Dementia is a progressive neurodegenerative disorder and a major contributor to disability and dependency among older adults worldwide, representing a major and growing global public health priority [1]. Recent meta-analyses have demonstrated the substantial burden experienced by dementia caregivers, particularly in relation to psychological distress and care demands [2]. In parallel, global epidemiological evidence indicates that the prevalence and societal burden of dementia are rapidly increasing, placing escalating pressure on health and long-term care systems worldwide [3]. Estimates from the China Alzheimer Report 2025 indicate that the prevalence has reached 13 million, representing a substantial and rapidly escalating public health and socioeconomic challenge, particularly in aging societies where long-term care systems remain underdeveloped [4]. Beyond the clinical burden of dementia itself, the associated caregiving demands represent a critical yet often underrecognized component of the overall disease burden. Importantly, the burden associated with informal caregiving has been increasingly recognized as a parallel public health challenge, with substantial implications for both caregiver health and health system sustainability [2]. Despite growing recognition, effective and scalable solutions to support caregivers remain limited.

Most people living with dementia receive care in home and community settings. In China, cultural norms emphasizing family responsibility, combined with the high cost and limited availability of institutional care, mean that informal caregivers, typically spouses or adult children, provide the majority of long-term care [5,6]. Recent large-scale randomized controlled trials suggest that existing dementia care models may have limited impact on key caregiver outcomes such as burden, despite improvements in care satisfaction [7]. These caregivers frequently manage progressive cognitive decline, neuropsychiatric symptoms, and functional deterioration, often without formal training or adequate support [8]. Informal caregivers are often described as the “second invisible patients” of the dementia care system [9]. Caregiver burden is a multidimensional construct encompassing psychological distress, physical strain, social restriction, and financial pressure. Elevated caregiver burden is associated with adverse outcomes including depressive symptoms, reduced quality of life, and increased health care utilization. In addition to objective caregiving demands, psychosocial factors, such as coping resources and access to support services, also influence caregiver outcomes [10,11].

eHealth is defined by the World Health Organization as the use of information and communication technologies for health and health-related fields, including health care services, health education, and research [12]. Building on this broad conceptualization, eHealth interventions, including web-based programs, mobile apps, and telehealth services, offer advantages in terms of accessibility, flexibility, and scalability compared with traditional face-to-face interventions [13-15]. Recent systematic reviews suggest that such interventions can improve caregiver outcomes, although the magnitude and consistency of effects vary across intervention types and delivery formats [16]. However, despite recognition of caregiver burden, effective and scalable intervention strategies remain insufficient.

Although recent reviews have examined the effectiveness of eHealth interventions for dementia caregivers, findings remain inconsistent. Reported effects on caregiver burden range from moderate reductions to nonsignificant outcomes [16-19]. These discrepancies may reflect heterogeneity in study populations, variation in intervention components, and the inclusion of nonrandomized designs. Furthermore, many meta-analyses have relied on conventional random-effects models, which may underestimate uncertainty when the number of included studies is limited. More conservative methods, such as the Hartung-Knapp-Sidik-Jonkman (HKSJ) approach, are recommended in such contexts to reduce the risk of inflated type I errors [20,21]. In addition, the relative contribution of specific intervention characteristics, such as delivery modality, intervention duration, and the inclusion of human-supported guidance, remains unclear. Elucidating these factors is essential to optimize the design and implementation of scalable digital caregiver support programs, particularly in the context of the rapid expansion of digital health services.

To address these gaps, this study focuses exclusively on randomized controlled trials (RCTs) and applies more conservative statistical approaches, including the HKSJ method, to provide more robust and reliable estimates. Furthermore, this study systematically examines key intervention characteristics, including delivery modality, intervention duration, and guidance level, to identify features associated with improved effectiveness. Compared with previous reviews that included heterogeneous study designs or did not explore intervention-level moderators, this study aims to provide a more rigorous and mechanism-oriented synthesis. By identifying the specific components that drive effectiveness, this work offers practical guidance for the design, implementation, and scaling of eHealth interventions for dementia caregivers.

This review aimed to evaluate the effectiveness of eHealth interventions in reducing caregiver burden among informal caregivers of individuals with dementia. The secondary objective was to assess the effects of these interventions on caregivers’ depressive symptoms. An additional aim was to examine potential moderators of intervention effectiveness through subgroup analyses and meta-regression, including intervention delivery modality, intervention duration, and the presence of human-supported guidance. Understanding these factors may inform the development of more targeted and scalable digital interventions for dementia caregiver support.

