high saturated fat intake increased breast cancer-specific mortality [27]. A meta-analysis of 56 observational studies in 1,784,404 cancer survivors (various types) showed greater adherence to a Mediterranean-style diet (largely based on vegetables, fruits, nuts, beans, cereal grains, olive oil and fish) was associated with lower all-cause cancer mortality for colorectal, breast, gastric, prostate, liver, head and neck, pancreatic and respiratory cancers [28]. Colorectal cancer survivors consuming a Western diet (high intake of meat, fat, refined grains and desserts) showed greater risk of recurrence and overall mortality compared to those with a prudent diet (high intake of fruits and vegetables, poultry and fish) in a prospective study of 1009 participants [29]. Similar findings have been shown in other prospective cohort studies of breast cancer survivors [30,31]. Breast cancer survivors with better overall diet quality also reported lower levels of fatigue, independently of PA participation, at 41 months post diagnosis in a prospective cohort study [30].
Longitudinal studies have shown that obesity increases the risk of cancer recurrence among prostate [32], colorectal [33] and breast [34] cancer patients.
Despite the wealth of evidence, cancer survivors' engagement with health behaviours and adherence to lifestyle guidelines for cancer survivors are remarkably poor [35,36]. The English Longitudinal Study of Ageing demonstrated that the proportion of cancer survivors who engaged in self-reported MVPA at least once per week fell from 13% before their cancer diagnosis to 9% after their cancer diagnosis (compared to a fall of 16 to 15% in the group who did not receive a cancer diagnosis between data collection waves) [37]. Wang et al. found that cancer survivors were less likely to engage in self-reported PA (adjusted odds ratio = 0.79, 95% CI 0.67, 0.93) compared to those without a cancer diagnosis [38]. Furthermore, few cancer survivors meet the minimum recommended guidelines of 150 min of MVPA per week. A study of over 9000 survivors of six types of cancer found that adherence to PA recommendations varied from 30% (uterine cancer) to 47% (skin melanoma); however, this study did use selfreported PA measures [36]. While this study reported that 35% of breast cancer survivors were meeting guidelines, another study which used accelerometers to measure PA objectively found that this can be as low as 16% and those with highest levels of comorbidities were the least active [39]. Consequently, there is a need for evidence-based interventions that are easy to access, low-cost and which therefore have the feasibility to be rolled out to reach a large number of cancer survivors.
A move towards digital health behaviour change interventions (DBCIs) has been driven by widespread and rising use of the Internet, smartphones and mobile technology [40,41]. The most recent Ofcom Communications Market report for the UK has shown that the proportion of adults going online using a mobile phone has risen from 20% in 2009 to 66% in 2016 and 71% of adults own a smartphone [41]. DBCIs use technologies such as text messaging, email, mobile applications (apps), video-conferencing (e.g. Skype), social media, websites and online patient portals increasing access to information, connecting patients with health services and as an approach to remote delivery of behaviour change interventions. DBCIs have been used in the promotion of medication adherence [42], management of long-term conditions [43][44][45], promoting smoking cessation [46] and promoting PA participation and dietary behaviours [47][48][49][50]. A recent systematic review of 224 studies reported that Internet and mobile interventions improved diet, PA, obesity, tobacco and alcohol use up to 1 year [51]. Among cancer survivors, a recent systematic review of 27 non-face-to-face intervention studies found telephone interventions as an effective approach to delivering PA and dietary interventions [52]. However, newer digital technologies should now be evaluated in this population as only three of the studies in that systematic review included used web-based methods to deliver the intervention [53][54][55]. No systematic review or meta-analysis has assessed the efficacy of DBCI interventions targeting PA, diet and/or sedentary behaviour among cancer survivors specifically.
Therefore, the primary objective of this study was to perform a systematic review and meta-analysis of health behaviour interventions using digital technologies in cancer survivors in order to assess their efficacy in promoting PA, reducing sedentary behaviour or improving dietary quality. Secondary aims were to explore any effects of DBCIs on BMI/weight, other cancer-relevant outcomes and the theoretical underpinning of included studies.

Search strategy
A systematic literature search was conducted from database inception to November 8, 2016, of the following databases: Medline, EMBASE, PsycINFO and CINAHL. Full details of the search strategy/terms used can be found in Online Resource 1. Broadly, the search strategy combined synonyms for PA, diet and/or sedentary behaviour; with types of DBCIs (e.g. website, mobile app, text messaging); and with words for cancer survivor(ship).
Limits included peer-reviewed, English language articles in human subjects. Forward and backward citing of included studies and hand-searching of relevant journals were also conducted to identify relevant articles. The protocol was registered in the PROSPERO database (CRD42016026956). After piloting of the search strategy, no new or relevant articles from other databases specified in the protocol (Cochrane Library, Web of Science, ACM Digital Library, or IEEE Xplore) were identified so these databases were excluded for the final search. As specified in the protocol, the ProQuest database (grey literature) was searched; however, this resulted in >60,000 search results. Results were sorted by relevance, and the first 200 titles were reviewed. No additional, relevant papers which met criteria were identified throughout this process so grey literature was not included.

Study selection
Studies were selected in line with the search strategy shown in Fig. 1. Eligible studies included DBCIs delivered remotely and targeting at least one of the following health behaviours: PA, diet and/or sedentary behaviour in adults (≥18 years) who had a cancer diagnosis of any type. There were no restrictions on quantitative study designs, so both randomised and nonrandomised controlled trials and one-arm pre-post comparison studies could be included. However, qualitative studies and protocols were excluded. Studies must have measured at least one of the target health behaviours (PA, diet and/or sedentary behaviour) at baseline and follow-up, but there were no limits on length of follow-up for inclusion.

Data extraction and quality assessment
Two authors (AR and AF) independently reviewed 109 fulltext articles screened for eligibility and extracted the data for included studies including author, country of study, study design, sample size, retention rate, population studied, age of participants, study duration, intervention type (i.e. type of DBCI), description of intervention content (including incorporated behaviour change techniques (BCTs)), approaches to measurement of engagement/adherence to the intervention, control group treatment and outcomes measured. Any discrepancies were resolved through discussion. Michie et al.'s BCT Taxonomy (v1) [56,57] was used to code BCTs based on information provided in the included studies (and any supplementary material). The Cochrane Collaboration's tool for assessing risk of bias was used to evaluate methodological quality of included studies [58], and Michie and Prestwich's Theory Coding Scheme was used to evaluate the theoretical basis of the included studies [59].

Statistical methods
Where possible, findings from both RCTs and one-arm prepost studies were synthesised in random effects meta-analyses using Stata. Effect sizes for the intervention were calculated using the difference in final values between experimental and control groups in RCTs and the change in scores before and after the intervention in pre-post studies. It is not recommended to combine studies using a mixture of final values and change scores when using standardised mean differences (SMDs) across studies using different measurement units/ tools to assess an outcome [58]. Therefore, outcomes using the same measurement unit were chosen wherever possible so non-standardised mean differences could be used and RCTs and pre-post studies could be combined in the meta-analyses [58]. Where this was not possible (i.e. fatigue outcomes), SMDs and their associated 95% CIs were calculated, and meta-analyses were conducted for the RCTs only (where the effect size reflects difference in final values between groups). As BMI is largely influenced by weight, the variability in reliability was judged to be similar for weight and BMI. Therefore, SMDs were used to pool the effect of BMI and weight across both the RCTs and pre-post studies reporting these outcomes. For PA outcomes, MVPA was chosen as the outcome measure of interest due to the American College of Sports Medicine's recommendation that cancer survivors follow the PA guidelines for the general population of at least 150 min of at least moderate intensity PA per week [24]. Studies reporting MVPA duration in minutes were pooled in the meta-analysis, so studies with differences in final values and change scores could be used using mean differences. Studies that did not report moderate and vigorous PA separately or MVPA combined in minutes were not included in the meta-analysis of PA outcomes. For the studies that reported minutes of moderate and vigorous PA separately, a new combined MVPA variable was calculated. To combine the means for moderate and for vigorous PA, the following formula was used: To combine the standard deviations for moderate and vigorous PA, the following formula was used: Publication bias was explored using funnel plots prepared in Stata. Due to the small number of included studies for each outcome, tests for funnel plot asymmetry (e.g. Egger's

Articles included in review n=15
Records identified through database searching n=7280 Articles screened after removal of duplicates n=6301 Articles remaining for scrutiny of abstracts n=733 Full-text articles screened for eligibility n=109

Removal of duplicates n=979
Articles excluded based on title n=5568 Additional records identified through forward-citing and hand-searching relevant journals n=15 Articles excluded based on abstract n=639 Full-text articles excluded: n=94

Primary outcomes Physical activity and sedentary time
All 15 included studies measured the impact of DBCIs on PA [53][54][55][61][62][63][64][65][66][67][68][69][70][71][72]. All used self-reported PA as outcomes: five used the Godin Leisure-Time Exercise Questionnaire (GLTEQ) [53,61,63,69,72], two the International Physical Activity Questionnaire (IPAQ) [66,70], one a 7-day PA recall [54] and one the Short Questionnaire to Assess Health Enhancing Physical Activity (SQUASH) [71]; two identified the number of days in the last seven that the participant engaged in moderate and/or vigorous PA [55,62]; three studies reported percentage of participants meeting PA guidelines (150 min of MVPA per week) [54,63,67] and two reported stages of change for PA [65,67]. Short et al. [72] also reported a resistance training score. Hoffman et al. [64] reported the number of minutes walked, steps walked and number of balance exercises completed. McCarroll et al. [68] reported number of minutes of PA completed and the number of calories expended as logged via the participant using the mobile app used for their intervention.
Of the other eight studies which could not be included in the meta-analysis, four reported a significant effect, [55,[65][66][67], two did not report significant findings [62,68] and two did not conduct significance testing due to small sample sizes [64,70]. No studies reported effects on sedentary time.

Diet
Five studies measured the impact of DBCIs on dietary intake [61,67,68,70,71]. Due to the variation in approaches to assessment and measurement of dietary outcomes, a metaanalysis was not considered appropriate. Three studies [61,67,71] were RCTs and two were pre-post studies [68,70]. Only two of the studies reported a significant effect on dietary outcomes [67,71]; however, this no longer remained significant after correcting for multiple testing in Kanera et al.'s study [71]. Quintiliani et al. [70] did not conduct significance testing, due to the very small sample (n = 10).

Other cancer-relevant outcomes
Fatigue Seven studies measured the impact of DBCIs on fatigue [54,61,63,64,67,69,70]. Of these, three used the Brief Fatigue Inventory (BFI) [61,64,67], two used the Functional Assessment of Chronic Illness Therapy-Fatigue (FACIT-F) [63,69], one used the Profile of Mood States-Fatigue (POMS-Fatigue) [54] scale and one used a 0-10 scale [70]. SMDs were required to pool effects across studies due to the variation in measurement tools. Therefore, meta-analysis was only conducted on the three RCTs where appropriate data could be extracted [54,61,63], using data from 406 participants (see Fig. 3, Online Resource 2). DBCIs resulted in a decrease in fatigue, but this was not significant (SMD = −0.23; 95% CI −0.51, 0.05; p = 0.103; I 2 = 78%). Once again, very high levels of heterogeneity were displayed for fatigue. A funnel plot revealed no evidence of publication bias for fatigue outcomes. Of the remaining four studies, only one reported a significant reduction in fatigue [67]. Two of these studies did not report significance testing [64,70] due to very small sample sizes (7 and 10, respectively).

Cancer-specific QoL
Five studies assessed cancer-specific measures of QoL [53,63,[67][68][69]. Four studies used the FACT-G [53,63,68,69] and one used the Quality of Life Questionnaire-Core 30 (QLQ-C30) [67]. The four studies using the FACT-G (two RCTs [53,63] and two pre-post studies [68,69]) were pooled using data from 198 participants (152 participants in RCTs; 46 from pre-post studies) (see Fig. 4 [69] and a seven-item non-validated measure [65]). Two studies found no significant changes in any QoL domains [63,69]. One study found a significant improvement in role functioning-emotional and mental health for those during treatment and a significant improvement in social functioning for those who had finished treatment [66].  [65].

Mental health
Three studies measured the impact of DBCIs on depression [61,67,69], none of which reported any significant impact. Two studies assessed the impact on anxiety [67,69], neither of which reported a significant effect. Rabin et al. [54] also measured Profile of Mood States (anger, depression, tension/anxiety, vigour, confusion) and did not find a significant change in scores between groups.

Sleep disturbance
Two studies measured the impact of DBCIs on sleep disturbance [61,69]. Both studies reported a significant improvement in sleep quality; Bantum et al. [61] showed a significant reduction in insomnia, and Puszkiewicz et al. [69] showed a significant improvement in sleep quality.

Theoretical underpinning
Twelve studies reported some level of theoretical basis to their intervention design [53-55, 62, 64-68, 70-72]. Of those that did mention a theoretical influence, Social Cognitive Theory (SCT) was most frequently reported [53,55,66,68,71,72], and sometimes used in combination with other theories (i.e. Transtheoretical Model (TTM) [54] or the Theory of Planned Behaviour (TPB) [66]). Other theories included the Theory of Reasoned Action [62], Theory of Symptom Self-Management [64], Theory of Goal Setting [65], the TTM alone [67] and the Social Contextual Model [70]. The description of the theoretical underpinning of the DBCIs varied across studies, but was relatively poor. Only seven studies explicitly reported how theory/predictors were used to select/develop intervention techniques [53-55, 66, 67, 71, 72], and only six studies used theory/predictors to tailor intervention techniques to participants [54,64,66,67,71,72]. Six studies measured theoryrelevant constructs and reported outcomes pre and post intervention [53,62,[66][67][68]70]. Of these six studies, three reported no significant change in measures related to self-efficacy [62,66,70]. Lee et al. [67] reported that the 'stage of change' and self-efficacy for PA and fruit and vegetable consumption was significantly higher in the intervention group, and McCarroll et al. [68] reported a significant increase in selfefficacy between pre and post intervention. Valle et al. was the only study which conducted mediational analysis of theoretical constructs/predictors (presented in an associated paper [73]). However, this study showed that the intervention group reported lower self-efficacy for PA and social support from friends on social networking websites compared to the control group. Changes in social support from friends on social networking sites were positively related to changes in MVPA; however, it was those in the control group who reported increased social support from friends on social networking sites compared to the intervention group [73].  N.B. p-values shown in Figure 2 represent significance for testing of heterogeneity

Risk of bias in included studies
See Fig. 3 for an assessment of the risk of bias in the included studies. Study quality was deemed to be low for the majority of included studies. For the RCTs, there was adequate randomisation sequence generation in the majority of the studies; however, allocation concealment was much less clear. The lack of a control group in the seven pre-post studies highly increased the risk of bias, reducing the confidence which can be placed on the observed effects. All of the studies were judged to be high risk for other sources of bias, for instance all 15 studies used self-reported PA, as opposed to objective measures (e.g. accelerometry). Some of the RCTs did not report (or it was unclear) whether baseline outcome measures and/or participant characteristics were similar between groups [53,54,61,63,71,72]. Furthermore, the representativeness of the sample was fairly narrow in that the majority of the study samples which consisted of predominantly female, middleaged, English-speaking, breast cancer survivors of high educational/income level, which likely limits the generalizability of the findings across the wider population of cancer survivors. Most of the studies had small sample sizes (generally ranging between 20 and 100), and only three studies had larger samples [61,71,72] (303, 462 and 492, respectively). Despite the largest sample size at baseline in Short et al.'s study [72], retention was extremely low (32% at 3 months and 11% at 6 months of follow-up).

Discussion
The current meta-analysis found that DBCIs resulted in an increase in MVPA participation of approximately 40 min per week. While meta-analysis was not possible for dietary outcomes, there was mixed evidence for an effect on dietary intake. No studies assessed sedentary behaviour. Metaanalyses also revealed a significant reduction in BMI, a reduction in fatigue which did not reach statistical significance, and no change was seen in cancer-specific measures of QoL. For other secondary outcomes where meta-analysis was not possible, there was mixed evidence for the effect on domains of generic QoL measures and theoretical constructs (e.g. selfefficacy). There is no evidence for an improvement in anxiety or depression, and while only two studies assessed sleep disturbance, both reported a significant improvement. To our knowledge, this is the first meta-analysis to assess the current evidence with regards to DBCIs targeting PA and/ or diet among cancer survivors. An increase in approximately 40 min of MVPA per week is important given that this represents a substantial proportion (27%) of cancer survivors' recommended weekly MVPA participation [23][24][25]. While there is limited evidence on the dose-response relationship between MVPA and mortality and recurrence outcomes, Schmid et al. estimated that each 10 MET-hour per week increase in postdiagnosis PA (approximately equivalent to the 150-min MVPA/week guideline) was associated with 24% (95% CI 11-36%) and 28% (95% CI 20-35%) decreased total mortality risk for breast and colorectal cancer survivors, respectively [18]. Schmid et al. also reported that breast or colorectal cancer survivors who increased their PA by any amount between pre and post diagnosis showed a decreased total mortality risk (RR = 0.61, 95% CI 0.46-0.80) compared to cancer survivors who did not change their PA level or who were insufficiently active pre diagnosis [18]. Therefore, even small increases in MVPA post diagnosis are likely to be beneficial for cancer survivors.
It is noteworthy that a third of the included studies were published in 2016, illustrating the rise in research interest of the effectiveness of DBCIs. It is interesting to compare the findings of this study with other meta-analyses using nondigital PA interventions among cancer survivors. While we did not find a significant reduction in fatigue (SMD = −0.23), Mishra et al. did find a significant improvement, both for survivors who had completed treatment (SMD = −0.82) [19] and who were still undergoing active treatment (SMD = −0.73) [74]. We also did not find any improvement in cancer-specific measures of QoL, while Mishra et al. reported positive improvements in both of their meta-analyses. Mishra et al. also found significantly improved anxiety and depression, sleep quality and improvements in some domains of generic measures of QoL (e.g. social, physical and role functioning) [19,74]. The small number of studies assessing these outcomes meant that meta-analyses for these outcomes were not possible in the current study. However, in the studies that did assess these outcomes, we did not find any evidence for an improvement in anxiety and depression, but both studies assessing sleep reported significant improvements. We also found mixed evidence for individual domains of generic QoL measures. It is possible that the non-digital interventions included in Mishra et al.'s studies result in larger effect sizes as many of the interventions are supervised by trained staff or involve some level of human interaction, which may foster higher levels of engagement and adherence to the intervention. However, due to the small number of low-quality studies included in this review, there is a need for more high-quality RCTs, with objective measures of PA, long-term follow-up and larger sample sizes before reliable comparisons between non-digital and DBCIs can be made. Sustained engagement with DBCIs was a significant problem for a number of the studies included in this review. For instance, in the study conducted by Short et al., retention at the 3-month follow-up was only 32% (156/ 492) and 11% (53/492) at the 6-month follow-up [72]. Furthermore, while 75% of the sample completed at least one action plan, the average number of action plans completed was only 2.2. Similarly, 50% of participants completed the week 1 module compared to 10% for the week 9 module in Forbes et al.'s study [63]. A systematic review has shown that there is a positive relationship between participants' adherence to/engagement with digital interventions and positive physical health outcomes across a range of populations and behaviours, suggesting that efforts to improve effective engagement with DBCIs could improve behaviour change outcomes [75]. It is possible that suboptimal engagement with the DBCIs in the studies included in this review may explain the reduced effects on outcomes compared to those observed in Mishra et al.'s review of non-digital PA interventions [19,74]. Future DBCI studies should integrate techniques or components that maintain effective engagement with the intervention for its duration. There is some evidence that technology-based strategies (e.g. reminders, prompts) can encourage user engagement [76]. Other aspects which have been identified as important for engagement include ease of use, design aesthetic, feedback, function, ability to change design to suit own preferences, tailored information and unique mobile phone features [77]. Similarly, less time consumption, user-friendly design, real-time feedback, individualised elements, detailed information and health professional involvement may also improve effectiveness of DBCIs, in particular mobile apps [78]. Furthermore, future studies should aim to better understand the link between engagement and effectiveness of DBCIs targeting PA and diet in cancer survivors and define, evaluate and report engagement more consistently so as to better understand techniques that foster effective engagement and mechanisms of action [79].
The majority (10/15) of the studies used an online portal or website to deliver the intervention, and, while one of these websites was mobile-enabled [65], only two studies used mobile apps [68,69]. This is interesting given the findings of a recent review of 23 interventions using mobile apps that found that 17 of the included studies reported a significant effect on behaviour change in the general population [78]. It is possible that interventions using mobile apps may be more effective than other types of DBCIs due to the widespread usage of and constant access to smartphones and the Internet. The most recent Ofcom Communications Market report conducted in the UK reports that 71% of UK adults own a smartphone and 66% use their smartphone to access the Internet most frequently [41]. Therefore, it would be interesting for future studies to use mobile apps as a mode of intervention delivery and compare the effectiveness of mobile app interventions compared to other DBCIs as they may foster higher levels of engagement.
