The Flipped Classroom in Medical Education: Systematic Review and Meta-Analysis

Abstract

Background

The concept of flipped classrooms (FCs) is gaining attention in medical education as it aligns with the digital age’s demand for more interactive and accessible learning experiences. By shifting the delivery of instructional content outside of the classroom, an FC allows students to engage with materials at their own pace, thereby maximizing in-class time for discussions, problem-solving, and other active learning activities.

Objective

This study aimed to conduct a comprehensive meta-analysis to appraise the comparative effectiveness of FC instruction in contrast to traditional pedagogical modalities, with a particular focus on postepidemic analyses within specific subfields of medical education.

Methods

The PubMed, Web of Science, and Scopus databases were systematically searched for studies comparing academic outcomes between the FC and traditional learning approaches in medical education. The primary outcome measures were knowledge assessment and students’ satisfaction. The standardized mean difference (SMD) was used as a measure of the overall effect, and subgroup analysis was performed according to the study design (randomized controlled trial [RCT] vs observational). The Cochran Q test and Baujat plots were used to estimate heterogeneity, coupled with I². Highly influential studies were identified; sensitivity analyses and metaregression were performed.

Results

In total, 141 studies were included in the systematic review; 127 (90.1%) studies with 21,171 participants were included in the meta-analysis of students’ knowledge assessment, of which 37 (29.1%) were RCTs. FCs had significantly better outcomes than the traditional method in knowledge test scores in both observational studies and RCTs (SMD 0.90, 95% CI 0.59-1.20, P<.001 and SMD 0.93, 95% CI 0.65-1.22, P<.001, respectively). There was substantial heterogeneity among included studies (I²=95.2%, τ²=1.614; P<.001). The funnel plot showed high asymmetry with significant small study effects (P<.001). However, the effect estimate remained robust to the exclusion of highly influential studies in the sensitivity analysis. In total, 27 (21.3%) studies with a total of 5842 participants reported students’ satisfaction. Higher student satisfaction scores for FCs were demonstrated in contrast to control groups (SMD 0.82, 95% CI 0.45-1.19; P<.001). There was substantial heterogeneity among the included studies (I²=97.8%, τ²=0.913; P<.001) but no evidence for publication bias, and no studies were found to be influential.

Conclusions

The FC method is associated with better knowledge achievement and greater student satisfaction than the traditional approach in medical education, paving the way for its broader integration into medical school curricula. However, it is essential to consider various factors, such as the availability of resources, faculty readiness, and student preferences when implementing any new educational approach. This study holds promise for advancing medical education by exploring innovative teaching methodologies that leverage technology to enhance learning outcomes.

Introduction

The rapid advancement of emerging technologies has prompted a reassessment of prevailing educational methodologies [1-3]. The wide accessibility of modern information and communication (IC) technologies has fundamentally altered the means of information retrieval, collaborative endeavors, and communication channels. These technologies have facilitated a novel learning paradigm wherein students enjoy and benefit from immediate and constant access to educational resources. While conventional pedagogical approaches have historically served to disseminate information broadly, they may exhibit shortcomings in cultivating solution-oriented competencies and practical knowledge application among students [4]. Consequently, there has been a rising scholarly interest in the flipped classroom (FC) model, also referred to as “flipped learning,” which represents a pedagogical framework that disrupts conventional learning modalities by integrating traditional face-to-face instruction with online educational platforms [5]. It involves the delivery of foundational content outside of traditional class hours, often via digital mediums. Conversely, activities traditionally designated as homework, such as problem-solving exercises and application-oriented tasks, are relocated to the classroom environment [2]. This instructional approach typically necessitates students’ independent engagement with prerecorded lectures or online materials before in-person instructional sessions, which are structured around student-centered activities aimed at fostering the application of previously acquired knowledge [6].

Moreover, the inconsistency inherent in traditional formative assessment practices and the limited avenues for communication between educators and students present challenges in addressing the entirety of the learning continuum. However, the FC model may offer solutions to these challenges associated with foundational learning and motivational impediments [7,8]. As traditional paradigms of medical education demand considerable temporal and resource allocations from both instructors and learners, the assessment of the potential impact of an FC and its practical use within the current medical education curricula has become paramount in enhancing educational efficacy [9-12]. Nonetheless, entrenched inertia within large academic institutions often impedes the adoption of innovative pedagogical approaches, notwithstanding meta-analytic findings substantiating the efficacy of FCs in health-related education [13,14]. This resistance to change is partly attributable to historical gaps in training directed toward students and educators in leveraging IC technologies within instructional contexts [15]. Successful navigation of technologically enriched educational milieux within the medical domain necessitates the acquisition of IC competencies, compounded by challenges uniquely relevant to medical education.