This systematic review and meta-analysis was conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement [22], PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses—Search extension) recommendations [23], and the Cochrane Handbook for Systematic Reviews of Interventions. The study protocol was prospectively registered with PROSPERO (ID: CRD42024625273).

A comprehensive search was performed across 8 electronic databases: PubMed (via NCBI), Embase and Scopus (via Elsevier), Web of Science Core Collection (via Clarivate), CENTRAL (Cochrane Central Register of Controlled Trials), ProQuest Dissertations & Theses Global, and CINAHL and PsycINFO (via EBSCOhost). In addition, ClinicalTrials.gov was searched to identify ongoing or unpublished studies. Reference lists of included studies were manually screened to identify additional relevant studies. All databases were searched from inception to March 10, 2026. Reference lists of included studies and relevant systematic reviews were manually screened to identify additional eligible studies. No additional sources such as conference proceedings, organizational websites, or gray literature databases were searched, as the database strategy was considered sufficient to capture relevant peer-reviewed evidence. No contact was made with study authors, experts, or manufacturers for additional or unpublished data.

The literature search was conducted and reported in accordance with the PRISMA-S extension to ensure transparency and reproducibility [23]. The search strategy combined controlled vocabulary terms (eg, MeSH [Medical Subject Headings] and Emtree) and free-text keywords related to dementia, caregivers, and eHealth or mHealth interventions. The strategy was developed de novo by the research team and iteratively refined. The search was limited to English-language studies due to resource constraints for translation and to RCTs to ensure methodological rigor. No date restrictions were applied. No validated external search filters were used; instead, study design restrictions were implemented through a combination of indexing terms and keywords. The full search strategies for all databases are provided in Table S1 in Multimedia Appendix 1. Any deviations from PRISMA-S checklist items due to methodological constraints are explicitly reported in the manuscript.

Eligibility criteria were determined according to the PICOS (Population, Intervention, Comparison, Outcomes, and Study Design) framework. (1) Population (P): studies were included if they involved informal caregivers (eg, family members, spouses, or friends) providing unpaid care for individuals with a clinical diagnosis of dementia. (2) Intervention (I): were defined as eHealth-related interventions (eg, web-based platforms, mobile apps, or telehealth programs) targeting informal caregivers of individuals with dementia. (3) Comparison (C): eligible comparators included usual care, waitlist controls, or active nondigital interventions (eg, printed educational materials). (4) Outcomes (O): the primary outcome was caregiver burden, assessed using validated standardized instruments such as the Zarit Burden Interview. To avoid selective reporting bias, all studies were included providing extractable data for burden, regardless of its status as a primary or secondary outcome in the original trial. If a study reported multiple burden scales, the Zarit Burden Interview was prioritized for pooling to enhance consistency. Secondary outcomes included depressive symptoms measured by validated scales such as the Center for Epidemiological Studies Depression Scale or equivalent, following the same inclusion as the primary outcome. (5) Study design (S): peer-reviewed RCTs published in English.

Studies were excluded if they (1) focused solely on formal or professional health care providers, (2) provided only passive web-based informational content without interactive components, or (3) were conference abstracts, editorials, or protocols without extractable outcome data.

All retrieved records were imported into EndNote (version 21; Clarivate Analytics) for deduplication using both automated and manual processes. Title and abstract were screened independently by 2 reviewers (ML and JXR). Screening was conducted in a 2-stage process to manage the high volume of retrieved records (n=7449). In the first stage, 2 reviewers (ML and JXR) independently performed a preliminary triage based on titles to remove clearly irrelevant records (eg, animal studies and unrelated medical conditions). To ensure a highly sensitive and overinclusive process, records were excluded at this stage only if both reviewers independently deemed them unequivocally unrelated to the research objective. In the second stage, all remaining records underwent a formal title-and-abstract screening process, in which both title and abstract information was considered, conducted independently by 2 reviewers (ML and JXR) in accordance with Cochrane recommendations [24]. This stage aimed to ensure an overinclusive and sensitive selection process. Any record considered potentially relevant by either reviewer was retained for full-text assessment. Full-text papers of potentially eligible studies were then assessed for inclusion. Discrepancies were resolved through discussion or consultation with a third reviewer (XYW).

Data extraction was conducted independently by 2 researchers (ML and JXR) using a standardized extraction form. Disagreements were resolved through discussion with a third reviewer (XYW). Extracted data included (1) study characteristics (author, year, and country), (2) participant demographics (age, gender, and relationship to care recipient), (3) intervention characteristics (modality, duration, guidance, and theoretical framework), and (4) outcome data (sample size, means, and SDs). For studies with multiple intervention arms, the control group was proportionally divided to prevent double-counting.

