Use of mHealth solutions in managing chronic conditions is increasing; however, the effective and long-term engagement (ie, attrition rate) has been attributed to various factors.

After database searches were performed, a total of 1420 articles were identified. After excluding duplicates, 831 unique records remained. Of these, 797 were excluded prior to review. A further 2 records were identified through reference screening from included studies, leaving a total of 36 studies to be included in the review (see Figure 1). Of the 36 studies included in the final review, 19 were excluded based on the exclusion criteria of studies including children and studies that only looked at acute or infectious diseases. Studies that were purely online, telephone and texting interventions, and studies that did not have any measurement of rates of dropout and attrition were also excluded. This left 17 studies to be included in the final qualitative and quantitative synthesis.

Included studies were published between 2011 and 2019 [14,16-31]. Of these, most (14/17, 82%) examined a range of chronic diseases, including single studies targeting lower back pain, chronic kidney disease, pain, dysmenorrhea, and HIV medications; the remainder (9/17, 53%) looked at more general lifestyle improvement, such as eating behavior and physical activity. Out of 17 studies, 3 (18%) that were included in the review looked specifically at diabetes. There were 9 RCTs (53%) included in the final synthesis and 8 observational trials (47%). These are summarized in Table 1.

Studies ranged significantly in duration, size, attrition rate, methodology, and other areas. The shortest trial included in this review lasted 2 weeks, and a total of 5 out of 17 (29%) lasted one month or less. Out of 9 RCTs, 2 (22%) looked at 1 year of data, and a number of observational trials were conducted over a period of 6-10 months. The lowest attrition rate in any study was 9% in an RCT lasting 1 year [29]; the highest attrition rate was 82% in an observational trial lasting 6 weeks [26]. The largest trial was an observational study lasting 24 weeks, with nearly 200,000 participants; the smallest trial was a small cohort study including just 20 people.

Results from the meta-analysis are presented in Figure 2. The average attrition rate overall was 43% (95% CI 29-57), with very high between-study heterogeneity indicated by an I² statistic of >99%. The very high heterogeneity is not unexpected in this case, as studies were extremely varied in terms of time, implementation, and the disease state that they were examining.

Looking at the breakdown of results by the type of study, there was a higher degree of attrition in the observational real-world studies (49%, 95% CI 27-70) than in the RCTs in more controlled scenarios, with only 40% (95% CI 16-63) dropping out. Sensitivity analyses looking at differences in length of study (ie, short vs long), diabetes versus other chronic diseases, or whether the studies were numerically large did not find any similar differences in attrition rate between trials.

One reason associated with lower attrition rates was the behavioral characteristics of the included participants. Low attrition rates were characterized by reasons such as the perception of own health as poor—thus, incentivizing the need to change [18]—and those who wanted to be involved in their health care [20]. Other factors that were associated with attrition in included studies were health literacy; age, with younger participants dropping out less; and postgraduate education [18,22]. Very low attrition was also reported among those who were on strict diets or who had been healthy eaters prior to the initiation of the study [14]. Another association with low attrition was with those engaged in multiple interventions. Those engaged in internet or phone programs as well as apps were more likely to remain in the research study [23]. Conversely, there did not appear to be much influence on attrition rates in terms of length of study, the disease studied, or the size of the app trial.

Findings from these studies suggest ways to improve attrition rate and long-term engagement by using varying message contents or formats to maintain users’ interests; for example, tailored messages may have the potential to improve adherence to a clinically significant degree [16]. Other studies suggest less of a benefit from tailoring messages to maintain users’ interest; despite a low attrition rate of 22% at 4 months and 1 year in two studies, respectively, an app and health counseling did not reduce hemoglobin A1c levels between the intervention and usual care groups [17,30]. In addition, self-management skills and the ability to contact health professionals were found to increase engagement, while users’ feedback input improved usability of apps and enhanced user experiences for daily self-reports [17,19]. Classifying different types of users may be important in improving long-term engagement. Low retention rate might also be due to an unguided self-help approach, and further engaging those who need self-monitoring remains challenging.

Another issue with attrition was that definitions varied significantly. While some studies reported users who only logged in a single time as dropouts, others expanded the definition to include those who only used an app once or twice. For example, the RCT with the lowest dropout rate overall included patients who only sent a single report through the app during the entire follow-up time, which did not indicate sustained, long-term use [29]. While these users may have been nondropouts by the definitions used in that study, they had received significantly lower benefits from the intervention, and would likely be considered dropouts had the analysis been less broad.

Attrition in app-based interventions is an important and yet underresearched element. For these interventions to work, it is a necessary component that people use and continue to use the app; however, there appears to be evidence that this is not always the case. In this systematic review and meta-analysis, the pooled estimate of dropout rates was 43%, with higher rates seen in real-world research and lower rates in highly supported RCTs. This may indicate a very serious underlying issue, as high dropout rates in these interventions will limit their use and uptake in health care across a range of chronic diseases.

