This systematic review and meta-analysis report adheres to the 2020 PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) statement (Multimedia Appendix 1) and PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension) to enhance transparency (Multimedia Appendix 2) [58,59]. The review protocol has been registered with PROSPERO (International Prospective Register of Systematic Reviews; CRD420251233352).

We conducted a search of PubMed, Web of Science, the Cochrane Library, Embase, CINAHL, Scopus, IEEE Xplore, and gray literature by October 2025. The search was updated in April 2026 to identify any newly published relevant studies. First, the search strategy and search terms were finalized through group discussion. Subsequently, preliminary database searches were performed, and the strategy and terms were further revised based on the retrieved results. Formal searches were then conducted in each database using the finalized search strategy and keywords. In addition, we manually reviewed the reference lists of included studies to identify any potentially omitted studies. The search strategy combined subject headings and free-text terms. The main subject terms included breastfeeding, AI, SMS text messaging, wearable sensors, telemedicine, mobile apps, internet-based interventions, RCTs, and non-RCTs. This review did not use published search filters. We did not adopt or reuse search strategies from previously published systematic reviews during our literature search. We manually screened lists of eligible studies and cited references via Google Scholar to identify additional relevant literature. We searched the US clinical trials registry (ClinicalTrials.gov) and the Chinese clinical trials registry to identify eligible studies. For studies where full-text articles and data were unavailable, we contacted the corresponding authors to attempt to obtain the full text. Given the availability of academic resources and the language proficiency of the research team, only studies published in English were included, with no publication date restrictions. The complete search strategy is available in Multimedia Appendix 3.

We included the following studies:

The following studies were excluded: studies involving mothers with infectious diseases (eg, HIV) or severe substance use disorders, studies published in languages other than English, studies for which full-text articles were unavailable, studies with missing or unavailable data, animal studies, letters or conference proceedings, and studies rated as “high risk” [65].

All retrieved articles were imported into EndNote software (Clarivate Analytics). After removing duplicates, two researchers (JS and YW) independently reviewed the article titles and abstracts and subsequently read the full texts to determine eligibility based on the inclusion criteria. Disagreements were resolved through discussion with a third researcher (SH), who made the final decision.

Two researchers (JS and YW) independently extracted data using a customized data extraction form. The details extracted included but were not limited to study details (eg, title, authors, year, and country), study design (study type, objective, measurement tools, and inclusion/exclusion criteria), participant characteristics (sample size, age, gestational age, number of fetuses, mode of delivery, etc), breastfeeding outcomes (EBF rates, breastfeeding self-efficacy, and breastfeeding knowledge), and intervention characteristics and dose (intervention type, period, frequency, and duration).

There were no issues of double counting in this meta-analysis, as the participants in each included study constituted independent samples [66]. Therefore, for studies with multiple intervention groups, we did not combine the interventions; instead, we extracted data only from the groups directly relevant to the study objectives for analysis. For studies reporting the same outcome at multiple time points, to ensure consistency of results and minimize the potential influence of time on the findings, we included only data from the first time point following intervention in the meta-analysis [66].

Quality assessment was conducted during the full-text screening phase to inform study eligibility. Two researchers (JS and YW) independently assessed the quality of RCTs using the Cochrane Risk-of-Bias Tool version 2 (RoB2) [63,67], which evaluates random sequence generation, allocation concealment, participant and staff blinding, outcome assessor blinding, missing outcome data, selective reporting, and other sources of bias. We used the Risk of Bias in Nonrandomized Studies of Interventions (ROBINS-I) tool to assess the quality of quasi-experimental studies [68]. Based on the overall assessment results of ROBINS-I, the risk level for each study was categorized as “critical risk, serious risk, moderate risk, low risk, or insufficient information to assess risk of bias.” During the screening phase, we first assessed the methodological quality of included studies using the RoB2 [67,69] and ROBINS-I [68]. Studies with at least one item or an overall bias risk rated as “high” were excluded. To minimize the systematic impact of biases such as selective reporting, inadequate blinding, and incomplete data on effect estimates, which could compromise the robustness of the pooled results, these studies were excluded from the primary analysis. Risk of bias results for the included studies were visualized using Robvis [70].