The risk of bias for each included RCT was assessed using the Cochrane Risk of Bias 2 (RoB 2) tool. Five domains were evaluated: bias arising from the randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. Two reviewers (ML and JXR) independently assigned a rating of “Low risk,” “Some concerns,” or “High risk” to each domain. Two reviewers (ML and JXR) conducted assessments independently, with disagreements resolved through consensus.

The systematic search across 8 electronic databases and registers identified 7444 records. After removing duplicates and records marked as ineligible by automation tools, 2846 records remained for screening. Following title and abstract screening, 113 reports were sought for retrieval, of which 7 could not be retrieved. A total of 106 reports from databases and registers were assessed for eligibility. In addition, 5 reports were identified through citation searching and assessed separately for eligibility. After excluding 75 reports from database and register searches and 1 report identified through citation searching, 35 RCTs were included in the review. The study selection process is detailed in the PRISMA 2020 flow diagram (Figure 1).

The 35 included RCTs, published between 2005 and 2026, were conducted across North America, Asia, Europe, and Oceania [27-61] (Table 1; details in Table S2 in Multimedia Appendix 1 [27-61]). The trials collectively enrolled 3388 informal caregivers (intervention: n=1700; control: n=1688). Caregivers were predominantly female (mean proportion 73.06%, SD 14.73%) and were generally middle-aged to older adults, with reported mean or median ages ranging from approximately 30 to 73 years. Most participants were spouses or adult children of individuals with dementia. Interventions varied in format and duration. Early studies primarily used web-based platforms, whereas more recent trials increasingly adopted mobile or hybrid digital delivery models. Intervention duration ranged from 2 to 26 weeks. Approximately 85.71% (30/35) of interventions were explicitly grounded in established theoretical frameworks, such as the Stress and Coping Model or the World Health Organization’s iSupport program. Detailed study characteristics are presented in Table 1 and Table S2 in Multimedia Appendix 1.

Risk of bias was assessed using the RoB 2 tool (Figure 2 [27-61]). Three (8.6%) studies were rated as low risk, 24 (68.6%) studies had some concerns, and 8 (22.9%) studies were judged to have high risk of bias. The most common sources of bias were related to deviations from intended interventions and outcome measurement. These issues primarily reflected the inherent difficulty of blinding participants in behavioral and digital interventions and the reliance on self-reported outcome measures. Most studies showed low risk of bias for randomization and selective outcome reporting.

Univariate and multivariate meta-regression analyses were conducted examining caregiver age, proportion of female participants, and attrition rates (Table 4).

For the primary outcome (Caregiver Burden; k=35), caregiver age emerged as a significant moderator (β=.02, 95% CI 0.0008 to 0.0392; Z=2.044, P=.049; Figure 5). In contrast, neither the proportion of female participants (P=.44) nor the study attrition rate (P=.36) significantly influenced the treatment effect. In the multivariate model, age showed borderline significance (β=.0209, 95% CI −0.0009 to 0.0427; P=.07), and the intercept remained significant (β=−1.542, 95% CI −2.7495 to −0.3347; P=.02).

For the secondary outcome (Depressive Symptoms) (k=23), the meta-regression results were consistent with the burden analysis in terms of direction but did not reach statistical significance. The intervention effect on depression was not significantly moderated by caregiver age (β=.0162, 95% CI −0.0199 to 0.0523; P=.39), female percentage (P=.83), or attrition rate (P=.22). The multivariate model confirmed that these factors collectively did not explain the substantial residual heterogeneity.

Egger tests indicated no significant funnel plot asymmetry for caregiver burden (t₃₃=−0.808; P=.42) or depressive symptoms (t₂₁=−0.954; P=.35). Contour-enhanced funnel plots (Figure 6A for caregiver burden and Figure 6B for depressive symptoms) confirmed symmetric distributions, suggesting that observed heterogeneity is attributable to study characteristics rather than publication bias.

Sensitivity analyses were conducted by focusing on 27 high-quality studies (N=2885) for the primary outcome and 19 studies (N=2059) for the secondary outcome, after excluding those identified with a high risk of bias. For the primary outcome (Caregiver Burden), the pooled effect for caregiver burden remained statistically significant and increased in magnitude (SMD −0.31, 95% CI −0.50 to −0.13; t₂₆=−3.47; P=.002; 95% PI −1.28 to 0.59; τ²=0.169, τ=0.412) (Table S3 in Multimedia Appendix 1). Heterogeneity remained substantial (I²=77.0%; Q=113.19; P<.001). This PI indicates that while the average effect is beneficial, the outcome in a specific clinical setting could vary.

