Chronic conditions are the most significant contributors to the global burden of disease, which is continuously increasing [1]. Living with ≥1 chronic conditions significantly impacts a person’s life, family, and society [1]. Physical inactivity is a major risk factor for at least 35 chronic conditions, which has become a global health concern [2]. Physical inactivity is expensive, constituting up to 4.6% of a country’s national health care expenditure [3]. In contrast, being physically active can benefit physical and psychosocial health [4]. In fact, following a structured exercise program can reduce the symptoms of chronic conditions with no additional risk of serious adverse events [5]. However, a high proportion of people living with chronic conditions do not meet the recommended levels of physical activity [6,7], possibly because of various reasons such as lack of access; travel distance; time inconvenience; cost; or internal factors such as pain, lack of motivation, or reluctance to engage in group activities [8].

Regarding physical activity interventions, it is essential to understand and adapt care according to the individual’s needs, preferences, and abilities [9]. One way to deliver and make physical activity available, individually or as an add-on, is through digital health interventions [10]. Digital health involves multiple solutions, including telephone calls, SMS text messages, mobile apps, websites, videoconferencing, and wearables (eg, activity trackers). The potential advantages of digital health solutions are convenience in terms of reduced distance, time, and cost, thereby limiting some of the more traditional barriers to physical encounters [10]. Digital health solutions enable personalized and person-centered tailored health interventions. Such personalized care empowers people to manage their health, increases health care efficiency, and reduces health care costs [11]. Consequently, the focus on digital health interventions has steadily grown over the last 2 decades [12], accelerating further in response to the COVID-19 pandemic [13]. This has led to digital health interventions being initiated without fully knowing their effect, potentially introducing people to ineffective interventions [14].

Several systematic reviews have investigated the effect of digital health interventions on health outcomes in people with individual chronic conditions (eg, chronic obstructive pulmonary disease [COPD], heart disease, and diabetes) [15-26]. Some reviews have focused on physical activity [19-21,23,25,27], but these included few studies, as they focused on a single condition, solution (eg, only telephone calls or apps), or outcome measure (eg, pain). This challenges the assessment of the benefits and harms of digital health intervention. Furthermore, the proportion of people with >1 chronic condition is increasing. More than half of the people with chronic conditions have at least 2 conditions [28]; therefore, to increase the power of the results and better resemble real-life clinical practice, including studies on several chronic conditions is relevant. Therefore, a more comprehensive approach is needed to provide clinicians and decision makers with the best current evidence.

With this systematic review and meta-analysis, we investigated the benefits and harms of digital health interventions promoting physical activity among people with ≥1 chronic conditions on objectively and subjectively measured physical activity and physical function, depression, anxiety, health-related quality of life (HRQOL), and nonserious and serious adverse events.

The study was preregistered at PROSPERO (CRD42020189028) and the study protocol (including amendments) is available in Multimedia Appendix 1 [4,26,29-44]. We followed the Cochrane Handbook for Systematic Reviews [45] and reported according to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [46]. The PRISMA checklist is available in Multimedia Appendix 2 [46].

As this research design only investigates aggregated data from published randomized controlled trials (RCTs) with no individual data, no ethics approval was required.

We included all peer-reviewed RCTs that compared an intervention with a digital health component in promoting physical activity to usual care or minimal intervention (eg, education or blinded wearables). Digital health interventions were defined as interventions in which any solution or technology (eg, website, app, and wearable) was used to deliver health information between health care providers and participants over a distance [47]. Studies were included if (1) physical activity (any activity that was up to the participants to choose), exercise (any prescribed activity), or exercise therapy (any planned and structured prescribed activity) was used or promoted; (2) the primary or secondary aim was physical activity promotion; and (3) digital health intervention was delivered with or without additional pharmacotherapy or other nondigital adjuvant interventions (eg, rehabilitation, education, or counseling).