Two authors (JS and YW) independently scored the certainty of evidence using the GRADE (Grading of Recommendations, Assessment, Development, and Evaluation) criteria [71,72]. Disagreements were resolved by consulting a third researcher (SH).

We conducted the meta-analysis using Review Manager (version 5.4; Cochrane Collaboration) and R software (version 4.6.0; R Foundation for Statistical Computing). The restricted maximum likelihood estimator was used to estimate the between‑study variance (τ²) [73]. The Hartung-Knapp-Sidik-Jonkman random-effects model was used to reanalyze the data and estimate CIs [74]. EBF rates were estimated using relative risk (RR) and 95% CI, with data pooled across three time periods: <3 months, 3-6 months, and ≥6 months. Breastfeeding self-efficacy and breastfeeding knowledge levels were estimated using the standardized mean difference (SMD) and 95% CI. Due to significant differences in participant characteristics, intervention methods, and outcome measurement methods, the data were analyzed using a random-effects model. To assess statistical heterogeneity, τ² was calculated. We also derived Cochran’s Q test and the I² statistic to jointly quantify the extent and significance of observed heterogeneity. For results with a sufficient number of studies (n≥3), we calculated a 95% prediction interval (PI) to quantify the practical clinical significance of heterogeneity [75]. Sensitivity analyses were conducted by sequentially excluding all included studies. Potential small-sample effects were assessed using funnel plots combined with the Egger regression test [76].

Subgroup analyses were conducted to explore potential sources of heterogeneity. Prespecified subgroup analyses were conducted to identify potential sources of between-study heterogeneity and examine whether intervention characteristics moderate the effects of DHIs on breastfeeding-related outcomes. Prespecified subgroup analyses were performed stratified by (1) DHIs type (mHealth: SMS text messaging and apps; eHealth: internet-based, websites, and videos; or digital devices: wearables and VR) [77]; (2) intervention stage (prenatal vs postnatal); (3) intervention duration (<3 months vs ≥3 months); (4) intervention frequency (<3 times/week vs ≥3 times/week); in addition, to further investigate the substantial heterogeneity observed in the pooled analyses, we conducted supplementary exploratory subgroup analyses stratified by (5) national income level (developed countries vs developing countries, with grouping standards following the official country income classification criteria released by the World Bank); and (6) population characteristics: gestational age (term vs preterm), age (adults vs adolescents), number of fetuses (singleton vs multiple), mode of delivery (vaginal vs cesarean). All findings from exploratory subgroup analyses should be interpreted with caution and regarded only as exploratory evidence to generate new research hypotheses.

This meta-analysis was a secondary analysis of published literature and did not require ethics approval. As there is no direct contact with subjects and no individual-level data are collected, informed consent is waived. This analysis strictly adheres to the principles of patient privacy protection: all data used are aggregate statistics or fully anonymized data that contain no personally identifiable information (eg, names, initials, and hospital numbers), such information having been omitted as it is not essential for scientific purposes. All original studies included in this meta-analysis obtained their own ethical approval and informed consent from subjects within their respective research contexts.

Initially, 2353 articles were identified from the databases. After excluding 508 duplicate articles, 2 researchers independently screened the titles and abstracts of the remaining 1845 articles. Of these, 1684 articles were excluded for irrelevance. Full texts of the remaining 161 articles were retrieved and assessed for eligibility. Studies rated as high risk of bias were excluded during this stage. Ultimately, 52 articles were included in the meta-analysis [34,35,42,43,49,56,57,78-122]. The literature screening process is shown in Figure 1.

The 52 included studies were published between 2015 and 2026 [34,35,42,43,49,56,57,78-122], with 15 conducted in Turkey [34,35,49,78,81,83,84,86,87,97-99,101,108,121], 8 in China [43,56,90,94-96,119,120], 6 each in the United States [79,85,92,93,100,114] and Iran [82,88,111,113,115,122], 2 each in India [91,109] and Saudi Arabia [80,110], and 1 each in Nigeria [116], Japan [106], Malaysia [103], Colombia [104], Spain [104], Kenya [107], Uganda [105], Vietnam [89], Myanmar [57], Australia [118], Israel [102], Ethiopia [42], and Indonesia [112]. This study included 42 RCTs [34,35,42,43,49,56,57,78,79,81,83-85,87, 89-95,97,99-105,107,109,110,113-122] and 10 quasi-experimental studies [80,82,86,88,96,98,106,108,111,112]. Among these, 3 studies recruited adolescent mothers [34,87,116]; 3 studies included mothers of critically preterm infants [100,110,111]; 3 studies included women who underwent cesarean section [89,97,121]; and 2 studies included women with multiple pregnancies [103,109]. The total sample size for this study was 11,704, with individual trial sample sizes ranging from 23 to 1717 participants. Interventions lasted from 1 day to 9 months, with a frequency ranging from once a week to 10 times a week.

