This systematic review adheres to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) checklist (Checklist 1) and has been registered in PROSPERO (CRD42024522261) [15].

The initial literature search strategy for this systematic review was designed according to the preregistered protocol. However, the original protocol included only Chinese and English studies and used insufficient search terms, resulting in limitations of the search strategy. Therefore, we optimized the search strategy in the final draft to enhance comprehensiveness and reduce language bias. These modifications have been explicitly reported as post-hoc adjustments in accordance with research standards.

This systematic review’s literature search has been reported according to the PRISMA-S statement [16]. A combination of subject headings and free-text terms was used for systematic retrieval, covering the following databases and registries: PubMed (NCBI), Embase (Embase.com), Web of Science (Clarivate), Cochrane Library (Wiley), CINAHL (EBSCO), China National Knowledge Infrastructure (CNKI), CqVIP (VIP Information), Sinomed (China Biology Medicine Disc; CBMdisc Platform), and ClinicalTrials.gov (NLM). The literature search timeframe was set from database inception to June 28, 2025. To ensure the timeliness of the search results, the retrieval strategy was optimized and rerun on December 31, 2025, to incorporate newly published literature.

The search strategy for this systematic review was developed by refining and optimizing strategies from prior studies [17], and it underwent peer review by experts in library and information retrieval. Search parameters were set to include only randomized controlled trials (RCTs) and peer-reviewed articles, with no restrictions on publication year or language, and no predefined filters were applied. To maximize retrieval volume, snowball citation searching was performed. This involved screening the reference lists of the included studies (backward citation searching) as well as identifying and reviewing studies that cited the included articles (forward citation searching). It should be noted that the literature search for this review was conducted exclusively using the aforementioned method and did not include searches of printed conference proceedings or government websites. Detailed search strategies are provided in Multimedia Appendix 1.

The mHealth interventions included in this review varied in their implementation formats. To facilitate subsequent integration and comparison, we predefined the following classification criteria based on the primary delivery platform and user interaction model of the intervention content [18]:

After importing the retrieved studies into EndNote 20.6 (Clarivate Analytics) and removing duplicates using the software, 2 researchers (JHY and QYZ) were invited to cross-review the titles and abstracts to complete the initial screening and full-text reading of the remaining literature. When disagreements arose between the 2 researchers regarding inclusion decisions, a third researcher (WCL) was consulted to participate in discussions. Where necessary, the third researcher made the final determination on whether to include a study. For studies lacking medication adherence data or where full-text access was unavailable, we first attempted to contact the corresponding author to obtain the data. If the author did not respond or the data could not be obtained, the study was excluded from the meta-analysis.

For included studies, data were extracted using Microsoft Excel software, encompassing the following: (1) general information (eg, first author’s name, nationality, publication year, theoretical framework, study design, and background); (2) characteristics of patients with cancer (eg, sample size, mean age/age range, and disease type); (3) intervention characteristics (eg, intervention format, intervention content, intervention duration, and intervention frequency); and (4) primary outcomes and measurement tools.

Two researchers (QYZ and JHY) assessed the risk of bias in RCTs using the 7 domains of The Cochrane Collaboration’s tool for risk of bias assessment (version 2.0) [23]. The Cochrane risk of bias assessment tool encompasses 5 key domains: randomization process, intervention deviation, missing data, outcome assessment, and selective reporting. Risks within each domain are categorized as low risk, high risk, or unclear risk. Studies meeting low risk in all domains are classified as grade A, those meeting criteria in some domains are classified as grade B, and those failing to meet criteria in any domain are classified as grade C. This systematic review included only grade A or B literature, and evidence quality was assessed using GRADE Pro software (Evidence Prime Inc). This tool categorizes evidence quality into 4 levels: high, moderate, low, and very low. Factors leading to downgrading encompass 5 aspects: risk of bias, imprecision, inconsistency, indirectness, and publication bias. High-quality evidence indicates very high certainty that the true effect is close to the estimated effect; moderate-quality evidence reflects moderate confidence in the effect estimate, with the true effect potentially close to but possibly substantially different from the estimate; low-quality evidence indicates limited certainty about the effect estimate, with the true effect likely substantially different from the estimate; and very low–quality evidence indicates almost no confidence in the effect estimate, with the true effect highly likely to be substantially different from the estimate [24]. When 2 researchers (QYZ and JHY) disagreed on the risk of bias or evidence grade of a study, a third researcher (WCL) was consulted. If necessary, the third researcher made the final decision on whether to include the study.