Regarding the secondary outcome (Depressive Symptoms), the pooled effect yielded a borderline significant reduction (SMD −0.31, 95% CI −0.63 to 0.01; t₁₈=−2.06; P=.05; 95% PI −1.63 to 1.01; τ²=0.371, τ=0.609). While heterogeneity remained high (I²=88.1%, P<.001), the 95% PI and the borderline significant P value support that eHealth interventions provide a nonsignificant trend for depressive symptoms. Collectively, these results derived from the conservative HKSJ model confirmed that the observed clinical benefits of eHealth interventions are not driven by low-quality studies and remain robust even under the most rigorous inclusion criteria.

Using the GRADE approach, the certainty for both caregiver burden and depressive symptoms was rated as moderate (Table 5). Evidence was downgraded 1 level due to risk of bias associated with lack of participant blinding and reliance on self-reported outcomes. Despite high heterogeneity, consistent effect direction across studies supported a moderate rating.

This systematic review and meta-analysis of 35 RCTs suggests that eHealth interventions are associated with a small but statistically significant reduction in caregiver burden and depressive symptoms among informal caregivers of people with dementia. Although pooled estimates favor eHealth interventions, the 95% PI indicates that the effect in a future specific clinical context could range from a substantial benefit to no effect. This distinction suggests that while eHealth interventions are validated at the population level, their effectiveness in real-world settings is highly context-dependent. The moderate certainty of this evidence, assessed by GRADE, reflects consistent effects direction alongside recognized limitations of psychosocial trials, particularly the inability to blind participants and outcome assessors.

A key observation is that eHealth interventions incorporating human support or active monitoring tend to yield larger effect sizes than fully self-guided approaches. Although subgroup analyses were not statistically significant, consistent within-group patterns suggest that human involvement (eg, professional feedback, coaching, or monitoring) may enhance engagement and adherence [62,63]. The magnitude of effect observed in this review exceeds the negligible effects reported in some prior meta-analyses (eg, SMD≈−0.06) [19], yet remains smaller than those reported for structured, education-based interventions (eg, g=−0.45) [2], and is consistent with recent syntheses of home-based digital interventions reporting small-to-moderate but heterogeneous effects [64]. The wider literature, however, remains inconsistent, with some studies reporting moderate short-term benefits and others demonstrating small or nonsignificant effects, particularly for caregiver burden and depressive outcomes [17-19,21]. This heterogeneity likely reflects variation in intervention intensity, degree of human support, and methodological differences across trials, including attrition, outcome measurement, and duration of follow-up. Importantly, while confidence intervals represent the average effect, the wide PIs underscore substantial variability in outcomes across contexts. Collectively, these findings indicate that eHealth interventions yield modest and heterogeneous effects, driven more by intervention characteristics than by uniformly strong efficacy.

Subgroup analyses further suggest that short-term (≤8 weeks) and mobile-based interventions tend to demonstrate relatively stronger effects on caregiver burden reduction. While consistent with prior evidence that time-limited digital interventions can improve proximal psychosocial outcomes, effects on caregiver burden remain inconsistent [19]. Compared with earlier reports of larger effects (eg, d=−0.65) [17], and more recent estimates showing smaller effects (eg, SMD=−0.21) [18], the pooled estimates in this review are comparatively conservative, potentially reflecting methodological differences, including the use of the HKSJ adjustment. Although shorter-duration (≤2 months) and mobile- or web-based interventions appeared more likely to yield statistically significant reductions in caregiver burden, these findings were not robust under substantial heterogeneity (I²≈75%‐92%) and variations in study design [17,18,65]. Overall, effects on caregiver burden remain small or nonsignificant, particularly for long-term outcomes and quality of life. One plausible explanation is that short-term or mobile-based interventions primarily influence proximal outcomes, while failing to produce sustained changes in more complex outcomes such as caregiver burden and quality of life, potentially due to limited intensity or insufficient personalization [66]. Taken together, these findings suggest that intervention effectiveness is determined more by design features, outcome selection, and intensity than by delivery modality alone [16,67].