A recent systematic review and meta-analysis assessing the effects of an FC on students’ learning demonstrated a large variance across various disciplines and academic levels, with medicine being among those scientific fields having the lowest effect [16]. These findings further support the role of assessing the possible moderators and confounds [16] to explain the true impact of the FC model on advanced learning in a particular academic field. However, among the 46 identified meta-analyses, just a few examined student performances in an FC compared to the traditional learning environment for health care professionals (in a pre−COVID-19 pandemic setting), and none with a specific focus on higher medical education, excluding related fields, such as dentistry, pharmacy, nursing, and others. In addition, the postpandemic landscape warrants a reevaluation of the prospects for FC implementation, especially in medical fields. In response to the pressing circumstances during the COVID-19 pandemic, medical faculties have undergone adaptations to operate within pandemic-induced restrictions, bolstering capacities for digital engagement through intensive training initiatives targeting students and faculties on IC technologies use within instructional frameworks [17]. Thus, in this study, we aimed to conduct a comprehensive systematic review and meta-analysis to appraise the comparative effectiveness of FC instruction in contrast to traditional pedagogical modalities, with a particular focus on postepidemic analyses within specific subfields of medical education. This approach aims to address gaps identified in the published literature and highlight the relevance of an FC in evolving medical education contexts.

Methods

The systematic review was conducted per the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines of Observational Studies in Epidemiology [18,19]. The search strategy was developed by a biostatistician and informatician with expertise in conducting systematic reviews and meta-analyses and experience using IC technologies as educational tools (NM and SM).

Search Strategy

The PubMed, Web of Science, and Scopus databases were searched until October 10, 2024, for studies containing the following search terms: “(flip* OR invert*) AND (classroom OR learn* OR instruction* OR course*) AND medic*.”

Eligibility Criteria

Rayyan, a web-based technology platform for conducting systematic reviews, was used to store publications, delete duplicates, record the reviewer’s decisions, and help resolve conflicts. After removing duplicates, 2 reviewers independently screened potentially relevant titles and abstracts. The papers were equally distributed among reviewer pairs (D Spaic and NR; KM and MS; JM-L and NG; and Nikola Milic and AD). Discrepancies during the screening process were resolved through discussion, either with the involvement of a third reviewer (D Stanisavljevic, DB, JVM, VJ, SM, and Natasa Milic) or by reaching a consensus among the reviewers. Before proceeding with the full-text screening process, studies underwent inclusion based on eligibility criteria and exclusion if the title and abstract lacked sufficient information. The following inclusion criteria were used: (1) studies that compared FCs with any other learning method, (2) studies focusing on medical students, and (3) original articles.

Articles containing any of the following were excluded: (1) studies exclusively evaluating the efficacy of the FC approach (ie, studies without control group) or studies that did not evaluate the FC approach; (2) studies that did not report outcomes of interest (ie, knowledge assessment or students’ satisfaction); (3) studies that included health care professionals or students of disciplines other than medicine; (4) studies not written in the English language; and (5) publications that were book chapters, reviews, case reports, editorials, letters to the editor, conference abstracts, and theses. The full-text papers were equally distributed among reviewer pairs (D Spaic and NR; KM and MS; JM-L and NG; and Nikola Milic and AD), and each reviewer independently screened the article’s text. In the case of disagreements, a third reviewer (D Stanisavljevic, DB, JVM, VJ, SM, and Natasa Milic) independently evaluated the articles and made the final decision regarding their inclusion or exclusion.

Data Extraction

Two independent reviewers extracted the following data: authors, publication year, country of origin, study population, study design (randomized controlled trial [RCT] or non-RCT design and whether baseline testing was conducted), duration of intervention, sample size, age, and sex. The following relevant outcomes were extracted: student knowledge assessment and student satisfaction with used learning approach.