Moreover, studies had to include adults (aged ≥18 years) with ≥1 of the following chronic conditions: depression or anxiety (mental health condition), ischemic heart disease or heart failure (heart disease), COPD, osteoarthritis of the hip or knee, hypertension, or type 2 diabetes. The inclusion criteria focused on these specific conditions, as they are among the most common worldwide according to the global burden of disease [1]. Furthermore, they share the risk factors of physical inactivity and systemic low-grade inflammation, which can cause a chain of events that can lead to worse outcomes and other chronic conditions [4,48].

Studies were excluded if they included angina, valvular heart disease, congenital heart disease, cardiomyopathy, hypertension if systolic blood pressure ≤139 mm HG or diastolic blood pressure ≤89 mm Hg, mixed populations of type 1 and 2 diabetes, tested the same digital health intervention but in different settings, or used the digital health solution for monitoring without any interaction with or feedback to the participants (ie, electronic health records or passive use of telemonitoring or a blinded actigraphy device).

The MEDLINE, Embase, CINAHL, and Cochrane Central Register of Controlled Trials databases were searched via Ovid from inception to October 28, 2022, with no language restrictions. The search string was developed based on previous searches used in systematic reviews for people with chronic conditions [49,50], and validated RCT search filters were used if available and applicable in each database [51]. All search strings are presented in Multimedia Appendix 3 [52-191]. In addition, studies found during full-text screens were manually added.

Initially, search duplicates were removed using the reference manager tool EndNote (version 20; Clarivate) [192] and the management tool Covidence (Veritas Health Innovation;) [193]. In total, 5 authors (AB, NN, FD, AL, and GZ) independently screened titles and abstracts, and 2 authors (AB and GZ) independently screened the full texts. Discrepancies were solved during regular meetings and resolved by discussing with a third study team member (CBJ). Although no restrictions were applied to the search, we only included studies if available in 1 of the following languages: English, Danish, Norwegian, Swedish, or Italian. Covidence was used throughout the selection process.

A standardized data extraction form (piloted in 5 studies) was used. A total of 4 authors (AB, BL, NN, and GZ; in pairs) independently extracted data at two time points: (1) immediately after the end of the intervention and (2) at follow-up, the time point closest to 12-month follow-up. Discrepancies were resolved by consensus or by referring to a third author (CBJ). WebPlotDigitizer (version 4.5; Ankit Rohatgi) was used to extract data from the figures [194]. The authors were contacted if the studies had missing or unclear information or data [195]. If the authors did not respond, the study eligibility was reassessed and consequently excluded.

Outcomes were hierarchically ordered (Table 1), with objectively measured outcomes prioritized over subjective and generic tests over disease-specific tests. The primary outcomes were objectively measured physical activity and physical function. The secondary outcomes were subjectively measured physical activity and physical function, depression, anxiety, HRQOL, and adverse events (nonserious and serious).

Extracted data included characteristics of the (1) trial—study design, study authors, publication year, country, and setting; (2) participants—number randomized, age, sex, BMI, socioeconomic status, and index conditions (if a study included participants based on them having ≥2 chronic conditions, the study was categorized as multimorbidity); and (3) intervention—type, frequency, duration, delivery method (categorized as mobile health [mHealth; ie, telephone calls, SMS text messages, and apps], eHealth [eg, internet based, website, and emails], or digital device [eg, wearables and Bluetooth-connected devices]), type of comparator, physical activity (categorized as physical activity [any activity that was up to the participants to choose] or exercise therapy [any planned and structured prescribed activity]), phase (categorized as without, with, or after any other rehabilitation), use of theory or framework, intervention components (categorized as digital only or combined in-person and digital [ie, other in-person adjuvant intervention, as rehabilitation, education, or counseling]), and adherence.

