This review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines [7]. This systematic review adhered to a predefined registered protocol in the International Prospective Register of Systematic Reviews (PROSPERO) under registration number CRD420251142906. The methodology followed the PRISMA 2020 and PRISMA-S guidelines for transparent reporting of systematic reviews and search strategies.

A comprehensive search using PubMed, Scopus, Web of Science, Google Scholar, and the World Health Organization (WHO) Global Health Library databases was conducted to identify studies between 01 January 2010 and 30 December 2025. The literature search aimed to identify articles that reviewed crowd management, mitigation, and safety plans for stampedes and mass gatherings. Only studies published in English were included due to feasibility constraints in screening and data extraction. The usage of relevant keywords and controlled vocabulary was considered in all the databases as follows: (“stampede” OR “crush” OR “crowd” OR “mass gathering” OR “large events”) AND (“safety” OR “crowd management” OR “risk mitigation” OR “emergency preparedness” OR “public safety”). The detailed search strategy was based on the PRISMA-2020, for which the checklist is provided in Supplementary Material S1, and the search strategy is provided in Supplementary Material S2, with the databases searched, date limits, and the number of records retrieved in each database. Reviews, editorials, opinion pieces, and conference abstracts without full empirical data were excluded.

The PICOS framework was used to determine the scope and relevance of studies conducted to follow eligibility and selection procedure regulations, and to define the scope of this review. The Population (P) consisted of individuals attending mass public gatherings, such as religious pilgrimages, music festivals, sporting events, and political rally settings, where the risk of crowd crushes or stampedes is high [8]. The Intervention (I) focused on preventive or crowd-management strategies designed to reduce the likelihood or severity of stampedes. These included physical design modifications (e.g., one-way flow systems, barrier placements), operational strategies (e.g., crowd density control, entry/exit flow management), technological tools (e.g., real-time surveillance, AI-based crowd simulations), and communication-based interventions (e.g., public announcements, signage, and emergency drills). The Comparator (C) included standard practices in crowd management, or, where applicable, specific contexts that lacked formal intervention. The Outcomes (O) included both direct and proxy measures of safety effectiveness, such as reduced incidence of crowd crushes or stampedes, improved crowd flow, increased evacuation efficiency, and reduced density-related congestion [9]. Lastly, the Study design (S) encompassed a broad range of empirical evidence, including observational studies, quasi-experimental designs, simulation-based modeling studies, and retrospective analyses of real-world events. Simulation and modeling studies were included to gain insights into crowd dynamics and testing of preventive strategies under controlled and high-risk conditions that are not feasible in real-world experimental settings. Such an extensive set of PICOS structures ensured the inclusion of a wide range of applicable evidence in the synthesis of best practices for stampede prevention and crowd safety.

Following the literature search, all retrieved records were managed systematically to ensure transparency and reproducibility. The screening was conducted using a multi-phase approach: Phase 1 involved the removal of duplicate records across databases; Phase 2 involved screening of titles and abstracts against eligibility criteria and predefined keywords (e.g., “stampede,” “crowd management,” “risk mitigation”); Phase 3 involved further screening of the contents of articles by reading the full text using the PICOS framework; Phase 4 involved the assessment of outcome levels to select those that reported primary or secondary outcomes related to crowd safety (e.g., evacuation time, crowd density); and Phase 5 involved the inclusion of relevant gray literature (e.g., WHO Global Health Library) that reported applicable prevention strategies. The screening process was conducted independently by two reviewers (RSC and AS) to identify studies that met the inclusion criteria. A third reviewer was consulted to resolve any discrepancies or disagreements between the two reviewers. Reasons for exclusion (e.g., wrong context, no evaluation of prevention strategy) were recorded at the full-text stage. The study selection process adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines. The study selection process is summarized in the PRISMA flow diagram (see Figure 1).

In this review, the methodological quality of the articles was evaluated using the Cochrane Risk Of Bias In Non-randomized Studies-of Interventions (ROBINS-I) [10]. The considered biases included random participant assignment, departures from intended interventions, missing outcome data, outcome measurement bias, bias in the selection of reported results, and the risk of bias in general.

Emerging applications of artificial intelligence (AI) and machine learning (ML) have shown potential in enhancing crowd monitoring and prediction capabilities. Evidence from recent studies suggests that AI-based surveillance and predictive modeling can help identify high-density zones and improve crowd-flow management. For example, in a simulated case study, a convolutional neural network (CNN)-based crowd surveillance system accurately detected areas of critical density in real time with 93% accuracy [11]. Another modeling-based study has applied a predictive ML model trained on historical stampede data from Mecca and India to optimize gate scheduling during entry/exit events, resulting in a 27% reduction in simulated congestion time [12].