A narrative synthesis approach was used to summarize study characteristics, including research design, publication year, country, participant characteristics, and intervention features. For outcome measures unsuitable for meta-analysis, a narrative synthesis approach was also applied for analysis.

Although the preregistration protocol specified the use of Stata (Version 17.0; StataCorp) and RevMan (Version 5.4; Cochrane) for statistical analysis, the final statistical analysis was conducted in RStudio (Posit) using a conservative random-effects model based on the assumption of between-study heterogeneity. Effect sizes were pooled using the Hartung-Knapp-Sidik-Jonkman (HKSJ) method [25]. To enhance methodological rigor and clinical interpretability, multiple analytical measures were introduced during the optimization phase: the Egger test was used to assess small-sample effects, 95% prediction intervals (PIs) were incorporated, τ² was used to quantify heterogeneity, and the GRADE (Grading of Recommendations Assessment, Development, and Evaluation) approach was used to assess evidence quality [26].

For continuous variables, the SMD was used to calculate the pooled effect size, interpreted according to Cohen d criteria: 0.2, 0.5, and 0.8 represent small, moderate, and large effects, respectively. For dichotomous variables, the OR was used. Each pooled effect size is accompanied by a 95% CI. Additionally, we report PIs to quantify the variability of actual effects across different populations or settings [27].

The Cochrane Collaboration recommends the HKSJ method for random-effects meta-analysis [28]. When facing insufficient statistical power due to a small number of included studies, random-effects meta-analysis based on the HKSJ method outperforms the standard DerSimonian-Laird method [27]. By using a t-distribution and adjusting SEs using the q statistic, the HKSJ method provides more robust effect size CI estimates. Particularly with limited studies, HKSJ-estimated CIs are more accurate as they better reflect the inherent uncertainty and imprecision of effect estimates [29]. Model selection should be based on conceptual assumptions about the true effect distribution rather than descriptive statistics like I² [26]. Given the presence of differences in participant characteristics, interventions, and measurement tools across the included studies, a conservative random-effects model was used for all analyses. This systematic review used restricted maximum likelihood estimation to account for between-study variance, thereby more accurately reflecting the impact of true heterogeneity on effect estimates. Heterogeneity was quantified using Q-tests, I², and τ² measures [27]. All statistical analyses were conducted in the R Studio environment using R software (version 4.5.2), utilizing packages including metafor, forestplot, and ggplot2.

Subgroup analyses of postintervention data were based on priority intervention criteria: (1) intervention duration, (2) cancer type, and (3) intervention type. Sensitivity analyses were conducted to assess result stability. Funnel plots were generated, and Egger regression intercept tests were applied to examine small-sample effects. Note that these methods aim to detect small-sample effects rather than directly assess publication bias [30,31]. Small-sample effects may stem from publication bias, selective reporting, genuine heterogeneity between studies, and methodological differences. According to methodological standards, funnel plots require at least 10 studies for reliable interpretation; otherwise, the reliability of interpretations is low [30]. Egger regression intercept analysis was applied to all outcome measures in this systematic review.

This systematic review strictly followed the protocol registered in advance on the PROSPERO platform (CRD42024522261). During methodological refinement and peer review, we identified the following limitations in the original protocol and made two subsequent revisions accordingly:

This literature search identified a total of 848 studies, including 846 from databases and clinical trial registries and 2 from reference lists. After deduplication, 703 studies were ultimately retained. Following preliminary screening of titles and abstracts against the inclusion criteria, 30 studies underwent full-text assessment. After full-text evaluation, 13 studies were excluded: 6 due to incomplete data, 5 due to unavailable full texts, and 2 due to differing outcome measures. Ultimately, 17 studies [19-22,32-44] were included (Figure 1).

This systematic review included 17 RCTs. Six studies were published in Simplified Chinese [20,21,36,37,42,44], while the remaining 11 were English-language publications [19,22,32-35,38-41,43], spanning publication dates from 2016 to 2025. Studies originated from 8 countries: the United States, China, Turkey, Australia, the Netherlands, South Korea, Iran, and Finland, and all were conducted in hospital and home settings.

The studies included a total of 1309 participants, with sample sizes ranging from 28 to 166 cases. Cancer types included breast cancer (3 studies) [33,39,43]; leukemia (4 studies) [22,35,40,41]; unspecified cancer types (7 studies) [19,32,34,37,38,42,44]; and thyroid cancer, pheochromocytoma, and prostate cancer (1 study each) [20,21,36].