Individual Effect Size Pooling, Study Heterogeneity, and Risk of Bias

The primary outcome measures were knowledge assessment and students’ satisfaction, presented as means with SD. Students’ satisfaction reported as a proportion of students satisfied with the applied teaching approach was also assessed. The standardized mean difference (SMD) was used as a measure of the overall effect, and subgroup analysis was performed according to study design (RCTs vs observational). Random-effects models via the Der Simonian-Laird method were used to pool individual trial results, examine differences between the means of the FC and control group in units of SDs, and estimate SMDs. GetData Graph Digitizer (version 2.26.0.20 [20]) was used to read knowledge test scores and students’ satisfaction levels when figures but no tables were available in the original article. In cases where means and/or SDs were not reported, the median was used to approximate the arithmetic mean, and IQR/1.35 was used to approximate the SD. If SEs were reported, the SD was calculated as SD = SE × √n. In cases of reporting ranges, the SD was estimated as (maximum−minimum)/4. Cochran Q test and Baujat plots were used to estimate heterogeneity. Furthermore, we have reported I² heterogeneity measures considering the known limitations of Cochran Q tests. According to the Cochrane Handbook, I²<30%, I²=30%-60%, and I²>60 correspond to low, moderate, and high heterogeneity, respectively.

A separate forest plot was presented for each analysis, showing the SMD (box), 95% CI (lines), and weight (size of box) for each publication. We used the diamond symbol to represent the overall effect size (ES). Prediction interval, based on t-distribution, was also estimated and presented in addition to SMDs. Furthermore, SMDs were interpreted using Cohen d ES, and common language ESs were calculated to better communicate the differences between the FC and control groups. A P value of <.05 was considered statistically significant for all analyses. Data visualization and statistical analysis were done using R meta and metafor packages (version 4.0.0; R Foundation for Statistical Computing).

Influence Analysis

A separate analysis was conducted to identify highly influential studies and quantify their impact by evaluating the standardized residuals, difference in fits (DFFITS), Cook distance, leave-one-out 𝜏² (LOO-𝜏²), and Cochrane Q metrics for differential heterogeneity assessment. Influence analysis was performed using the dmetar package.

Sensitivity Analysis

Sensitivity analysis was conducted to examine the effects of (1) the exclusion of highly influential studies and (2) the replacement of knowledge test values in studies where multiple tests were performed.

Publication Bias

Publication bias was assessed by funnel plots. An Egger test was performed and reported alongside the funnel plots to quantify the amount of asymmetry. The trim-and-fill method was used to identify and impute hypothetically missing studies to correct for funnel plot asymmetry, providing an adjusted estimate of the effect.

Risk of Bias

Two reviewers independently evaluated the potential for bias in each RCT and the overall quality of the collected evidence using the Risk of Bias 2 tool, a component of the Cochrane Collaboration’s framework for assessing bias in randomized trials. The Risk of Bias 2 tool examines multiple domains of bias, including those related to the randomization process, deviations from intended interventions, missing outcome data, outcome measurement, and the selection of reported results. For observational studies, the reviewers independently assessed the risk of bias within each study using an adapted version of the Newcastle-Ottawa tool and the guidelines outlined by the Grading of Recommendations, Assessment, Development, and Evaluations Working Group.

Metaregression

Univariate and multivariate metaregression analyses were performed for final knowledge assessment as an outcome measure. Due to a lot of missing data on predictors within studies, we performed predictor selection (1) based on the frequency of reporting in a minimum of 100 studies and (2) by performing forward stepwise selection on a subset of studies with complete cases, optimizing for the Akaike Information Criterion.

Results

Study Selection

A total of 6940 potentially eligible publications were found. After removing duplicates, 4307 abstracts were screened for eligibility. Of the 325 publications sought for retrieval, 257 full-text articles were reviewed and 141 were selected for inclusion in the systematic review [21-161]. A flowchart illustrating this selection process is presented in Figure 1.

Figure 1.

Flowchart of study selection process. WoS: Web of Science.

Study Characteristics

Among the 141 studies in the systematic review, 69 (48.9%) were conducted in the United States and China. Studies were published between 2013 and 2024, with sample sizes ranging from 18 to 1016 participants. The study groups’ ages ranged from a minimum average age of 18.9 years to 38.5 years. In 3 (2.1%) studies, all participants were female. Of the 141 studies, 111 (78.7%) included undergraduate students, while 29 (20.6%) involved postgraduate students. More than half of the studies (78/141, 55.3%) focused on clinical subjects, 45 (31.9%) on basic science subjects, and 9 (6.4%) on public health. Overall, 104 (73.8%) were observational studies, and 37 (26.2%) were RCTs. A baseline knowledge assessment was performed in 44 (31.2%) studies, while 91 (64.5%) studies included only a final knowledge assessment. Detailed breakdowns of all eligible studies included in the systematic review are presented in Table S1 in Multimedia Appendix 1 [21-161].