Among the 52 included studies, DHIs types could be categorized into 3 groups: mHealth (n=30) [42,43,56,57,78,81-83,85,89,90,92-96, 100,102,103,105,107,109,112-118,121], eHealth (n=17) [34,35,79,80,84,86-88,91,97-99,101,106,108,110,119], and digital devices (n=5) [49,104,111,120,122]. A total of 30 studies primarily implemented interventions during the prenatal period [34,35,42,43,49,56,57,80,81,84,85,89,91,92,94-96,98,101,103-106, 108,112,114,117-120], while the remaining 22 initiated interventions postpartum [78,79,82,83,86-88,90,93,97, 99,100,102,107,109-111,113,115,116,121,122]. Two studies were 3-arm RCTs [85,118]. For studies with multiple intervention groups, we extracted data only from the groups directly relevant to the study objectives for analysis. To ensure consistency of results and minimize the potential impact of time on the findings, we adopted the methodological protocol used in a previous meta-analysis [123]. The outcomes we examined included EBF rates (n=37) [34,42,43,56,57,78-85,87,89-97, 99,100,102,104-106,109,110,112,114,117-120], breastfeeding self-efficacy (n=28) [34,35,49,81,83,85-88,90,91, 93,97,98,101,103,104,106,108-111,113-115,119,121,122], and breastfeeding knowledge (n=30) [86,94,103,107,115,116]. Although the outcome assessment tools varied across studies, all were valid scales, and data collection was successfully conducted. A total of 24 studies used the Brief Breastfeeding Self-Efficacy Scale-Short Form [34,35,81,83,86-88,90,91,93, 97,98,103,104,106,108-111,113,114,119,121,122]. Two studies used the full Breastfeeding Self-Efficacy Scale [85,115], 2 used the Prenatal Breastfeeding Self-Efficacy Scale [49,101], 4 used a self-designed Breastfeeding Knowledge Questionnaire [94,103,107,116], 1 used a Knowledge, Attitudes, and Practices questionnaire to assess this outcome [115], and 1 used the Breastfeeding and Breastfeeding Knowledge Test [86]. Detailed information is presented in (Multimedia Appendix 4) [34,35,42,43,49,56,57,78-121,124].

Among the included RCTs, 12 studies were classified as low risk [42,43,56,89,93,95,100,105,109,113,117,120] and 30 had some concerns [34,35,49,57,78,79,81,83-85,87,90-92,94,97,99, 101-104,107,110,114-116,118,119,121,122]. Most studies showed some concerns for deviations from intended interventions, primarily because blinding of participants and personnel was impractical in DHIs, and adherence to the intervention could not always be fully ensured. Three of these studies had dropout rates exceeding 5%, raising concerns about attrition bias [34,91,110]. Overall, 12 of the studies were rated as having some concerns because the outcomes were self-reported and outcome assessors were not blinded [35,83,84,90,92,97,101,114,115,119,121,122] (Figure 2 [34,35,42,43,49,56,57,78,79,81,83-85,87,89-95,97,99-105,107,109,110,113-122]).

Among the quasi-experimental studies included, 7 had a moderate risk [80,82,96,106,108,111,112], and 3 had a low risk [86,88,98]. In addition, 6 of the included studies did not report or use matching, stratification, or statistical adjustment to balance baseline differences between groups, resulting in a moderate risk of bias [80,82,96,106,108,112]. Regarding bias due to selection of participants, 2 studies were rated as having a moderate risk of bias due to their small sample sizes and insufficient representativeness [106,125]. Regarding selection bias, 2 studies were assessed as having a moderate risk of bias because the study sample did not represent the expected characteristics of the target population [80,111]. Regarding attrition bias, 5 studies had high participant dropout rates, indicating a moderate risk of bias [82,108,125-127]. All studies had a low risk of bias in terms of classification of interventions, outcome measurement, and selective reporting (Figure 3 [80,82,86,88,96,98,106,108,111,112]).

