We performed this systematic review and meta-analysis in accordance with the PICOS (population, intervention, comparator, outcome, and study design) framework and PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses–Search) guidelines [24] (Checklist 1). The study protocol was registered with PROSPERO (CRD42024622807). All databases were searched individually on their native platforms (no multidatabase platform searching was used). Two researchers (XY and YW) independently conducted a thorough search of all relevant studies published in the following databases: PubMed, Embase, Web of Science, Cochrane Library, WanFang Database, China National Knowledge Infrastructure (CNKI), CQVIP, and Chinese Biomedical Literature (CBM) from inception to February 2026 to ensure no recent studies had been published since the initial search. No study registries (eg, ClinicalTrials.gov and World Health Organization International Clinical Trials Registry Platform) were searched. No additional online resources or print sources (eg, conference proceedings and websites) were purposefully searched or browsed. Medical subject headings terms and free words were used in combination to search the relevant studies. No published search filters were used in any of the database searches. The search strategies were developed de novo by the authors based on the PICOS framework and were not adapted from prior reviews. To maximize sensitivity for the population concept, the chronic disease component was searched using both generic terms (eg, chronic disease or illness, noncommunicable disease, and multimorbidity or comorbidity) and comprehensive controlled vocabulary or keywords covering major chronic conditions (cardiovascular diseases, cancers or neoplasms, chronic respiratory diseases, diabetes, hypertension, stroke or cerebrovascular disease, chronic kidney disease, and chronic liver disease). In addition, the language of the studies was set as English and Chinese. The references of pertinent previous systematic reviews were also screened as supplementary sources. We did not contact authors, experts, or manufacturers to seek additional studies or data. No additional search methods beyond those described were used. Following the completion of the initial draft, a supplementary literature search was conducted in February 2026 to identify any newly published studies (as noted in Deviations from the Registered Protocol). No automatic email alerts were used. The search strategies were developed by 2 reviewers (XY and YW) and reviewed by a third reviewer (XL) for completeness and accuracy; however, no formal peer review of the search (eg, using PRESS [Peer Review of Electronic Search Strategies]) was conducted. The full search strategies were shown in Tables S1 to S8 (Multimedia Appendix 1).

Eligibility criteria were set based on the PICOS framework:

A thorough screening procedure was conducted in compliance with PRISMA 2020 and 2 reviewers (XY and YW) screened the included studies independently. Among the initial 4057 records imported into EndNote X9 software, a total of 1142 duplicate publications were identified through a combination of EndNote software and manual verification. Subsequently, studies were excluded based on inclusion and exclusion criteria by reviewing titles and abstracts. Finally, full-text articles were reviewed for eligibility. Discussions or negotiations with a third reviewer (XL) were used to settle disagreements.

XY and YM extracted data from each study separately. Study authors, publication date, country, population (population and sample size), study design (description of numerical interventions, characteristics of intervention groups, and duration), theoretical framework, intervention site outcome indicators (measurement tools), and primary research outcome were extracted. The author team engaged in discussions and reexamined the original study in cases where there were disagreements over data extraction until an agreement was reached. The quality of evidence was assessed using the GRADE (Grading of Recommendations, Assessment, Development, and Evaluation) professional guidance development tool. Evidence quality was determined by evaluating five aspects: (1) risk of bias, (2) inconsistency, (3) indirectness, (4) imprecision, and (5) small-study effects [26].

This study used the risk-of-bias assessment tools recommended by the Cochrane Collaboration, evaluating studies according to their specific design types. For RCTs, the Cochrane Risk of Bias 2 (RoB 2) tool was used [27]. This tool evaluates six domains: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome data, selective reporting, and other bias. Each domain is rated as “Low risk of bias,” “High risk of bias,” or “Unclear risk of bias.” Bias in nonrandomized studies (eg, quasi-experimental designs) was appraised using the ROBINS-I tool specifically designed for such intervention studies [28]. The 7 domains of bias addressed in the ROBINS-I assessment tool include bias due to confounding, bias arising from measurement of the exposure, bias in the selection of participants into the study, bias due to post-exposure interventions, bias due to missing data, bias in measurement of the outcome, and bias in selection of the reported result. Each domain is addressed using a series of signaling questions that aim to gather relevant information about the study and analysis being assessed. Most signaling questions offer five response options: “Yes,” “Probably yes,” “Probably no,” “No,” and “No information.” Responses of “Yes” and “Probably yes” are treated equivalently in risk-of-bias judgments, as are “No” and “Probably no.” The categories for risk of bias judgments are “No information,” “Low risk,” “Moderate risk,” “Serious risk,” and “Critical risk” of bias. All assessments were conducted independently by 2 reviewers (XY and YT). Disagreements were resolved through joint discussion or consultation with a third reviewer (XL).