Knowledge Assessment Between the FC and Control Group

In total, 127 studies with 21,171 participants (FC: n=9865; control: n=11,306) reported final knowledge scores. The meta-analysis results identified significant differences in final knowledge scores between the FC and control groups (SMD 0.90, 95% CI 0.68-1.13; P<.001), thus favoring an FC (Table 1). However, there was substantial heterogeneity among included studies (I²=95.2%, τ²=1.6136; P<.001; Figure S1 in Multimedia Appendix 1) and the funnel plot showed high asymmetry (Figure S1 in Multimedia Appendix 1) with significant small study effects (P<.001). After applying the trim-and-fill method, the adjusted ES was attenuated (SMD 0.30, 95% CI 0.002-0.59; P=.049; Table 2).

Table 1.

The standardized mean difference (SMD), Cohen effect size, and common language effect size (CLES) for final knowledge score and student satisfaction between the flipped classroom (FC) and control group.

Analysis			SMD (95% CI)		Effect sizea		CLES (95% CI)
Final knowledge score between the FC and control group
	All	0.90 (0.68-1.13)		Large		74 (68-79)
	Observational	0.90 (0.59-1.20)		Large		74 (66-80)
	RCTsb	0.93 (0.65-1.22)		Large		74 (68-81)
Final knowledge score in studies reporting baseline knowledge test
	All	0.97 (0.53-1.40)		Large		75 (65-84)
	Observational	0.99 (0.22-1.76)		Large		76 (56-89)
	RCTs	0.97 (0.50-1.44)		Large		75 (64-85)
Student satisfaction assessment
	All	0.82 (0.45-1.19)		Large		72 (62-80)
	Observational	0.87 (0.39-1.35)		Large		73 (61-83)
	RCTs	0.69 (0.17-1.22)		Medium		69 (55-81)

^aCohen d standardized effect size.

^bRCT: randomized controlled trial.

Table 2.

Meta-analysis of studies reporting knowledge scores (overall effect, influence, sensitivity, and trim-and-fill analysis).

Analysis	SMDa (95% CI)	95% prediction interval	I2 (95% CI)
Overall	0.90 (0.68-1.13)	−1.62 to 3.43	95.2 (94.6-95.6)
Influential study removedb	0.71 (0.55-0.86)	−0.98 to 2.40	94.3 (93.7-94.9)
Sensitivity analysis	0.84 (0.61-1.07)	−1.75 to 3.43	95.3 (94.8-95.8)
Trim and fill	0.30 (0.002-0.59)	−3.44 to 4.04	96.8 (96.5-97.0)

^aSMD: standardized mean difference.

^bAristotle et al [31], 2021; Cai et al [45], 2022; Feng et al [57], 2022; Khojasteh et al [81], 2021; and Peterson et al [118], 2017.

The observational study design was used in 90 studies with a total of 16,676 participants (FC: n=7613; control: n=9063), while the RCT design was used in 37 studies with a total number of 4495 participants (FC: n=2252; control: n=2243). The final knowledge scores were higher in an FC in contrast to the control group in both observational studies and RCTs (SMD 0.90, 95% CI 0.59-1.20; P<.001 and SMD 0.93, 95% CI 0.65-1.22; P<.001, respectively; Figures 2 and 3). In both observational and RCT studies, there was a substantial heterogeneity among the included studies (I²=95.6%, τ²=2.1195, P<.001 and I²=93.2%, τ²=0.7142, P<.001, respectively). Regarding the assessment of risk of bias in the RCTs included in the meta-analysis, there were some concerns about bias in most of the included studies (Figure S2 in Multimedia Appendix 1). For observational studies, there was an overall moderate risk of bias, with the quality of included studies ranging from 4 to 9, the highest rating being 9. Additional subgroup analyses of studies reporting knowledge scores based on study level, study subject, course duration, and differences among instructors reveal the same large heterogeneity among the included studies (Table S2 in Multimedia Appendix 1).

Figure 2.

Forest plot comparing final knowledge scores between the flipped classroom and control group in observational studies. Forest plot of final test score-01 in observational studies.

Figure 3.

Forest plot comparing final knowledge scores between the flipped classroom and control group in randomized controlled trials (RCTs). Forest plot of final test scores in RCTs.

The SMD for final knowledge scores in the overall meta-analysis, as well as in both observational studies and RCTs, indicated a large ES (≥0.8), with the FC group outperforming approximately 74% of individuals in the control group. Additional details on the common language ESs are provided in Table 1.