The risk of bias was assessed for each outcome in pairs by 7 authors (CD, SRM, TGH, RMA, LHT, GZ, AB) independently using the Cochrane Risk of Bias 2 tool [30]. Disagreements were resolved by reaching a consensus or by including a third author (GZ or AB). The quality of evidence of each outcome was assessed using the Grading of Recommendations Assessment, Development, and Evaluation approach by 2 authors (GZ and AB) [31]. An agreement was reached through consensus.

Meta-analysis was conducted using a random effects model, adjusted to Hedges g and using a restricted maximum likelihood method, as heterogeneity was expected. The effect of the intervention on continuous outcomes was expressed as standardized mean difference (SMD) for outcomes reported with different measures and with 95% CIs. The mean differences with 95% CIs were calculated for data on the same metric scale. Risk ratios were calculated with 95% CIs for dichotomous outcomes (ie, adverse events). Meta-analyses were visually presented as forest plots and were performed using the meta command in Stata (version 17.0; StataCorp LLC) [196]. The pooled SMD effect sizes were interpreted as 0.20, 0.50, and 0.80, representing small, moderate, and large effects, respectively [32]. The inconsistency of the results was evaluated using the I² and adjusted R² statistics and by visual inspection of the bubble plots and meta-regressions [52].

Univariable and multivariable meta-regression analyses were performed if ≥10 studies were included in the analyses to identify factors that could influence the effect estimates. For each outcome at the end of the intervention, the impact of (1) participant characteristics—age, sex, BMI, socioeconomic status, depression, and anxiety levels—and (2) intervention characteristics—number of digital and in-person sessions, frequency, duration, type of physical activity, phase, use of theory or framework, adherence, digital only, other in-person adjuvant intervention, and financial incentive—were assessed. Meta-regression analyses were performed using the meta regress command in Stata version 17.0 [196].

Subgroup analyses investigated whether the effects of digital health interventions varied among the included index conditions on all outcomes. Moreover, the delivery method and methodological quality according to the Cochrane Risk of Bias 2 tool (high, some concern, or low risk of bias) were assessed by subgroups for the primary outcomes.

Small study bias was assessed using funnel plots and Egger test (or Harbord test for risk ratios) by 2 authors (GZ and AB), and agreement was reached by consensus.

The search retrieved 14,078 hits; 427 papers were full-text screened, and 140 papers were included, representing 130 RCTs (see Figure 1 for flowchart and refer to Multimedia Appendix 3 for the list of excluded studies). The included RCTs allowed for 136 comparisons, representing 20,094 participants (mean age 60.68, SD 7.17 years; 8440/20,094, 41.63% female). The interventions were conducted in 31 countries, mainly in the United States (24 RCTs) [53-75] and Australia (21 RCTs) [76-96]. Over half (91/140, 65%) of the RCTs were published between 2018 and 2022; however, the publication years ranged from 2003 to 2022. For study characteristics, refer to Multimedia Appendix 3.

In total, 80 RCTs used a digital-only approach [53,54,56,59-62,65,66,68-74,76,79,80,83,85,86,88-93,95-144], whereas 56 RCTs combined digital solutions with nondigital interventions (eg, counseling or in-person guided exercise) [55,57,58,63,64,66,67,75,77,78,81,82,84,87,94,142,145-188]. Of the included studies, 7 offered a preintervention period as run-in exercise (a period before randomization and intervention start where all participants performed exercise) or rehabilitation [63,98,108,111,120,160,183,189]. Furthermore, 13 RCTs offered rehabilitation as part of the intervention [64,75,80,82,147,149,151,152,157,160,163,169,178].