As the funnel plot requires at least 10 original studies, this study used regression analysis with Egger test to assess small-study effects. The results indicated that the intercept term β₁ was not statistically significant (β₁=2.22, SE=1.83; P=.22), suggesting no significant small-study effects were detected in the current analysis, as shown in Table 1. For details on testing small-study effects in quasi-experimental studies, see Table S9 (Multimedia Appendix 1).

Although most included studies used Norman’s eHealth Literacy Scale [29], some used other validated instruments to measure eHealth literacy. Given the variability in assessment tools, the standardized mean difference (SMD) and its 95% CI were calculated to enable pooling of continuous outcome data. In line with best practice, meta-analyses were conducted separately by study design (like-for-like pooling): RCTs were pooled only with RCTs, and quasi-experimental studies were pooled only with quasi-experimental studies; different designs were not combined within the same meta-analysis. Given anticipated clinical and methodological diversity, we used a random-effects model for all quantitative syntheses. The Hartung-Knapp-Sidik-Jonkman method was used to construct the random-effects model for effect estimation and pooling. This method is strongly recommended in meta-analyses as it enables more reliable uncertainty estimation while reducing the rate of false positives [30]. All analyses were conducted in Stata 17 using the meta suite with the Hartung–Knapp–Sidik–Jonkman method. To investigate potential sources of heterogeneity, we conducted pre-specified subgroup analyses based on target of intervention (specific chronic disease vs chronic disease in general) and intervention duration (≥3 mo vs <3 mo). Studies were sequentially excluded in sensitivity analyses. Egger tests were used to assess small-study effects for meta-analyses that involved more than 10 studies [31]. P<.05 was set as statistically significant.

To quantify the real-world impact of heterogeneity, we computed the 95% prediction interval (PI) for the pooled overall estimate. Due to the limited number of included studies (k=6) and substantial between-study heterogeneity, we applied the confidence distribution method proposed by Nagashima et al [32], which is recommended for small-sample meta-analyses. PIs were presented for meta-analyses with a sufficient number of studies (≥3). When the number of studies in a subgroup was small, PI estimates were interpreted cautiously.

During the implementation of this study, several adjustments were made to the protocol pre-registered on the PROSPERO platform (CRD42024622807). Following the completion of the initial draft, a supplementary literature search was conducted in February 2026, resulting in the inclusion of 15 studies [16,20,33-45] that met the eligibility criteria. Given the limited number of included studies, small-study effects were assessed using the Egger test instead of the initially planned funnel plot analysis. For data synthesis, the Hartung-Knapp-Sidik-Jonkman random-effects model was used to pool effect sizes. The literature search was restricted to publications in Chinese and English. All adjustments were documented throughout the research process and were deemed unlikely to substantially affect the validity of the study conclusions.

The screening procedure for the subsequent subphases (title or abstract screening, full text assessment) was strictly followed by the PRISMA framework to ensure transparency and reproducibility. Initially, 4057 articles were searched, including PubMed (n=652), Embase (n=110 1), Web of Science (n=955), Cochrane Library (n=875), the Wan Fang Database (n=65), the CBM (n=144), the CQVIP (n=55), and the CNKI (n=210). Using EndNote software, 1142 duplicate references were excluded. The remaining 2915 papers were screened by title and abstract. A total of 2840 papers were excluded due to noncompliance with inclusion criteria or irrelevance to the research topic. After full-text review of the remaining 84 studies, 69 were excluded for reasons detailed in the supplementary file. (Multimedia Appendix 2). The final analysis incorporated 15 eligible studies [16,20,33-45], comprising 6 RCTs [33,38-40,44,45] for the primary quantitative synthesis, 5 quasi-experimental studies [34-37,43] for supplementary quantitative synthesis, and 4 qualitative studies [16,20,41,42] were included and synthesized narratively. Figure 1 shows the comprehensive flowchart of the literature search process.

15 clinical studies [16,20,33-45] were reviewed and 2884 participants with various chronic conditions were included. Geographically, the studies were predominantly conducted in China (n=7) [33-39], Australia (n=2) [40,41], Canada (n=1) [42], Denmark (n=2) [16,20], South Korea (n=1) [43], and the United States [44,45]. All studies were published between 2016 and 2026. Two of the papers were published in Chinese, while thirteen were published in English. Regarding intervention formats, the 15 included studies predominantly used blended online-offline models (n=8) [33-36,38-40,45], followed by purely online interventions (n=4) [20,41,42,44], and face-to-face interventions (n=3) [16,37,43]. Regarding intervention content, the primary focus encompassed eHealth literacy skills training (n=6) [33,34,36-39], disease self-management and monitoring (n=5) [16,20,35,39,40], health knowledge dissemination (n=4) [16,20,35,41], and social interaction and support (n=3) [16,39,44]. Regarding intervention methodologies, approximately half of the studies were designed based on explicit theoretical frameworks (n=9) [16,33,35,37-39,42-44], with widespread adoption of personalized feedback (n=4) [16,35,40,43], interactive learning (n=5) [16,33,35,36,43], and multimedia instruction (n=5) [16,20,36,37,44], to enhance patient engagement and learning outcomes. The key features of the 15 studies are presented in Table 2. A summary of intervention implementation is provided in Table S10 (Multimedia Appendix 1).