In the overall meta-analysis of the final knowledge score, 5 studies were found to be highly influential; however, the exclusion of these studies had no significant effect on the pooled effect estimate. The robustness of the effect estimate was also assessed in sensitivity analysis by rerunning the meta-analysis with outcomes substituted with alternatives in studies where they were available. The effect estimate remained robust to these changes (Table 2; Figure S3 in Multimedia Appendix 1).

Knowledge Assessment in Studies Reporting a Baseline Knowledge Score

In total, 40 studies with a total of 4542 cases (FC: n=2221; control: n=2321) implemented baseline knowledge assessment at study entry. The results of the meta-analysis presented differences in knowledge scores between the FC and control group at baseline (SMD 0.32, 95% CI 0.11-0.53; P=.003; Figure 4). The subgroup analysis revealed a significant difference in baseline knowledge between the FC and control group in observational studies (SMD 0.36, 95% CI 0.03-0.70), while in RCTs, there were no substantial differences (SMD 0.30, 95% CI 0.00-0.59; Figure 4).

Figure 4.

Forest plot for baseline knowledge scores between the flipped classroom and control group according to study design. Forest plot of baseline test scores.

The final knowledge scores in studies reporting baseline knowledge assessment were higher in an FC in contrast to the control group overall (SMD 0.97. 95% CI 0.53-1.40; P<.001; Figure 5), as well as in both observational studies and RCTs (SMD 0.99, 95% CI 0.22-1.76 and SMD 0.97, 95% CI 0.50-1.44, respectively; Figure 5). The SMD in the overall meta-analysis, as well as in both observational studies and RCTs, indicated a large ES (≥0.8), with the FC group outperforming approximately 75% of individuals in the control group in the overall analysis (Table 1).

Figure 5.

Forest plot comparing final knowledge scores between the flipped classroom and control group in studies reporting baseline knowledge test scores according to study design. Forest plot of final test scores in studies reporting baseline test.

Substantial heterogeneity was identified among included studies in both baseline and final analyses (I²=88.7%, τ²=0.41, P<.001 and I²=94.9%, τ²=1.90, P<.001, respectively; Figure S4 in Multimedia Appendix 1). The funnel plot found no evidence for publication bias in the assessment of baseline knowledge (P=.98; Figure S4 in Multimedia Appendix 1), while publication bias was found for the final knowledge assessment (P<.001; Figure S4 in Multimedia Appendix 1). The results of the trim-and-fill analysis presented an unchanged ES for the assessment of baseline knowledge, while an attenuated ES was present for the final knowledge assessment (Table 3).

Table 3.

Meta-analysis of studies reporting baseline and final knowledge score (overall effect, influence, and trim-and-fill analysis).

Analysis		SMDa (95% CI)	95% prediction interval		I2 (95% CI)
Baseline knowledge score as outcome measure
	Overall	0.32 (0.11-0.53)	−0.99 to 1.63		88.7 (85.5-91.1)
	Influential study removedb	0.22 (0.07-0.37)	−0.64 to 1.08		85.5 (81.0-88.9)
	Trim and fill	0.32 (0.11-0.53)	−0.99 to 1.63		88.7 (85.5-91.1)
Final knowledge score as outcome measure (in studies reporting baseline knowledge test)
	Overall	0.97 (0.53-1.40)		−1.86 to 3.79	94.9 (93.8-95.8)
	Influential study removedc	0.73 (0.42-1.04)		−1.20 to 2.67	93.8 (92.4-95.0)
	Trim and fill	0.64 (0.04-1.24)		−3.41 to 4.70	95.7 (94.8-96.4)

^aSMD: standardized mean difference.

^bAristotle et al [31], 2021; Malik et al [103], 2020.

^cAristotle et al [31], 2021; Khojasteh et al [81], 2021.

Among studies that reported baseline knowledge scores, 2 studies were detected as influential in the assessment of baseline knowledge (Table 3; Figure S5 in Multimedia Appendix 1), and 2 highly influential studies were identified in the final knowledge assessment (Table 3; Figure S6 in Multimedia Appendix 1). However, sensitivity analysis presented no significant effect of these studies on the pooled effect estimates in both analyses.

Student Satisfaction Assessment

A total of 27 studies with 5842 participants (FC: n=2464; control: n=3378) reported students’ satisfaction scores. The results of the meta-analysis overall effect showed higher students’ satisfaction scores for an FC in contrast to the control group (SMD 0.82, 95% CI 0.45-1.19; P<.001; Figure 6). There was substantial heterogeneity among the included studies (I²=97.8%, τ²=0.91, P<.001; Figure S7 in Multimedia Appendix 1) and borderline significance for publication bias (P=.051; Figure S7 in Multimedia Appendix 1), and no studies were found to be influential (Figure S8 in Multimedia Appendix 1). However, the results of the trim-and-fill analysis presented an attenuated ES (SMD 0.07, 95% CI −0.40 to 0.54; P=.76). The results of the subgroup meta-analysis showed significant differences in final knowledge scores between the FC and control group in both observational studies and RCTs; however, a lower ES was demonstrated in RCTs (SMD 0.87, 95% CI 0.39-1.35 and SMD 0.69, 95% CI 0.17-1.22, respectively; Figure 6).