Regarding intervention formats, the primary methods used were mobile apps (10 studies) [19,33,34,36,37,40-44], text messaging services (5 studies) [20,32,35,38,39], and websites (2 studies) [21,22]. Additionally, 2 studies utilized smart pill bottles as supplementary tools, integrated with either mobile apps or text messaging services [39,43]. Regarding intervention content, mobile apps and websites primarily enhanced medication adherence through online Q&A sessions, multiplayer online games, and science-popularization video pushes, while text messaging services focused on delivering personalized medication reminders.

Regarding intervention providers, 7 studies explicitly reported that physicians and/or nurses delivered interventions via mHealth platforms [20,21,37,39,40,42,44], while the remaining 10 did not specify the type of health care professional [19,22,32-36,38,41,43]. The intervention duration ranged from 3 weeks to 9 months, with only 3 studies reporting frequency (daily to 3 times weekly) [20,33,40]. Regarding theoretical frameworks, only 1 study designed interventions based on social cognitive theory [32].

All studies assigned control groups to receive usual care. Twelve studies described usual care components, including health education, medication guidance, and regular follow-up [19-21,32-40], while the remaining 5 did not provide specific details [22,41-44]. Detailed study characteristics are summarized in Table 1.

Ten studies reported randomization methods, while 5 described allocation concealment methods. Since improving medication adherence among patients with cancer using mHealth technology constitutes a behavioral intervention, blinding participants and researchers proved challenging. Consequently, none of the studies used blinding, resulting in a high risk of performance bias. Furthermore, as outcome measures were primarily assessed using subjective scales, the risk of measurement bias remains unclear. Only 2 studies used intention-to-treat analysis to address missing outcome data. No studies demonstrated selective reporting bias or other types of bias, resulting in an overall quality rating of grade B (Figure 2).

This systematic review used the GRADE evidence rating tool to assess the quality of evidence for each outcome measure. Overall, evidence quality ranged from moderate to very low. For the impact of mHealth interventions on medication adherence rates, the evidence quality was moderate, downgraded primarily due to risk of bias and imprecise data. For the effect of mHealth on the mean medication adherence score, the evidence quality was very low, downgraded due to high risk of bias, inconsistent results, and imprecise data. For the effect of mHealth on symptom burden, the evidence quality was low, downgraded due to risk of bias and imprecise data.

Regarding the effect of mHealth on self-efficacy, the evidence quality was also low, downgraded due to risk of bias and imprecision. The GRADE evidence summaries for each outcome are presented in Table 2, with detailed downgrade justifications available in Multimedia Appendix 2.

To assess the small-sample effect, we followed the methodological guidance by Sterne et al [30], combining visual inspection of funnel plots with the Egger regression intercept test to analyze 4 outcome measures: mean medication adherence scores, medication adherence, symptom burden, and self-efficacy.

It should be noted that, except for the mean medication adherence score, fewer than 10 studies were included for each of the other outcome measures, which is below the ideal threshold for reliable assessment proposed by Lau et al [45]. Nevertheless, funnel plots were constructed for each measure (Figures S6-S9 in Multimedia Appendix 3) for preliminary exploration. Visual inspection revealed no obvious asymmetry in the funnel plots; however, this visual assessment itself carries considerable uncertainty due to the limited number of studies.

Subsequent quantitative analysis using the Egger test also failed to indicate statistically significant small-sample effects (medication adherence average score: z=1.64; P=.75; symptom burden: z=0.18; P=.91; self-efficacy: z=1.29; P=.78; and medication adherence rate: z=1.53; P=.26).

However, these findings must be interpreted with caution. The primary limitation lies in the insufficient number of included studies. First, a small sample size significantly reduces the statistical power of the Egger test, increasing the risk of type II error (ie, failing to detect a genuinely existing small-sample effect) [30]. Second, neither the symmetry of the funnel plot nor the nonsignificance of the Egger test provides sufficient evidence to rule out publication bias [45]. Furthermore, funnel plot asymmetry itself may stem from clinical heterogeneity or differences in methodological quality rather than specifically indicating publication bias. Therefore, based on the current limited data, the findings of this systematic review cannot rule out the possibility of potential publication bias or other forms of selective reporting bias.