This systematic review applied the GRADE framework to assess the certainty of evidence for the effects of DHIs on breastfeeding outcomes. The certainty of evidence was moderate for EBF at <3 months and 3-6 months, downgraded by 1 level because of serious inconsistency (I²=84% and 75%, respectively). For EBF at ≥6 months, the certainty of evidence was low, downgraded for serious inconsistency (I²=86% and a wide 95% PI) and potential small-study effects identified by the Egger test (P=.006), although the pooled effect remained robust after trim-and-fill correction. The certainty of evidence for breastfeeding self-efficacy was moderate, downgraded because of serious inconsistency (I²=95% and a 95% PI crossing the null value). The certainty of evidence for breastfeeding knowledge was low, downgraded for serious inconsistency (I²=98% and a 95% PI crossing the null value) and serious imprecision due to the small number of studies (k=6), limited sample size, and an imprecise CI (Multimedia Appendix 5).

This meta-analysis used the Egger regression test to assess funnel plot symmetry and investigate potential small-study effects. The results showed that the funnel plots for EBF rates at <3 months and 3-6 months, breastfeeding self-efficacy, and breastfeeding knowledge were symmetrical, whereas the funnel plot for EBF rates at ≥6 months indicated asymmetry, suggesting a potential small-study effect. This may stem from publication bias, genuine heterogeneity, or methodological discrepancies across studies. The trim-and-fill method was used to assess potential small-study effects. Results showed that the adjusted effect size remained significant (RR 1.70, 95% CI 1.29-2.24; P<.001; I²=93%) even without adding missing studies (Multimedia Appendix 7). Notably, heterogeneity remained high after adjustment, indicating substantial between-study variability; therefore, the results should be interpreted with caution. The sensitivity analysis, conducted by sequentially removing each study, confirmed that no single study had an excessive influence on the overall pooled estimate of breastfeeding outcomes, demonstrating the robustness of the findings (Multimedia Appendix 8).

The meta-analysis confirmed that DHIs have a statistically reliable average effect on EBF rates and breastfeeding self-efficacy, but no statistically significant effect was observed for breastfeeding knowledge. Notably, these pooled estimates represent average effects across highly diverse study contexts. A wide 95% PI, however, indicates substantial real-world variability in individual DHI responses, restricting broad clinical deployment. Subgroup analyses also reflected this inconsistency, although intervention duration, country, gestational age, and number of fetuses provided some explanatory power; heterogeneity remained high in some subgroups and sample distribution was uneven, necessitating further research to elucidate the underlying causes of heterogeneity. These findings underscore that the potential of DHIs to improve breastfeeding outcomes still requires further validation.

Research has found that DHIs can increase EBF rates and breastfeeding self-efficacy, which is consistent with previous conclusions regarding the role of digital or remote support in promoting breastfeeding outcomes. These benefits may result from personalized guidance, timely feedback, and emotional support provided by DHIs [128]. Existing evidence suggests that prenatal and postpartum breastfeeding education, particularly for first-time mothers, is critical in improving early breastfeeding practices such as latching technique and feeding management [5,129]. By overcoming temporal and geographic barriers, DHIs extend traditional support through online consultation [56,130,131], multimedia education [132], and interactive feedback [133], enabling mothers to identify and correct breastfeeding problems promptly. In addition, peer support and online counseling components may reduce postpartum anxiety and strengthen maternal confidence [134]. Emerging technologies, such as VR-based demonstrations, may further facilitate skill acquisition [25,122]. Collectively, these integrated functions may contribute to both improved EBF practices and enhanced maternal breastfeeding self-efficacy.

It is important to note that the 95% PI for the effects of DHIs on both EBF rates and breastfeeding self-efficacy are wide and cross the line of no effect. This suggests substantial inconsistency in the expected effects of DHIs in future real-world settings, indicating that while the average effect is positive, the intervention may yield null or even negative outcomes in some contexts or subgroups. Such variability likely reflects true differences across populations, intervention designs, and implementation conditions, including differences in user engagement, intervention intensity, duration, and technological complexity. In addition, variations in outcome measurement tools may further contribute to inconsistencies across studies. Therefore, these findings highlight that the pooled estimates should be interpreted as average effects across highly diverse settings, rather than uniform effects applicable to all populations.