The scope of AI has recently expanded beyond physical sensors to include the sentiment analysis of social media data, offering a transformative tool for predicting crowd behavior by detecting shifts in the “emotional state” of the gathering. Furthermore, recent field evaluations have transitioned these technologies from simulation to large-scale reality; the 2025 Maha Kumbh Mela successfully utilized AI-powered density surveillance and drone-assisted evaluations, integrated into a centralized command center, to manage record-breaking crowd densities [13, 14].

While these findings highlight the potential of AI/ML to improve anticipatory and adaptive crowd safety strategies, they were interpreted as predictive or exploratory evidence rather than directly generalizable real-world outcomes. During thematic synthesis, AI/ML-based studies were primarily mapped to the theme of crowd monitoring and technology, while also contributing insights into predictive and adaptive crowd management strategies.

From each included study, we extracted details on authors, year, country, setting/event type, study design, sample size, description of the intervention(s) or strategy, comparator (if any), outcomes measured (e.g., crowd density metrics, casualty reduction, evacuation time), and key results are provided in Supplementary Material S3. One of the reviewers extracted the data using a standardized extraction form, which was verified by the other. Given the diversity of included studies, outcome measures were heterogeneous. They included both direct indicators (e.g., incidence of crowd crush or injury) and proxy indicators (e.g., crowd density, evacuation time, congestion levels, and situational awareness). These outcomes were synthesized narratively to allow comparison across varied study designs.

Quality appraisal was conducted using two design-specific tools. The ROBINS-I tool was applied to seven non-randomized intervention and simulation studies, generating domain-level ratings from Low to Critical risk of bias [13, 15–20] (see Supplementary Material S5). The JBI Critical Appraisal Checklist for Analytical Cross-sectional Studies was applied to six observational and descriptive studies, producing criterion-based Yes/No/Unclear judgments [14, 21–25] (see Supplementary Material S6). Results from both tools were synthesized qualitatively to inform study weighting in the narrative synthesis (see Supplementary Material S7).

A quantitative meta-analysis could not be conducted because the included studies were highly heterogeneous in design, context, methodology, and outcome metrics. Hence, a thematic (narrative) synthesis method was used. The interventions were grouped into engineering controls, operational interventions, technology-enabled solutions, and communications/awareness measures. Each theme followed a deductive coding approach in which intervention categories were predefined based on commonly reported domains in the crowd-management literature. Themes were interpreted to determine the direction and strength of the reported effects, allowing commonalities across interventions to be identified. When available, relevant quantitative information (e.g., mean differences, simulated density reductions, evacuation times, or effect sizes) was summarized to strengthen and accompany the qualitative synthesis. Two reviewers, using a deductive approach based on predefined, theory-driven themes identified from the literature, independently performed thematic coding of the included studies. Data extraction and quality assessment were also conducted independently to ensure methodological rigor. Discrepancies at any stage were discussed and resolved through consensus; where disagreements persisted, a third reviewer adjudicated. to ensure consistency in classification.

A study-weighting framework was applied to reflect each study’s relative contribution to the overall synthesis. Each study’s weight was derived from (1) methodological quality based on JBI or ROBINS-I appraisal, (2) data richness (empirical or high-fidelity simulation evidence), and (3) relevance to the review objectives. Scores were assigned on a 1-3 scale for each factor, normalized to a total weight of 1.00 across all included studies. These weights were not used for statistical pooling but to visually indicate each study’s influence within the narrative synthesis (see Figure 5).

In total, 833 articles were identified through database and gray literature searches, comprising 146 from PubMed, 182 from Scopus, 134 from Web of Science, and 371 from gray literature sources, including the WHO (57) and Google Scholar (314). After removing 18 duplicates, 815 records remained for screening based on titles and abstracts. During this phase, 424 records were excluded because 240 had an irrelevant topic or context, 90 were not intervention-focused, 56 were duplicates or non-English studies, and 38 were commentaries or editorials lacking empirical data. Following this screening, 391 reports were sought for retrieval, but 366 were not retrieved due to a lack of accessible full text (178), a lack of author contact response (95), language or translation limitations (60), or incomplete or retracted records (33). Consequently, 25 reports were successfully retrieved and assessed for eligibility. Twelve citations were excluded at the full-text stage: eight failed to address stampede prevention strategies specifically, and four were excluded due to low methodological quality. Ultimately, 13 studies met all inclusion criteria and were included in the systematic synthesis (see Figure 1).