This systematic review aimed to systematically evaluate the impact of mHealth interventions on medication adherence among patients with cancer. The results indicate that mHealth interventions can improve medication adherence rates and mean adherence scores among patients with cancer. They also demonstrate positive effects in reducing symptom burden, enhancing self-efficacy, and increasing service satisfaction. However, mHealth showed no significant impact on health literacy. Subgroup analysis results indicate that the effectiveness of mHealth interventions is influenced by intervention timing, intervention type, and cancer type. Furthermore, variations in the intervention protocols and assessment tools used across the included studies may have impacted the reliability of the findings.

Regarding the medication adherence rate, the 95% PI following the mHealth intervention ranged from 1.46 to 8.21, indicating that the benefits of mHealth for the medication adherence rate are reproducible in future studies. In terms of the mean medication adherence score, the mHealth intervention demonstrated a significant effect, consistent with previous research findings [46]. From the perspective of specific effects and real-world effect variability, the 95% PI for the mean medication adherence score ranged from −0.50 to 2.52, encompassing zero. This indicates that while this study demonstrates a positive impact of mHealth interventions on the mean medication adherence score, the effect may not be significant in certain populations or different contexts.

Furthermore, the included studies utilized specific channels, such as mobile apps, websites, and text messaging services, to deliver knowledge and skills related to oral medication through diverse formats, including real-time online Q&A sessions, multiplayer online games, educational videos, and push notifications. These interventions also provided timely medication reminders. Both the content and methods effectively enhanced medication adherence among patients with cancer by addressing intrinsic knowledge and skill acquisition as well as extrinsic behavioral supervision. Based on this, it can be inferred that improved medication adherence contributes to optimizing the overall effectiveness of cancer treatment, such as reducing symptom burden and enhancing patient self-efficacy. Data analysis results also support this inference, showing that mHealth interventions not only significantly reduced patients’ symptom burden but also improved their self-efficacy and service satisfaction. However, the impact of mHealth interventions on health literacy was not statistically significant (P=.29), contradicting previous research findings [47]. This discrepancy may stem from the inclusion of only 2 studies in the health literacy outcome analysis, severely limiting statistical power and reducing the reliability of conclusions. Finally, this systematic review used subgroup analyses to examine the effects of intervention format, timing, and cancer type on mHealth intervention outcomes.

Mobile apps and websites offer health education information in diverse formats, including videos, audio, and text. Moreover, the information exchange between these platforms and patients with cancer—that is, between information providers and recipients—is often bidirectional and interactive. In contrast, text messaging services, which are somewhat outdated, provide information in a relatively monotonous format dominated by text, and the information transmission through this channel is predominantly one-way [48]. Theoretically, mobile apps and websites should yield superior intervention outcomes compared with text messages. However, our subgroup analysis revealed that among the 3 mHealth intervention formats, text messaging services demonstrated greater effectiveness than mobile apps and websites. The findings of this systematic review align with the research by Boima et al [49], which demonstrated that text messaging interventions were more effective than mobile apps and web-based interventions in reducing medication nonadherence events among patients with cardiovascular diseases.

The underlying reason for this discrepancy may be related to the age characteristics of patients with cancer. As shown in Table 1, the average age of patients included in this systematic review was relatively high, with 6 studies having an average age exceeding 55 years [20,32,34,38,39,44] and 2 studies having an average age exceeding 70 years [20,38]. Although text messaging services represent a simple, one-way communication method, their content is often concise, clear, and straightforward, making it easier to understand and implement. This characteristic reduces cognitive load during information reception and processing, potentially aligning better with the information format preferences and cognitive functional characteristics of older patients with cancer. Furthermore, compared with mobile apps and websites, text messaging services impose lower demands on mobile devices and network connectivity, enhancing their practical feasibility and applicability. These factors may explain why text messaging interventions outperformed mobile app and website interventions among patients with cancer, who are predominantly older adults.

This finding also raises the following questions: Have existing mHealth interventions delivered via mobile apps and websites sufficiently considered the cognitive characteristics and information preferences of older adults during their design and development? Could this oversight inadvertently contribute to and exacerbate the so-called digital divide?