Figure 6.

Forest plot comparing students’ satisfaction scores between the flipped classroom and control group according to study design. Forest plot of satisfaction scores.

The SMD in the overall analysis and in the analysis of observational studies indicated a large ES (≥0.8), with the FC group outperforming approximately 72% of individuals in the control group in the overall analysis. The SMD in the analysis of RCTs indicated a medium ES (0.69), with the FC group outperforming approximately 69% of individuals in the control group (Table 1).

A total of 4 studies with 786 cases (FC: n=404; control: n=382) reported a proportion of students satisfied with the teaching approach. The results of the meta-analysis showed borderline differences in the proportion of satisfied students between the FC and control group (OR 3.1206, 95% CI 0.97-9.99; P=.06; Figure 7). There was substantial heterogeneity among the included studies (I²=79.2%, τ²=1.156, P<.001).

Figure 7.

Forest plot comparing the proportion of students satisfied with the teaching approach between the flipped classroom and control group.

Metaregression for Final Knowledge Test Scores as the Outcome Measure

Metaregression (Table 4) showed a significant correlation (P=.01) between the course students have undertaken and ES (basic science subjects having a larger ES) but no significant relationship between study design (RCT or observational), knowledge assessment reporting (baseline/final vs final), having same versus different instructors in compared groups, study level (undergraduate vs postgraduate), duration (semester vs block), students’ age, sex, period (before, during, or after the COVID-19 pandemic), and the outcome (P>.05 for all analysis).

Table 4.

Univariate metaregression for final knowledge test scores as the outcome measure.

Variables		Studies, n	Coefficient (95% CI)	P value
Study design (RCTa/observational)		127	0.069 (−0.428 to 0.566)	.79
Knowledge assessment (baseline/final vs final)		123	0.076 (−0.419 to 0.573)	.76
Instructors (same vs different in compared groups)		41	0.091 (−0.991 to 1.173)	.87
Study level (undergraduate vs. postgraduate)		127	−0.003 (−0.569 to 0.563)	.99
Duration (semester vs block)		105	−0.025 (−0.530 to 0.481)	.92
Age (y)		44	−0.061 (−0.185 to 0.062)	.33
Subject (clinic/public health vs basic)		122	−0.609 (−1.091 to −0.126)	.01
Sex (female; %)		44	0.001 (−0.018 to 0.020)	.92
Period
	Before the COVID-19 pandemic (reference)	88	—b	—
	During the COVID-19 pandemic	19	0.105 (−0.566 to 0.775)	.78
	After the COVID-19 pandemic	11	0.037 (−0.810 to 0.883)	.88

^aRCT: randomized controlled trial.

^bNot applicable.

Due to a lot of missing data on some predictors within studies, for multivariate modeling, we used only predictors that are reported in a minimal 100 studies; that is, study design (RCT/observational), knowledge assessment (baseline/final vs final), study level (undergraduate vs. postgraduate), duration (semester vs block), subject (clinic/public health vs basic), and period (before, during, or after the COVID-19 pandemic). The results of the multimodel inference coefficients (Table 5) and predictor importance analysis (Figure S9 in Multimedia Appendix 1) confirmed that the most important predictor of FC success is the specificity of the course students have undertaken, indicating that basic science subjects have a larger ES.

Table 5.

Multimodel inference coefficients for final knowledge scores as outcome measure.

Variables			Coefficient		P value
Study design (RCTa/observational)			0.040		.88
Knowledge assessment (baseline/final vs final)			0.178		.59
Study level (undergraduate vs postgraduate)			−0.007		.98
Duration (semester vs block)			−0.232		.47
Subject (clinic/public health vs basic)			−0.900		.008
Period
	Before the COVID-19 pandemic (reference)	—b		—
	During the COVID-19 pandemic	0.184		.62
	After the COVID-19 pandemic	−0.118		.78

^aRCT: randomized controlled trial.

^bNot applicable.