mHealth solutions (eg, telephone calls, text messages, and apps) were primarily used in 56 RCTs [53,56,57,59,63,67,69,78-80,​83-86,90,92,93,96,99,101,102,105,108,110,112-115,119,​120,124,128-130,134,136,140-142,146-148,150,155,157,161,163,​164,167,168,172,176,177,179,181,185], whereas 22 RCTs [58,60,61,65,70,81,87,111,116,117,126,133,134,145,149,151,154,156,​165,171,173,174] used mHealth combined with a device (eg, wearable, pedometer, activity tracker, or Bluetooth-connected pulse oximeter or glycosometer). Furthermore, 6 RCTs [75,89,118,158,169,183] combined mHealth and eHealth (eg, websites, web apps, and emails) technology. eHealth solutions were used by 29 RCTs [54,55,66,72-74,76,77,88,91,95,​97,100,103,104,106,122,125,127,131,132,138,143,144,152,159,175,184,190], whereas 6 RCTs [64,71,82,98,137,162,189] combined eHealth and a device. Digital devices alone were used in 14 RCTs [62,68,94,107,109,121,123,130,139,166,170,178,182,186,188,191]. Finally, 2 studies combined mHealth, eHealth, and a digital device [132,160].

The RCTs had a mean intervention duration of 22.94 (SD 18.72; range 4-104) weeks, and 60.9% (81/136) RCTs were placed in the physical activity category. A theory or framework most often (71/136, 52.6%) guided the interventions. The mean number of digital sessions was 97.98 (SD 109.29; range 2-377), whereas in-person sessions had a mean of 3.26 (SD 2.93; range 0-18). Of the included RCTs, 31 [54,58,62,65,70,76,​77,82,84,86,88,92,94,95,100,101,112,121,132,135,143,​144,146,150,151,155,163,164,174,182,186,189-191,197] had follow-up data, and the mean follow-up time was 45.38 (SD 21.74; range 12-104) weeks.

Type 2 diabetes [53,56,58,59,66,67,81,84,88,92,93,97,99,104,​107,114-116,119,122,123,126-128,130,131,140,145,146,155,157,159,165,168,​172,182,184] and heart disease [62,64,65,70,72,73,75,80,85,87,​96,98,102,106,109,111-113,121,132-135,139,142,147,154,160,161,169,171,​173-181,185,189] were the most frequent index conditions (41 RCTs each), followed by COPD (26 RCTs) [60,61,63,68,71,74,79,82,91,94,105,108,117,120,124,​136,141,149,151-153,156,163,164,167,170,183,186,190,191], osteoarthritis (18 RCTs) [54,55,57,76-78,86,89,90,100,110,​118,129,138,148,158,166,188], mental health (5 RCTs) [95,125,137,143,144,162,198], hypertension (3 RCTs) [83,101,103,197], and multimorbidity (2 RCTs) [69,150]. For the index conditions and comorbidities among the participants, refer to Multimedia Appendix 3.

At the end of the intervention, the digital health interventions, compared with usual care or minimal intervention, showed a small improvement in objectively measured physical activity SMD 0.29 (95% CI 0.21-0.37), corresponding to an average increase of 970.79 daily steps (95% CI 657.11-1284.47; I²=54.04%). Objectively measured physical function also showed a small improvement in favor of the digital health interventions (SMD 0.36, 95% CI 0.13-0.59) consistent with an improvement of 19.82 m (95% CI 9.45-30.02; I²=75.94%) on the 6-minute walk test (6MWT). There was no between-group difference in objectively measured moderate to vigorous physical activity (MVPA; SMD 0.03, 95% CI −0.31 to 0.37). However, objectively measured physical function and MVPA showed substantial heterogeneity. See Figure 2 for the overall forest plot and refer to Multimedia Appendix 4 [52-191] for a detailed forest plot of each outcome, including daily steps and 6MWT.

End-of-intervention effects were found in favor of the digital health interventions for subjectively measured physical activity and physical function, depression, anxiety, and HRQOL. Subjectively measured physical activity and HRQOL showed substantial heterogeneity (Figures 3 and 4; Multimedia Appendix 4).

There was an increased risk of nonserious adverse events (eg, musculoskeletal problems, soreness, or pain) in the intervention groups (risk ratio 1.31, 95% CI 1.11-1.55), whereas the risk of serious adverse events did not differ (risk ratio 0.89, 95% CI 0.76 to 1.04; Multimedia Appendix 4).