This systematic review conducted subgroup analyses based on different intervention durations using 3 months as the cutoff point. The selection of this cutoff point was primarily based on 2 considerations. First, from the perspective of clinical treatment patterns, the duration of oral therapy for patients with cancer typically ranges from 12 to 24 weeks. Three months corresponds to the midpoint of the treatment course, marking a critical transition from the early to the later stages of therapy [50]. Second, prior studies indicate that medication adherence often declines around 3 months after initiation due to factors like drug side effects and forgetfulness [8]. Subgroup analysis revealed that mHealth interventions lasting <3 months significantly improved the mean medication adherence scores of patients with cancer compared with those lasting ≥3 months. The findings of this systematic review align with the study by Al-Arkee et al [51], which demonstrated that mHealth interventions lasting <3 months were more effective than those lasting ≥3 months in improving the average medication adherence scores of patients with cardiovascular diseases.

From a behavioral science perspective, the superior effectiveness of mHealth interventions lasting <3 months, compared with those lasting ≥3 months, may stem from the behavioral habituation cycle. Research by Lally et al [52] indicates that the median time required to establish a new behavior is 66 days. This suggests that the critical phase for behavioral acquisition occurs mainly within the first 2 to 3 months. During this window, mHealth interventions can effectively promote the formation and consolidation of medication-taking behaviors. They do so by providing timely reminders, feedback, and external frameworks. Once the target behavior is automated, continued external prompts may offer diminishing marginal benefits [53]. After habit formation, long-term interventions (≥3 months) typically offer few additional benefits. In some cases, they may even reduce effectiveness by reinforcing external dependence or causing intervention fatigue. In summary, mHealth interventions of <3 months match the natural process of habit formation. By maximizing gains during the most sensitive adaptation period, they significantly improve medication adherence.

This systematic review included subgroup analyses for different cancer types. Notably, the results indicate that mHealth interventions demonstrate greater efficacy in improving medication adherence among patients with breast cancer compared to patients with leukemia and those with other cancer types. However, no studies have specifically examined the impact of cancer type on the effectiveness of mHealth interventions in enhancing medication adherence. Therefore, future research should conduct in-depth analyses of the differential effects of such interventions across various cancer types.

This systematic review has several limitations. First, it included only 17 studies, and the small sample size may lead to imprecise 95% CIs and biased estimates, thereby weakening the reliability of the results. Additionally, the limited number of studies resulted in insufficient statistical power, potentially overlooking potential small-sample effects. Although the existing evidence is insufficient to prove the existence of small-sample effects, we cannot completely rule out the influence of publication bias or selective reporting bias.

Second, among the 17 included studies, only the study by Spoelstra et al [32] used a theoretical model (social cognitive theory) to guide the design of the mHealth platform. Interventions targeting medication adherence are fundamentally interventions targeting medication-taking behavior. From a behavioral science perspective, it is necessary to leverage well-established health behavior theoretical models to explore and analyze the underlying logic and influencing factors of medication adherence among patients with cancer. This approach aims to uncover the core drivers behind medication-taking behavior, thereby enabling the development of more precise and effective medication adherence intervention strategies.

Furthermore, all 17 included studies used uniform interventions for all patients. No tailored interventions were designed or implemented based on individual patient characteristics. Future design of mHealth strategies for cancer medication adherence should incorporate methods, such as latent profile analysis, user profiling, and in-depth interviews, to thoroughly analyze patient traits and needs. Integrating patient profiles into platform design will enable personalized, precision-based scientific interventions [54,55].

Building upon previous systematic reviews, this systematic review systematically reviewed and comprehensively analyzed the impact of mHealth interventions on medication adherence among patients with cancer. Overall, existing evidence suggests that mHealth interventions hold potential advantages in improving medication adherence, self-efficacy, and service satisfaction, as well as reducing symptom burden. However, considerable heterogeneity exists across certain outcome measures, and the overall evidence quality remains moderate to low. Additionally, the included studies exhibited a moderate risk of bias, necessitating cautious interpretation of these findings. It is currently inconclusive to assert superiority over conventional care.

Despite these limitations, the significance of this systematic review lies not only in confirming the efficacy of mHealth interventions for improving medication adherence among patients with cancer but also in revealing their strategic value within the context of digital health care transformation. As chronic disease management shifts from hospital-centered to patient-centered models, mHealth care offers a viable pathway for achieving continuous care, personalized support, and low-cost, scalable interventions. Future research should enhance methodological rigor by conducting multicenter, large-sample, high-quality RCTs. These studies should explore the long-term efficacy and cost-effectiveness of mHealth care across diverse health care settings and patient populations to clarify its role and value within comprehensive cancer management systems.