Discussion

In this study, we conducted a comprehensive meta-analysis to appraise the comparative effectiveness of FC instruction in contrast to traditional pedagogical modalities in the context of medical education. The meta-analysis includes 127 studies, of which 37 are RCTs, thus enabling the assessment of a global effect measure of the FC strategy. Overall, we found a substantial positive effect (ES=0.90) of the FC approach. The pooled ES is comparably more pronounced than other fields (mathematics: ES=0.38; science: ES=0.50; science, technology, engineering, and mathematics: ES=0.90), suggesting that medical training is particularly amenable to the FC pedagogical strategy [162].

Most previously published meta-analyses included studies across various disciplines and academic levels, which limited the generalizability of their results to medical education. In addition, many studies from the medical educational context were observational in design, providing lower-quality evidence regarding the true effect of the FC model. In our study, we focused exclusively on studies related to medical education, excluding other disciplines and related fields such as dentistry, pharmacy, and nursing. We expanded the search strategy to include studies from after the COVID-19 pandemic period and presented ESs for the FC model in comparison to traditional learning methods, using high-quality study designs such as RCTs. It is important to note that the COVID-19 pandemic presented both novel challenges and opportunities for integrating technological innovations into medical pedagogy [163]. In our study, standardized ESs between RCT and observational studies were similar across medical disciplines and favored FC methods (observational: SMD 0.90, 95% CI 0.59-1.20, P<.001 and RCT: SMD 0.93, 95% CI 0.65-1.22, P<.001, respectively). However, between-study heterogeneity remained high irrespective of the type of study. This heterogeneity is likely a consequence of varied methodologies in FC application, partially because there is yet to be a universally accepted pedagogical implementation of the framework. Therefore, a distribution of ESs exists among the study population with differing success rates. Some inconsistently used sources of added heterogeneity could be preclass assessments, differing amounts of technology used, multiple types of handout material used, the use of group-based methods, the use of problem-based learning, the use of simulation exercises, and differences in outcome assessment. For example, the inconsistent use of preclass assessment (eg, quizzing) to maintain student engagement and enforce preparedness introduces added heterogeneity. Preclass assessment is a recognized mediator associated with larger ESs [162].

In addition, studies differed in the amount, quality, and type of technologies involved, some providing only prerecorded lecture material and others incorporating student lecture annotation with review, text-based handout material, or multiple types of student activities [164]. At other times, the scope of the applied FC model was significantly different, with some studies using complete curriculum redesigns and extensive lecture material having to be adapted [97]. In contrast, others (especially in the clinical setting) explored one-time learning opportunities tackling a single clinical concept [28,70]. The FC effect is highest when both quizzes and exercises (problem-based learning) are used together. However, the optimal mediator choice must be tailored to the curriculum at hand [162,165]. Studies also used group-based learning, project assignments, team-based assessment, and other ways to foster engagement [165].

Conceptually, the FC model should strive to replace instructor-driven passive learning with student-centered active learning. Acquiring clinical knowledge and skills requires active learning with an emphasis on problem-based learning, simulation sessions, clinical cases, and group-based learning—all activities where the FC approach is advantageous [166-168]. Given that active learning has proven benefits for knowledge acquisition and application, the question then becomes of identifying the salient features of the FC approach [162,169]. Material availability, self-paced learning, personalized teaching, and student-driven activities are thought to predict better learning outcomes [110,162,165]. The critical question is maintaining student engagement to use class time most effectively. Quizzing students in an FC model at the beginning of class is an external motivator for them to engage with the material before class, suggesting that some of the model’s effectiveness can be attributed to factors enforcing motivation. This preparatory assessment drives students to study in anticipation of the quizzes, leveraging external motivation to ensure they come to class equipped with the necessary knowledge at the cost of increased workload [165,166,170]. Increased performance also reflects increased workload and opportunity to practice. In one study, the number of classes required was doubled in the treatment arm [28]. However, these increases in workload might only scale if applied to part of the school curriculum.

There is also the looming question of outcome measurement. Final assessments could suffer from “teaching to the test” and not reflect any real-world knowledge application [169,171]. For example, FC approaches could increase test performance. However, if the application is not evaluated (through simulations and otherwise), these differences may not reflect the sought outcome—creating high-quality medical professionals.