A total of 15 studies [16,20,33-45] were included in this review. Of these, 11 studies contributed quantitative data to the meta-analyses (the primary meta-analysis pooled 6 RCTs [33,38-40,44,45], and the supplementary meta-analysis pooled 5 quasi-experimental studies [34-37,43], while the remaining 4 qualitative studies [16,20,41,42] were synthesized narratively. Importantly, risk-of-bias assessments were conducted for all included quantitative studies, regardless of whether they were ultimately pooled in the meta-analyses. Specifically, we used the Cochrane risk of bias assessment tool to evaluate the methodological quality of the 9 included RCTs [16,33,38-42,44,45,] (Figure 2A). Five studies [16,33,38,40,41] were judged to be at high risk of bias due to deviations from intended interventions or implementation issues. Four studies [16,33,38,44] had some concerns related to the randomization processor selection-bias–related information. Two studies [16,45] presented an unclear risk of reporting bias. Overall, the primary limitations of the RCT evidence centered on inadequate blinding and implementation, whereas other domains were relatively consistent, leading to an acceptable overall risk-of-bias profile. For the 6 quasi-experimental studies [20,34-37,43] included, we used the ROBINS-I tool for bias risk assessment (Figure 2B). The assessment revealed that 5 studies [20,34,35,37,43] exhibited a core issue of serious risk of confounding bias, and 2 studies [20,37] presented a high risk of bias due to missing data.

The initial quality of evidence in this study was rated as high due to the inclusion of RCTs. However, the certainty of evidence was downgraded by one level owing to substantial heterogeneity between studies that could not be fully explained. Consequently, the overall quality of evidence supporting the conclusion that mHealth interventions can improve eHealth literacy among patients with chronic diseases is moderate. The GRADE evidence level for the pooled estimate from quasi-experimental studies was, as expected, rated as low due to the very serious risk of confounding bias inherent in their design. (Multimedia Appendix 3.)

The effectiveness of eHealth literacy interventions in patients with chronic diseases was systematically assessed in this review, with effectiveness primarily assessed using RCT evidence. As explained below, specific data could be found in Table S11 (Multimedia Appendix 1). Quantitative syntheses followed a like-for-like approach by design: the primary meta-analysis pooled RCTs only (Figure 3 [33,38-40,44,45]), and a separate supplementary meta-analysis pooled quasi-experimental studies only (Figure 4 [34-37,43]).

The meta-analysis of 6 RCTs [33,38-40,44,45] showed that mHealth interventions can enhance eHealth literacy among patients with chronic diseases. The pooled average effect was positive (SMD=1.19; 95% CI 0.14-2.23; P=.03; Figure 3). The 95% CI quantifies uncertainty around this pooled average effect. Between-study heterogeneity was substantial; while I² is commonly reported (I²=97.75%, P<.001), it has limited pragmatic utility for understanding how much true effects vary across settings. Therefore, to describe the expected distribution of true effects across different populations and settings, we calculated the 95% PI, which was extremely wide (−2.68 to 5.05; Figure 3). Thus, although the average effect is beneficial, a future implementation could plausibly observe little or no benefit or even harm, or conversely, a very large benefit, depending on context and implementation. This interpretation is consistent with the risk-of-bias assessment (several trials had limitations related to blinding and implementation) and with the GRADE rating of moderate certainty for the RCT evidence, downgraded for unexplained inconsistency.

This review aimed to understand whether mHealth tools could improve eHealth literacy among patients with chronic diseases. We used the results of RCTs as primary evidence. A combined analysis of 6 RCTs showed that the intervention had a positive effect on average. This study is consistent with previous research findings [46,47]. However, there are large differences between the studies. The 95% PI was wide. The wide pPI suggests that the effectiveness of the same intervention may differ substantially across implementation contexts, with true effects potentially ranging from negligible to clinically meaningful. Due to this large uncertainty and some problems with the study (such as the lack of blinding), we rated the overall quality of evidence from RCTs as moderate using the GRADE system [48,49]. We also synthesized 5 quasi-experimental studies, and the results of the quasi-experiment supported the results of RCT with large PI.

The findings indicated that mHealth interventions are effective in enhancing eHealth literacy among patients with chronic diseases. The effect was more robust in RCTs, while quasi-experimental studies showed more modest effects accompanied by higher uncertainty. This result is generally consistent with previous reviews, which have shown that mHealth interventions are generally beneficial for health literacy-related outcomes, but the effects vary widely [50]. The observed differences may stem from heterogeneity in patient characteristics, intervention design, and outcome measures [51]. Lyles et al [45], in a study of underserved populations, found that portal use among individuals with limited health literacy was substantially lower than among those with adequate health literacy. Nahm et al [44] reported more favorable outcomes in older adults with higher educational attainment, suggesting that socioeconomic status and access to resources are critical determinants of intervention effectiveness. Intervention intensity and theoretical grounding also appear to be pivotal [52,53]. Gao et al [39] demonstrated that an intervention based on the self-regulated learning model was more effective than one relying on a single theoretical framework. Cheng et al [38] showed that experiential learning significantly improved eHealth literacy. By contrast, the findings of Parker et al [40] did not achieve meaningful improvements in clinical endpoints, likely owing to insufficient intervention intensity and limited tailoring. Future research should prioritize the development of stratified, adaptive, and theory-informed multimodal interventions [54]. In particular, greater attention should be directed toward designing and validating more personalized and supportive strategies for vulnerable populations with limited health literacy and inadequate social resources [55,56].

This study conducted subgroup analyses by intervention target. The results showed that interventions designed for one specific disease (eg, diabetes alone or heart failure alone) appeared to work better than interventions designed for patients with all chronic conditions. This finding is consistent with previous research [57]. Interventions on patients with specific diseases showed greater effects on average. mHealth intervention programs targeting the general chronic disease population have smaller average effects, and their PIs suggest that they may not always be beneficial. When an intervention focuses on a specific disease, its content, tasks, and feedback closely match what patients need to do every day [58]. This may make the tool more useful to patients and keep them engaged [59]. In the field of mHealth, higher engagement generally leads to better outcomes [60,61]. For example, apps that help manage diabetes or blood pressure often include self-management tasks and provide direct feedback, which may explain why they work so well [62]. In contrast, trying to cover multiple conditions can make it difficult to provide in-depth, personalized support for each condition [51]. Recent research on tools for patients with multiple conditions showed that combining several disease-specific tools could feel burdensome and did not always provide useful, actionable recommendations [63,64]. Therefore, it may be wiser to start with small, sophisticated tools for a single disease and consider expanding to more diseases later.

We further compared studies with an intervention duration of less than 3 months versus an intervention duration of more than 3 months. Subgroup analysis showed a statistically significant combined effect of the less than 3-month intervention, which is generally consistent with the findings of Sevda and Mechelle [65,66]. However, the PI remained notably wide, suggesting considerable uncertainty in the effectiveness of the intervention when applied across diverse implementation contexts. In contrast, the subgroup with intervention for more than 3 months had larger point estimates, but its combined effect did not reach statistical significance, and the PI was extremely wide. Given that the test for subgroup differences did not reach statistical significance, it is unwarranted to conclude that long-term interventions are inherently less effective than short-term ones. A more plausible interpretation is that the effectiveness of longer-duration mHealth interventions may be more susceptible to contextual factors and the fidelity of implementation processes [67]. The possible reason is that the benefits of mHealth interventions depend not only on the duration of the intervention itself but also on whether patient engagement can be sustained over time [68]. Existing evidence suggests that user engagement in mHealth interventions often declines as time progresses, while higher levels of engagement are generally associated with better adherence and improved self-management outcomes [52]. For longer-duration interventions, factors such as treatment burden, reduced interaction, or implementation barriers may weaken sustained patient participation [69]. Therefore, future research should not only focus on extending the duration of interventions but also prioritize optimizing intervention content, feedback frequency, personalized support, and engagement strategies to enhance long-term adherence and participation, thereby maximizing the long-term effectiveness of mHealth interventions.

Four major limitations of this study should be mentioned. First, the restriction to English and Chinese language publications may introduce selection bias and limit the cross-cultural applicability of our findings. Second, there was considerable methodological heterogeneity among the studies included, particularly in terms of intervention protocols and outcome measures. Third, this study acknowledges certain limitations in the current analysis, primarily reflected in the relatively small sample size. Fourth, this study used eHealth literacy scores as the primary indicator of intervention effectiveness. While this metric directly reflects improvements in digital health capabilities, it struggles to comprehensively capture the intervention’s overall benefits regarding clinical symptoms, quality of life, and health care costs. Future research should adopt multidimensional indicators to conduct more comprehensive assessments.

Overall, current evidence suggests that mHealth interventions can improve eHealth literacy among patients with chronic diseases on average. However, the wide PIs indicate substantial between-study variability, suggesting that intervention effects are highly context-dependent and closely related to implementation factors. By focusing on eHealth literacy as a core capability outcome and integrating evidence from RCTs, quasi-experimental studies, and qualitative research, this review extends previous work in the field. These findings highlight the need for more rigorously designed and better-implemented mHealth interventions in future research.