While dropout rates in RCTs were notably lower than in observational trials, it is worth noting that attrition was often defined differently in this research. RCTs tended to describe all participants as users of the app unless they had ceased using it entirely; while this is in line with best-practice intention-to-treat analyses, it also presents an important limitation with the pooled results above. Including people in analyses who have been randomized is laudable; however, this also obscures the fact that large proportions of people, even in these randomized trials with detailed patient support and follow-up, barely used the app, if at all. This is worrying, because it implies that even with very high levels of support, apps are not an intervention with substantial staying power for people with various chronic diseases. It is also worth recognizing that combining estimates for both observational and RCT studies may have some drawbacks, given the differences in study design and participant retention. However, the difference in reported dropout rates was not significantly different, with only a small difference in point estimates and overlapping confidence intervals. This is likely because the main divergence was in definition of attrition, rather than the specifics of the study per se. As reported above, even among RCTs with low rates of dropout, there may have been a number of studies that would meet the looser definitions of attrition used in some observational research.

It was also concerning that there does not appear to have been much examination of the reasons behind this attrition in many studies. Few studies attempted to explain why people dropped out, with this being attributed to health literacy, age, and education, but it is unlikely that these are the only factors that would be related to attrition in the use of apps. For example, as mentioned in the results, people whose health saw greater improvements were more likely in some studies to keep using the app. It is likely that there are a range of unidentified issues that could potentially be targeted to ameliorate this problem, but thus far there has been little recognition of the issue formally, which may have limited the research that has been done to remedy the situation. Many studies do not even address the possibility that people dropping out of an app-based intervention at alarming rates could be an issue for the intervention’s adoption at scale, nor do they address the issue that this could cause in terms of aggravating health inequities depending on the reasons for dropout. This is especially concerning when considering that age and social status are likely to be barriers to app access—as some included studies hinted—which may further compound the issues caused by selection bias of those who use apps in the first place. If younger, healthier people are more likely to use apps overall, which is often the case [32], and are then more likely to use them long term, the apps may be less useful for the very populations that we most want them to help.

This is a common theme among trials, in which attrition or nonusage is barely addressed, or only given very surface-level appraisal. If there is a significant difference in the primary outcome between the intervention and control groups, there is a general attitude that the attrition is unimportant; this appears to be fairly common in RCTs, and may be because the aim of this research is specifically to evaluate the app in an intention-to-treat framework [29,30]. However, there are clear drawbacks to this, not least that we may be seeing a large underestimate in the literature of the efficacy of app-based interventions, caused by their generally low use in the populations who have been studied.

There are a number of very important limitations to this research. Firstly, the estimates produced are certainly not comprehensive. Many studies (N>30) that fit all of the inclusion criteria failed to report dropout or attrition in a way that could be extracted. Given the number of trials on app-based interventions, it was not considered feasible to follow up with every author group that had these figures, but it is worth noting that this best guess represents a relatively small number of trials within the total pool of potential evidence.

There is also the issue with heterogeneity. Given the nature of the included studies, it is not surprising to find very significant levels of heterogeneity statistically, but it is concerning for the meta-analysis as a reasonable estimate of attrition. These studies were conducted across different disease states, with highly variable interventions; the fact that they all included an app is a thin bond that did not overcome the vast differences that they had between them.

It is also worth noting that the research in this space is quickly evolving. We found no published research to be included before 2010, very little in the years leading up to 2015, and then an explosion of studies in the years after. It is likely that redoing this meta-analysis in 2025 will yield a much more reliable estimate of the figures. This may also allow for analyzing by disease state, which could prove to be a more accurate estimation of the rate of attrition.

There are a number of theories that might pertain to attrition in app-based interventions, with several different focuses. One area that could help inform attrition research in the future is behavioral theory, perhaps by examining the sociocognitive aspects of people who do and do not drop out of app-based studies. Integrated behavioral theories might also be useful in examining the relationship between social factors and the behaviors they cause, to come to an understanding of the process by which people decide to use or discontinue using apps.

This would ideally tie in to an examination of the broader social and demographic drivers of attrition. While few of these drivers have been identified in research so far—age, in particular—there remains a large evidential gap pertaining to how society influences behavior to prevent people from using app-based interventions. Future research should combine these two theoretical approaches to define the background reasons for attrition, so that interventions can be designed to minimize it.

Aside from the estimates of app attrition, there are some important implications of this research. Future studies looking at app-based interventions should include attrition as a secondary endpoint and develop methods to prevent it if possible. One important aspect would be to develop a standard measure of minimum use in app-based interventions; a reasonable example is the one used in some of the included trials of one or fewer log-ins to the app in any given period of time (ie, one log-in per month). Lower use than this basic threshold could then be considered attrition for the purposes of research studies. As well, there should be trials looking at ways to reduce the rate of dropouts, as well as the potential inequity in the rate of attrition, in app-based interventions. Without such research, we have no way of knowing if apps can be effective in the general population.

This systematic review and meta-analysis found that the pooled estimate for dropout rates in trials of app-based interventions for chronic diseases was 43% over a variety of timelines, with the length of time having little impact on the rate of dropout. Attrition was higher in observational real-world studies, with randomized clinical research seeing less than a third of patients drop out before the trials were completed. However, findings were limited by high heterogeneity and the lack of reporting in many trials on attrition rates. Future research should focus on how often patients drop out and examine reasons why, so that this important issue can be addressed in app-based interventions for chronic disease.