The impact of DHIs on breastfeeding knowledge remains uncertain, which differs from earlier systematic reviews that reported improvements in breastfeeding knowledge [135,136]. However, knowledge acquisition does not necessarily translate into behavioral change, as effective breastfeeding education requires sustained engagement, repeated reinforcement, and appropriate intervention intensity [137]. Excessive or poorly tailored information may lead to information overload and reduced maternal engagement [127,138,139], while factors such as baseline knowledge, educational level, prior breastfeeding experience, and lack of individualized intervention design may further weaken the effectiveness of standardized programs [139-141].

The 95% PI for breastfeeding knowledge indicates substantial variability in intervention effects across populations and settings, suggesting pronounced individual response differences. This uncertainty is further reflected by methodological limitations, including a small number of included studies, lack of standardized outcome measurement tools, and inadequate adjustment for key confounders, which may introduce bias and affect the precision of pooled estimates. Although the evidence certainty was rated as moderate and the P value approached statistical significance (P=.06), this does not provide conclusive evidence of ineffectiveness. Overall, these findings suggest that while the average effect may be limited, DHIs may still be beneficial in specific contexts or subgroups that are not captured by aggregate analyses, highlighting the need for larger, high-quality studies with standardized designs.

This meta-analysis followed rigorous methodological guidelines for systematic reviews, with comprehensive searches and sensitivity analyses supporting the robustness of findings. Furthermore, this review features broad inclusivity regarding study populations and interventions. Compared with previous reviews limited to telephone, SMS text messages, or mobile apps, our study included emerging DHIs, such as wearable devices and AI-based interventions. We also expanded the population scope to include women across pregnancy and postpartum stages and conducted comprehensive subgroup analyses to identify key technical parameters of the interventions and elucidated the varying effects of DHIs across countries with different levels of economic development and among different maternal populations. In addition, the inclusion of recent trials and GRADE assessment strengthened the reliability of evidence. These findings provide insights for optimizing digital breastfeeding support and suggest that DHIs may complement traditional care by improving accessibility and continuity of support.

However, this systematic review also has limitations. First, substantial heterogeneity remained in some outcomes despite predefined subgroup analyses, likely due to differences in intervention characteristics, user engagement, measurement tools, and participant populations. Second, the limited number of studies and uneven subgroup distribution reduced statistical power and estimate precision. Third, the inclusion of quasi-experimental studies may have introduced selection, performance, and measurement biases, while variations in outcome assessment tools and restriction to English-language publications may further affect comparability and increase bias risk. Moreover, the overall evidence certainty was low to moderate, and findings should therefore be interpreted cautiously [161]. Evidence for emerging technologies, such as AI-based chatbots and machine learning interventions, remains limited. Future well-designed, adequately powered RCTs with standardized outcomes and subgroup analyses are needed to clarify heterogeneity and evaluate effectiveness, acceptability, cost-effectiveness, and dose-response relationships.

This study comprehensively assessed the impact of DHIs on breastfeeding outcomes across multiple dimensions (type, duration, stage, frequency, country, and population characteristics) and outcomes (EBF rate, breastfeeding self-efficacy, and breastfeeding knowledge). Studies suggest that digital interventions may effectively improve EBF rates and breastfeeding self-efficacy, yet no statistically significant evidence currently supports their benefits for breastfeeding knowledge. Exploratory subgroup analyses revealed comparatively stronger intervention benefits in several specific subgroups, namely mHealth interventions, programs lasting a minimum of three months, studies carried out in developing countries, adult maternal samples, and women who underwent cesarean delivery. Our pooled analysis indicated that average pooled effects favor DHIs, but significant between-study heterogeneity and methodological differences among included studies may further limit the generalizability of pooled estimates, and applicability across settings remains uncertain. This study only provides preliminary exploratory evidence for the implementation of digital interventions in breastfeeding support. Larger, rigorously designed RCTs are warranted in future research to verify the efficacy of digital health technologies within real-world contexts and generate evidence to inform optimal intervention strategies tailored to diverse populations.