Meta-regression analyses demonstrated a potential moderating effect of caregiver age, with reduced effectiveness observed in older populations in univariate models. However, these findings suggest that intervention effectiveness is determined more by design features, outcome selection, and intensity than by delivery modality alone. This pattern aligns with existing evidence suggesting that digital interventions for caregivers generally yield small or nonsignificant effects and limited improvements in quality of life [66], and that benefits may not be uniformly distributed across subgroups. Age-related differences may be attributable to broader digital divide factors, including variability in digital literacy, usability challenges, and technology-related self-efficacy [68]. Evidence from gerontechnology and digital health research indicates that adoption and sustained engagement are strongly influenced by usability and contextual fit, with digital literacy demands and cognitive load representing persistent barriers [69,70]. In addition, broader syntheses of digital therapeutics adoption identify mistrust, usability constraints, and insufficiently tailored onboarding processes as key barriers among older populations [71]. Accordingly, differential effectiveness by age is more plausibly explained by variation in engagement and usability rather than by inherent differences in treatment response.

Overall, the findings of this review are consistent with recent umbrella and meta-analytic evidence suggesting that intervention heterogeneity, rather than delivery modality per se, is the principal driver of variability in effect estimates across studies [17,18,21]. This reinforces the interpretation that effectiveness is not determined solely by the presence of digital components but by how interventions are designed, implemented, and aligned with user needs. Accordingly, standardized, one-size-fits-all digital approaches are unlikely to achieve optimal outcomes. Future interventions should instead prioritize adaptability and personalization, alongside the deliberate integration of human support elements (eg, guidance, feedback, or monitoring), to enhance user engagement, adherence, and overall effectiveness.

Several limitations should be considered when interpreting these findings. First, the wide PI indicates substantial between-study variability, suggesting that intervention effects may differ considerably across real-world settings. Second, as common in psychosocial and digital health research, most studies were rated as having “some concerns” or “high risk” of bias, primarily due to the inability to blind participants (RoB 2 Domain 2) and reliance on self-reported outcomes (Domain 4). Third, high attrition rates in several pilot trials may have affected internal validity [35,38], although meta-regression did not identify attrition as a significant moderator. Fourth, while subgroup analyses identified potentially important moderators (eg, intervention duration, delivery modality, and level of human support), these findings should be interpreted cautiously due to limited statistical power and persistent heterogeneity. Furthermore, although the screening process was rigorous, the preliminary title-based triage may have introduced a minimal risk of inadvertently excluding relevant studies. Finally, this review was restricted to English-language peer-reviewed publications, potentially excluding relevant gray literature or region-specific implementations.

This review extends the existing literature in several important ways. Methodologically, the application of the HKSJ approach, alongside the reporting of PIs, provides more conservative and clinically informative estimates of intervention effects. Analytically, it moves beyond evaluating overall effectiveness by exploring potential moderators, offering insights into the conditions under which eHealth interventions may be more effective. Although subgroup differences were not statistically significant, the observed patterns highlight the potential importance of intervention characteristics such as human support and duration [10].

From a practical perspective, these findings have important implications for clinical practice, digital health development, and health policy. Clinically, guided eHealth interventions could be integrated into dementia care pathways as scalable support tools, particularly during high-stress periods. For digital health developers, platforms should be grounded in evidence-based psychological frameworks [17], and our results underscore the potential value of embedding human-supported components, such as coaching, feedback, or monitoring, within digital platforms, rather than relying on fully self-guided designs. At the health system level, the evidence points toward the potential of a hybrid care model in which digital infrastructure is complemented by training human facilitators or “digital coaches,” whose involvement appears linked to intervention effectiveness [62,63]. Usability and embedded support appear more critical than baseline digital literacy for successful adoption [71].

This review provides timely evidence that eHealth interventions are associated with modest reductions in caregiver burden and improvements in depressive symptoms among informal caregivers of people with dementia. By applying the HKSJ framework with PIs, this study offers more robust and practice-relevant estimates than prior reviews. Interpretation of these findings requires integrating multiple dimensions of evidence, as statistically significant average effects based on 95% CIs coexist with wide 95% PIs, indicating substantial heterogeneity and suggesting that real-world effectiveness may range from meaningful benefit to minimal impact. This variability, together with the identified risk of bias, particularly the lack of participant blinding, and an overall moderate level of certainty based on GRADE, indicates that intervention effectiveness is context-dependent and influenced by both methodological and implementation factors. Subgroup analyses indicate potential benefits of human-supported and shorter-duration interventions, although substantial heterogeneity persists. Future research should focus on optimizing hybrid digital-human intervention models, improving accessibility across diverse populations, and generating high-quality evidence with longer follow-up to enhance consistency and real-world applicability.

These include the following:

These include the following: