Studies were included in the review if they met the predefined eligibility criteria, which were operationalized according to the PICOS framework as follows:

Population: Studies focused on early-childhood or early primary education, corresponding approximately to children aged 3–8 years, including preschool education and up to Grade 3.

Application of the PRISMA 2020 protocol yielded a structured multi-stage screening process. A total of 395 records were initially identified across databases. After automated duplicate removal, 373 unique records remained for title and abstract screening, resulting in the exclusion of 145 studies. The remaining 228 records proceeded to full-text assessment, of which 46 articles were evaluated against predefined eligibility criteria using a standardized checklist.

During eligibility assessment, 22 studies were excluded due to language restrictions (Russian or Chinese), absence of the arts dimension (pure STEM focus), lack of empirical evaluation, or duplication of content.

A final set of 24 studies met all inclusion criteria. Of these, 17 empirical studies constituted the analytical corpus for qualitative synthesis and meta-analysis, while 7 systematic reviews and meta-analyses provided theoretical and contextual support for interpretation.

The complete selection process is illustrated in Figure 1, which presents the PRISMA 2020 flow diagram detailing identification, screening, eligibility, and inclusion stages.

The RoB 2.0 tool (Sterne et al., 2019) was applied to empirical studies employing experimental or quasi-experimental designs, particularly those involving AR/VR interventions with comparison groups. Overall, the pattern indicates a low-to-moderate risk of bias, consistent with methodological constraints typical of classroom-based research.

The strongest domains were random sequence generation and selective outcome reporting, supporting the internal consistency of the quantitative evidence. Allocation concealment showed moderate limitations, while blinding emerged as the most vulnerable domain. This limitation is structurally inherent in educational interventions, where participants and teachers are aware of the technological condition (e.g., use of VR headsets, AR devices, or makerspace activities), potentially introducing performance or expectancy effects.

Table 1 summarizes the distribution of risk levels across assessed domains.

The overall low-to-moderate risk profile supports the inclusion of these studies in the meta-analysis. Nevertheless, the elevated risk observed in blinding should be acknowledged as a structural limitation inherent to educational interventions and interpreted cautiously, particularly in short-duration designs where expectancy or performance effects may be amplified. To account for residual heterogeneity and between-study variability, a random-effects model was employed in the quantitative synthesis.

The Mixed Methods Appraisal Tool (MMAT; Hong et al., 2018) was applied to qualitative, case-study, and mixed-methods studies within the empirical corpus, particularly those examining AEI/robotics and FabLab/Maker interventions, where contextual implementation is central.

Overall, the corpus demonstrates strong methodological coherence. All studies clearly articulated research objectives, and a high proportion exhibited alignment between reported findings and stated conclusions, supporting the credibility of the qualitative synthesis.

Table 2 presents the distribution of methodological quality across assessed MMAT dimensions.

The strongest dimensions—objective clarity and results–conclusion coherence—support the robustness of contextual interpretations presented in the narrative synthesis. Moderate limitations were observed in data triangulation and analytical transparency, indicating that some studies provide limited detail regarding integration of multiple data sources or explicit analytical procedures.

These constraints are particularly relevant in FabLab/Maker research, where learning outcomes are highly dependent on facilitation style and task configuration. Accordingly, findings derived from single-context implementations should be interpreted cautiously and not generalized beyond comparable pedagogical settings.

To provide an article-level confidence statement aligned with high-impact systematic review standards, an education-adapted GRADE approach (Guyatt et al., 2011) was applied. Confidence was synthesized across five domains: risk of bias (RoB 2.0 + MMAT triangulation), consistency of findings, directness to early-childhood STEAM contexts, precision of estimates, and potential publication bias. The overall evidence assessment based on the education-adapted GRADE framework is summarized in Table 3.

Across domains, the evidence base demonstrates moderate-to-high overall confidence. AR/VR interventions show the most stable quantitative patterns, particularly for motivational and cognitive outcomes when pedagogically scaffolded. AEI and robotics interventions present consistent socio-emotional and engagement-related benefits, although effect magnitude appears sensitive to interaction design and adult mediation.

FabLab/Maker evidence remains credible but comparatively context-dependent, with fewer replicated experimental designs and greater reliance on case-based implementations. Consequently, confidence in these findings is moderate and contingent upon implementation quality, facilitation style, and curricular integration.

Taken together, the adapted GRADE synthesis supports the inclusion of these domains in an integrative framework while reinforcing the importance of developmental alignment and pedagogical mediation as primary determinants of effectiveness rather than technology type alone.

Effect sizes were calculated using Hedges’ g, which adjusts for small-sample bias and is therefore appropriate for early-childhood educational research contexts. When not explicitly reported, effect sizes were derived from published descriptive statistics following standard meta-analytic procedures. Magnitudes were interpreted using conventional benchmarks (0.20 = small; 0.50 = medium; ≥0.80 = large).

Given expected variation across technology type, intervention duration, educational context, and outcome measures, a random-effects model (DerSimonian–Laird estimator) was applied. This approach assumes that true effects vary across studies and provides more conservative and generalizable estimates than fixed-effect models. Heterogeneity was assessed using Cochran's Q and the I² index, interpreted as ≤30% (low), ∼50% (moderate), and ≥75% (high). Moderate-to-high heterogeneity was observed (I² ≈ 63%), reflecting substantive diversity in technological mediation and pedagogical implementation rather than methodological inconsistency.

Immersive technologies—Augmented Reality (AR) and Virtual Reality (VR)—constitute the largest subgroup within the empirical corpus, accounting for 7 of the 17 studies (41%) included in the descriptive and quantitative synthesis. The AR/VR empirical evidence included in the review is summarized in Table 5. These studies were conducted across diverse geographic and educational contexts and collectively provide robust evidence on the pedagogical potential of immersive environments in early-childhood STEAM education. Overall, the AR/VR interventions consistently report positive effects on both cognitive learning outcomes and motivational variables, with effect directions uniformly favoring technology-enhanced conditions over traditional instructional approaches.

Artificial Emotional Intelligence (AEI) constitutes a distinct and emerging domain within early-childhood STEAM education, represented by 5 of the 17 empirical studies (29%) included in the review. The empirical evidence on artificial emotional intelligence interventions is summarized in Table 6. These studies focus on emotion-aware educational agents, such as social robots, affective tutors, and adaptive robotic systems, designed to perceive and respond to children's emotional and behavioral cues during learning activities. Overall, the evidence indicates moderate but highly consistent positive effects of AEI-based interventions, particularly in socio-emotional and engagement-related outcomes, which are critical at early developmental stages.

FabLabs and Maker-based learning environments represent the third major technological domain analyzed in this review, accounting for 5 of the 17 empirical studies (29%). The empirical evidence on FabLab and Maker-based environments is summarized in Table 7. These studies focus on hands-on, fabrication-oriented STEAM learning, in which children engage in the design, construction, testing, and refinement of tangible artifacts using age-appropriate tools and materials. Overall, the empirical evidence highlights the distinctive contribution of FabLab/Maker environments to the development of creative, collaborative, and process-oriented competencies, which are central to the STEAM educational philosophy.

From the 17 empirical studies identified through the PRISMA selection process and summarized in Table 4, 10 met the criteria for quantitative comparability and were included in the meta-analysis. These studies employed experimental, quasi-experimental, comparative, or pre–post designs and provided sufficient statistical information to compute standardized effect sizes. The quantitative synthesis was based on 10 analytically comparable effect-size estimates rather than on a direct aggregation of participant counts. The meta-analysis summary of effects by technological subgroup is presented in Table 8.

A random-effects model was applied to account for expected conceptual and methodological variability across interventions. The analysis revealed a positive and statistically significant overall effect of technology-enhanced STEAM interventions:

Overall combined effect: Hedges’ g = 0.66, p < 0.001

Figure 2 presents individual study effect sizes (Hedges’ g) with 95% confidence intervals, along with the pooled estimate calculated using the DerSimonian–Laird random-effects model. The diamond represents the overall combined effect (g = 0.66), and the vertical reference line denotes the null effect (g = 0).

To further interpret sources of variability and to ensure alignment between quantitative findings and the descriptive synthesis, subgroup analyses were conducted by technological domain: AR/VR, Artificial Emotional Intelligence (AEI), and FabLab/Maker environments. Each study included in the meta-analysis contributed one analytically comparable effect-size estimate; therefore, the number of studies corresponds to the number of effect-size estimates included in each subgroup.

The subgroup analyses indicate variation in effect magnitude across technological domains. AR/VR interventions demonstrate the largest pooled effect (g = 0.78), suggesting comparatively strong associations with learning-related outcomes. The moderate heterogeneity observed in this subgroup likely reflects variation in implementation format, duration, and outcome focus rather than inconsistency in direction of effects.

AEI-based interventions yield a moderate-to-large pooled effect (g = 0.65) with comparatively lower heterogeneity (I² = 47%). This pattern suggests more stable effects across contexts, particularly for outcomes related to engagement, attention, and emotional regulation. The narrower variability may reflect the focused functional role of emotion-aware systems within guided classroom activities.

FabLab/Maker environments show a moderate pooled effect (g = 0.54), accompanied by the highest heterogeneity (I² = 63%). This dispersion is consistent with the contextual sensitivity of maker-based pedagogies, where instructional scaffolding, facilitation style, and task design substantially mediate outcomes.

Figure 3 presents pooled effect sizes and confidence intervals across technological domains. The visual comparison highlights both the comparatively larger magnitude associated with immersive technologies and the greater dispersion observed in maker-based interventions.

The analysis by educational stage revealed comparable effect sizes between preschool education (≤5 years) and early primary education (6–8 years). Pooled effects were approximately g ≈ 0.60 for preschool-aged children and g ≈ 0.65 for early primary students. This similarity indicates that the effectiveness of immersive technologies, AEI, and FabLab/Maker environments is not confined to a single early developmental stage, provided that interventions are appropriately adapted to children's cognitive, emotional, and motor capabilities.

These findings suggest that early exposure to STEAM-oriented technologies can be pedagogically beneficial across the transition from preschool to primary education, supporting continuity in learning trajectories rather than isolated, age-specific effects.

When studies were grouped by geographic context, a non-significant trend toward larger effect sizes was observed in Asian educational contexts (g ≈ 0.70) compared with Western contexts (g ≈ 0.55). Although this difference did not reach statistical significance, it may reflect contextual factors, such as variations in baseline instructional practices, curricular integration of technology, or teacher familiarity with structured technology-based interventions.

Importantly, the absence of statistically significant regional differences suggests that the positive effects of technology-enhanced STEAM education are broadly transferable across cultural and educational systems, reinforcing the global relevance of the findings.

The comparison between cognitive–academic outcomes (e.g., learning performance, comprehension, conceptual understanding) and creative–attitudinal outcomes (e.g., motivation, creativity, engagement, attitudes toward learning) revealed similar pooled effect sizes, both close to g ≈ 0.60. This result indicates that the analyzed interventions do not privilege one outcome domain at the expense of another.

Rather, technology-enhanced STEAM approaches appear to support multidimensional learning gains, simultaneously strengthening academic learning and key dispositions such as curiosity, persistence, and creative engagement—core objectives of the STEAM educational paradigm.

Structurally, the framework implies a balanced configuration of learning environments integrating physical, digital, and emotional mediation.

This configuration includes:

Hybrid environments combining tangible maker resources with developmentally appropriate immersive systems.