Self-monitoring, goal-setting, credible source and feedback on behaviour were the most frequently described BCTs used in the included studies. Due to the heterogeneity in intervention type, mode of delivery, behavioural outcomes and measurement approaches, it is difficult to interpret which BCTs were most effective at changing PA/ dietary behaviour. Kanera et al.'s study used the most BCTs (n = 16) but also used a tailored if-then algorithm within the intervention to automatically tailor content to participants; therefore, it is difficult to ascertain what intervention components or approaches to delivery lead to increased effectiveness. A recent meta-analysis revealed that theory-based interventions are significantly more effective at improving PA [80]. SCT was the most commonly reported theoretical basis of the interventions; however, several other behaviour change theories were used across the studies. The level of reporting of the extent to which theory was incorporated into the development of the interventions varied across studies, but was generally poor. Only three studies did not report any theoretical underpinning. Future studies should aim to explicitly report how theory is used to develop intervention techniques and tailor the intervention to participants. Measurement and exploration of changes in targeted theoretical constructs (e.g. via mediation analyses) can aid understanding of why interventions may or may not be effective [59] and can be used to refine theoretical models of behaviour change. The lack of clarity about the intervention content and theoretical underpinning and the lack of measurement on theoretical constructs mean it is difficult to unpick how or why the interventions which improved behaviour in this review were effective. Future DBCI studies should clearly report any theoretical underpinning and behaviour change techniques used, for instance by using Michie et al.'s Theory Coding Scheme [59] and Behaviour Change Technique Taxonomy [56].
There are several limitations to this review. Primarily, the data extracted for the meta-analyses reflect unadjusted models. While the forest plot for MVPA minutes/week (Fig.  2) illustrates that Bantum et al. [61], Rabin et al. [54] and Kanera et al. [71] found a significant effect, the original publications show that this no longer remained significant when adjusting for baseline values and/or other covariates (e.g. demographics, disease characteristics) [54,61], or when adjusted models are controlled for multiple testing [71]. Bantum et al. [61] did find a significant increase in vigorous PA participation in the adjusted model, but not when combined with moderate PA. Similar issues arise for the meta-analyses for fatigue and BMI/weight. The studies that show significant findings for fatigue and BMI/weight in the current metaanalyses do not report significant findings when adjusted for baseline and/or covariates in the original publications. Therefore, pooling the adjusted results would reduce the overall effect size substantially, and it is likely that this would no longer remain significant. Furthermore, a combined MVPA variable was chosen to assess PA outcomes to maximise the number of studies that could be included in the meta-analysis and to coincide with the American College of Sports Medicine's recommendation that cancer survivors should follow the PA guidelines for the general population (at least 150 min of at least moderate intensity PA per week) [25]. However, this means the current meta-analysis does not differentiate between different intensities of PA: it may be beneficial to evaluate the effect on outcomes at different intensities of PA. Visual inspection of funnel plots suggested possible publication bias for smaller studies assessing MVPA outcomes; it is possible that our choice to only include published studies may have increased publication bias. There was no suggestion of publication bias for BMI/weight, fatigue or cancer-specific QoL measures.
The risk of bias in included studies was high. The inclusion of one-arm, pre-post studies substantially increases the risk of bias; however, we felt that the novelty of this area of research warranted the inclusion of these studies and that this added valuable insight into the current state of the literature. Few studies assessed outcomes other than PA, where only half of the studies could be included in a meta-analysis. Heterogeneity across studies was very high, likely due to the variability of types of DBCIs, intervention content, cancer type and populations, outcome measurement tools, etc. With the addition of future studies, more specific inclusion criteria could be used to assess effectiveness of more similar studies. All of the PA findings used self-report data, which while easy to use, inexpensive and validated, often hugely underestimate or overestimate PA participation [81]. Therefore, we advise that these results are interpreted with caution until the number of published studies assessing DBCIs increases and inclusion criteria for future systematic reviews can be more stringent for low-quality trials.
To the best of our knowledge, this is the first meta-analysis of DBCIs designed to improve PA and diet among cancer survivors. While the review shows some evidence of an improvement in MVPA, a reduction in BMI and a trend towards significance for fatigue, large, high-quality RCTs, with objective measures of PA and long-term follow-up, are lacking. Future studies should aim to address these limitations, but the approach of using digital technology in this context appears promising.

Compliance with ethical standards
Conflict of interest HP has previously received consultancy from myownteam. AR, AF, LS and MH declare that they have no conflict of interest. Ethical approval This article does not contain any studies with human participants or animals performed by any of the authors. The current study was a review/meta-analysis of existing studies that did not involve collection of new data.
Open Access This article is distributed under the terms of the Creative Comm ons Attribution 4.0 International License (http:// creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.