At follow-up, objectively measured physical activity showed a small improvement (SMD 0.17, 95% CI 0.04-0.31; I²=50.85%) in favor of the digital health intervention groups but with no between-group difference in daily steps (mean daily steps 498.40, 95% CI −44.99 to 1041.80; I²=77.94%). For objectively measured physical function, a small improvement was found in favor of digital health interventions (SMD 0.29, 95% CI 0.01-0.57; I²=76.17%). The improvement in 6MWT was sustained (50.06 m, 95% CI 15.1-85.01; I²=71.28%). However, a small effect on objectively measured MVPA in favor of the intervention group was found at follow-up (SMD 0.23, 95% CI 0.06-.40, I²=29.45%). See Figure 5 for the overall forest plot and refer to Multimedia Appendix 4 for a detailed forest plot of each outcome, including daily steps and 6MWT.

Only the effect on subjectively measured physical activity was sustained at the follow-up of the secondary outcomes (SMD 0.36, 95% CI 0.05-0.66; Figures 6 and 7; Multimedia Appendix 4). However, in the follow-up period, neither nonserious (risk ratio 1.35, 95% CI 0.71-2.54) nor serious adverse events (risk ratio 0.76, 95% CI 0.56-1.03) showed any between-group differences (Multimedia Appendix 4).

For objectively measured physical activity, 6% (3/51) of the RCTs were classified as low risk of bias [58,84,160], 55% (28/51) of the RCTs had some concerns [53,61,​62,68,70,71,74,81,82,95,97,98,100,102,107,111,115,​117,136,143,144,147,148,155,156,166,180,182,183,188-191], and 39% (20/51) of the RCTs were classified as high risk of bias [60,65,69,81,91,94,99,104,125,130-132,145,146,149,151,​152,163,164,186] (Figure 8). For objectively measured physical function, 4% (2/47) of the RCTs were classified as low risk of bias [110,158], 70% (33/47) of the RCTs had some concerns [54,55,64,71,74,79,82,87,94,96,97,102,105,108,117,​118,120,121,124,133,136,147,156,159,167,171,175,178,183,​185,186,190], and 26% (12/47) of the RCTs were classified as high risk of bias [57,63,91,112,126,149,151-154,163,164,​170,176] (Figure 9). The risk of bias profiles were similar for the secondary outcomes (Multimedia Appendix 5 [52-191]).

Small study bias was judged as present at the end of the intervention for objectively measured physical function, subjectively measured physical activity, and HRQOL and at follow-up for subjectively measured physical activity (Multimedia Appendix 6 [52-191]).

The Grading of Recommendations Assessment, Development, and Evaluation assessment scores ranged from moderate to very low (Table 2). Downgrading was done owing to study limitations, imprecision of the findings (some conditions had a limited number of RCTs), and inconsistency (high statistical heterogeneity only partially explained by the meta-regression analyses).

Univariate meta-regression analysis for objectively measured physical activity showed that increasing age was associated with a lower effect size (slope −0.01, 95% CI −0.03 to −0.00), suggesting an effect reduction of 0.01 SDs per increased year. For objectively measured physical function, a higher effect was found for interventions that combined exercise therapy and physical activity (coefficient [β]=1.40, 95% CI 0.42-2.36). A similar effect of age was found for subjectively measured physical activity (slope −0.02, 95% CI −0.03 to −0.00), suggesting an effect reduction of 0.02 SDs per increased year; however, BMI was also associated with a lower effect size (β=.04, 95% CI −0.07 to −0.01) for this outcome. Furthermore, for subjectively measured physical activity, shorter intervention duration showed a lower effect size (β=−.01, 95% CI −0.01 to −0.00), and the use of a theory or framework showed a decreased effect (β=−.31, 95% CI −0.50 to −0.11), which persisted in the multivariate model (adjusted for intervention session, duration, and frequency; β=−.28, 95% CI −0.46 to −0.09). For mental health conditions, the univariate analysis showed that participants with higher depression levels had a greater effect on depression symptoms (slope −0.01, 95% CI −0.02 to −0.00), which persisted in the multivariate model (adjusted for age and sex). There were no differences in the rest of the meta-regression analyses (Multimedia Appendix 7) or in the subgroup analyses stratified by conditions, delivery method, or methodology quality (Multimedia Appendix 8 [52-191]).