Differences in learning objectives are the most striking delimiter between basic science and clinical subjects. Clinical subjects should teach students clinical skills in addition to clinical reasoning, compared to basic sciences, which do not have the added logistical hurdle of requiring simulation exams. Consequently, the ES of the flipped model is impacted by the discrepancy between the learning approach and assessment method (skill demonstration or test performance). For example, it would be impractical to assign students to make slides for their preclass assessment if the objective was to master the act of physical examination of the eye [23,93,169]. Curricula that provided students with examination stations and where students had access to simulation-based learning fared better in assessment [60]. In our study, subjects providing basic science skills presented a larger effect of the FC model in contrast to clinical subjects and the public health field. However, in some studies, problem-based learning was an excellent introductory approach to teach clinical skills, and studies that used case-based group discussions have noted increased scores in clinical case assessment [28,70], augmenting the goal of attaining relevant clinical skills [169].

Student engagement can be fostered through group-based discussions to diffuse student responsibility and grow independence. Problem-based learning and simulation can be used when clinical skills are the learning objective to develop clinical reasoning that can be generalized to the real-world setting. Recorded lecture material is vital as its availability makes it easy for self-paced learning and review for both physical examination and medical procedures [164]. However, the successful implementation of the FC approach requires careful planning, along with both technical and pedagogical support for instructors. Clearly defined learning objectives and aligning classroom activities with those objectives are critical for effective implementation [162]. One of the main challenges for instructors is the development of interactive content and structuring of lectures for independent student learning. These tasks demand additional time, technical resources, and a certain level of digital literacy from the teaching staff. Students may also face difficulties. Traditional laboratories and clinical practice offer direct sensory engagement and spontaneous learning opportunities through unforeseen situations—experiences that are difficult to replicate in a digital format. The absence of such practical, hands-on learning can weaken the connection between theoretical knowledge and real-world application. In addition, unequal access to digital resources due to unreliable internet connections or lack of appropriate equipment can exacerbate existing educational inequalities among students. This is why institutional efforts are also needed to ensure that students have equitable access to all necessary educational resources.

Student evaluations are commonly used to assess class and instructor quality. Indeed, student feedback showed significant satisfaction with FC learning models compared to traditional methods [67,156,159,172,173]. In both clinical and preclinical contexts, students expressed high levels of satisfaction, with flipped learning being the majority preferred strategy. Therefore, assessing student perception and its impact on teaching course satisfaction should be weighed when considering teaching reform. Importantly, current research demonstrates that student satisfaction may explain a minimal portion of learning variation [174]. There are concerns that student evaluations exhibit a systematic bias, making them poor markers of teaching effectiveness [175,176]. Some reports have found that despite the greater effectiveness of the flipped model, student satisfaction remained the same or even decreased [177,178]. This inconsistent effect of satisfaction could indicate that, for some learning objectives, the workload needed to achieve a better outcome might not align with students’ perceptions. The level of satisfaction could be related to factors irrelevant to the set objective [174,176,179]. The FC model tends to increase immediate workload in exchange for delayed but better outcomes. Further curricula optimization would entail differentiating excellent and bad workloads, which impacts learning and student satisfaction [170,180,181]. Allowing students to learn at their own pace and providing more opportunities for hands-on and engaging activities during class seem like good pedagogical tools [60,62,109,162]. Finally, the effectiveness of incorporating active learning seems to have the greatest impact on the overall effects of an FC [163].

Limitations

The findings of this study should be interpreted with caution because of the high methodological diversity related to study design and outcome measures, as well as statistical heterogeneity. Incomplete reporting was evident in the assessed literature; some studies may have reported only major findings, making it difficult to evaluate their methodology or to reuse their data. Our meta-analysis did not include conference papers, dissertations, and papers on languages other than English, which may have introduced a selection bias.

Conclusions

This study has the potential to advance medical education by exploring innovative, technology-enhanced teaching methodologies that improve learning outcomes. Its findings provide valuable insights for educators, curriculum developers, and policy makers regarding the efficacy of the FC model in medical education. The FC approach has been shown to promote better knowledge acquisition and higher student satisfaction compared to traditional methods, supporting its broader integration into medical school curricula, particularly for basic science subjects. However, successful implementation requires careful consideration of factors such as resource availability, faculty preparedness, and student preferences. To promote the FC model in low-resource settings, institutions should leverage free digital tools and open educational resources; prioritize low-cost, high-impact in-class strategies such as peer instruction and group discussions; and maximize existing infrastructure through community collaboration. In addition, future research should explore the integration of technology-assisted teaching tools, including artificial intelligence and adaptive learning platforms, to further enhance the effectiveness and scalability of FC methodologies in medical education.

Abbreviations

ES

effect size

FC

flipped classroom

IC

information and communication

PRISMA

Preferred Reporting Items for Systematic Reviews and Meta-Analyses

RCT

randomized controlled trial

SMD

standardized mean difference