In an unplanned sensitivity analysis of the primary outcomes of RCTs with a digital-only approach, the results at the end of the intervention were in line with the main analyses (Multimedia Appendix 8).

Overall, digital health interventions had a small effect on physical activity and physical function in people with chronic conditions compared with usual care or minimal intervention at the end of the intervention and follow-up. Furthermore, subjectively measured physical function, depression, anxiety, and HRQOL showed a small improvement but only at the end of the intervention. The certainty of the findings was downgraded because of study limitations, imprecision, and inconsistency and was moderate to very low.

This study had some limitations. The risk of selection bias within the RCTs, our inclusion criteria, which focused on 8 specific chronic conditions, thereby excluded, for example, cancer, and the fact that only a few RCTs targeted hypertension, depression, anxiety, and multimorbidity limits the generalizability of our results. Furthermore, few RCTs had follow-up data that induced uncertainty regarding whether the effect was maintained beyond the end of the interventions. In addition, the reporting of digital health solutions and the type and dose of physical activity were poor for many RCTs, preventing us from conducting further analyses on the intervention characteristics. Another limitation is the lack of measuring digital health literacy within the included studies, as it is known that low literacy (both health and digital) affects participation and health outcomes and deepens health inequities [222]. Therefore, we lack information on whether the participants in the included studies were the ones with high or low levels of digital health literacy, as we may suspect that in groups with lower literacy, the effect would become even smaller [223].

The quality of the evidence had to be downgraded to low or very low for objectively measured physical function, subjectively measured physical activity, anxiety, and HRQOL, implying that the actual effect might be different from the estimated effect. In addition, a small study bias cannot be ruled out for some outcomes, although other factors, such as high heterogeneity and low quality of evidence, may contribute to funnel plot asymmetry.

Nevertheless, the strengths of this systematic review are the meticulous, state-of-the-art design and methodology, following the Cochrane handbook and reporting using the PRISMA guidelines. In addition, we assessed the effects across several chronic conditions, breaking down the more silo-based thinking of chronicity.

More high-quality research with better reporting, available protocols, blinding of assessors and study participants (when possible), better handling of missing data, and the use of common and clearly described measurement methods are needed for future research. In addition, such research should assess whether higher and more sustainable effects of digital health interventions that promote physical activity can be achieved. It should also focus on whether any digital health solution or features within these are more effective than others and whether a dose-response relationship exists. Studies focusing on the impact of comorbidities or multimorbidities are also warranted. Furthermore, if digital health solutions are to live up to their potential and be part of the solution to the health care shortage, the digital divide must be addressed by assessing if and how the participants’ digital health literacy levels affect digital health interventions.

Engaging people in physical activity is considered a polypill that provides high health gains with low risk and low costs [224]. Digital health solutions as a delivery method can provide effective physical activity interventions across people with chronic conditions and may help make these types of interventions available for more people at a lower cost. Therefore, given the potential health benefits, digital health interventions promoting physical activity in people with ≥1 chronic condition should be considered in clinical practice, especially if the overall aim is to improve physical activity levels. However, educating participants about the potential risk of nonserious adverse events is essential. The findings of this study are also relevant for policy makers because easy and affordable access to high-quality health care services is at the top of the political agenda. However, it should be noted that the effects were small and, for most outcomes, limited to the end of the intervention, with analyses showing substantial heterogeneity between studies and many studies with some concerns for the overall risk of bias.