The essential characteristics of the 13 studies included in this review are presented in Supplementary Material S4. The publication years range from 2010 to 2025, and recent years show more publications. Geographically, studies originated mainly from Asia and the Middle East (n = 10; India, China, South Korea, Egypt, Saudi Arabia, and Iran), with two from the United States and one from the United Kingdom. Event settings included religious pilgrimages (e.g., Hajj, religious festivals), sports and music events, urban festivals, and transportation hubs. Study designs were heterogeneous: three simulation or modeling studies, two experimental or quasi-experimental evaluations, three retrospective observational analyses or reviews, and five descriptive case studies or field evaluations that used mixed-methods and contextual analysis. Most of the results presented in the articles considered composite outcomes, including measures of crowd density, evacuation time, or near-miss incidents; specific injury rates or stampede occurrences were reported in a minority of the studies. Common interventions included one-way pedestrian flows, physical barriers/fencing, crowd capacity limits, increased security staffing, crowd-monitoring systems (CCTV and mobile data), emergency preparedness drills, and emergency communication measures (signage and audio announcements). Randomized controlled trials were not identified; evidence was primarily derived from observational or modeling research.

Methodological quality appraisal results are presented in Supplementary Material S5 (ROBINS-I), which summarizes key study characteristics, and in Supplementary Material S6, which presents JBI checklist assessments. Overall, three studies were judged high quality, four moderate, and six low methodological quality. The most frequent limitations included the absence of control or comparator groups (particularly in observational and descriptive designs), small or non-random participant samples, and potential confounding factors such as environmental or organizational influences that were not statistically controlled.

Many studies did not use clearly defined data-collection procedures or validated measurement tools. Simulation-based studies demonstrated relatively greater methodological rigor, as their models were calibrated against real data [15, 18]. In contrast, descriptive and case-based studies [21, 22] provided contextual insights but were hampered by their non-experimental nature and limited methodological detail. Empirical field evaluations conducted under real conditions [17, 23] provided important evidence from a practical perspective, although they varied in terms of analytical transparency and external validity.

No randomized controlled trials (RCTs) were found; therefore, the focus of the methodological evaluation was on the appropriateness of the design, the reliability of the data, and the clarity of reporting, rather than on experimental comparisons. Detailed quality appraisal findings are available in Supplementary Materials S5, S6, and the ROBINS-I–based bias and narrative-synthesis results as illustrated. (see Figures 3, 4).

Through the studies, the analysis identified several thematic categories of preventive measures. Key findings are summarized below, with illustrative examples. An evidence-mapping matrix was also developed to visualize the distribution of study designs, intervention strategies, and reported outcomes across the included studies, as described in Supplementary Material S7. This provides a structured visualization of preventive interventions for crowd safety and management. The map organizes studies along two primary dimensions: intervention (rows), such as “Evacuation Algorithms and Simulation,” “Spatial Planning,” and “Inter-organizational Communication,” and outcomes (columns), like “Crowd Safety,” “Evacuation Efficiency,” and “Situational Awareness.” Each colored circle represents one or more studies that have investigated a specific intervention-outcome pair, with the color indicating the study design: green for Simulation (SIM), light blue for Experimental (EXP), dark blue for Observational (OBS), and purple for Review/Practice Survey (REV). It also reflects the confidence level of supporting evidence (see Supplementary Material S8).

The Risk of Bias In Non-randomized Studies-of Interventions (ROBINS-I) evaluation of seven studies revealed a range of methodological quality. Only one study (Darsena et al., [17]) was judged to have a low overall risk of bias due to its robust design and clear outcome measurement. Two studies (Wang et al. [15] and Verma and Bains [13]) were classified as moderate risk, primarily due to concerns regarding confounding and deviations from intended interventions in field or simulation settings. The remaining four studies (Ding et al.; Choi et al.; Zhu et al.; and Martella et al., [15, 18, 19, 20]) were found to have a serious risk of bias, often stemming from issues in intervention classification and outcome measurement. Simulation-based research frequently exhibited limitations in generalizability, while experimental and field designs faced challenges with reporting bias and deviations from standard protocols. The results are presented as traffic light plots and summary plots, as shown in Figures 3, 4, respectively, for critical appraisal of the studies (Insert Figures 3, 4 here).

The included evidence from 13 studies demonstrates that multi-modal crowd control frameworks combining structural, technological, and procedural strategies play a vital role in mitigating stampede risk during mass gatherings. Most studies reported measurable improvements in crowd flow indicators, such as reduced density peaks and shortened evacuation time, when interventions were applied in combination rather than in isolation. Quantitatively, simulation-based studies indicated reductions of 30%–40% in peak crowd density and 20%–35% in bottleneck duration under optimized management conditions.

The weight synthesis graph (see Figure 5) summarizes the relative contribution of each study to the overall narrative synthesis [30]. The weight of each study was determined by its methodological robustness, data richness, and relevance to the review objectives. The studies that offer empirical or high-fidelity data, or clearer intervention-outcome relationships, have been given more weight than those that do not. The highest contributions came from Darsena et al. (0.13) and Ding et al. (0.11), followed closely by Lu et al. (0.10) and Choi et al. (0.10) [18]. High-impact field evaluations and frameworks, such as Verma and Bains (0.08) and Singh and Kishore (0.07), also provided substantial evidence to the synthesis [13, 16–18, 24, 25]. Comparative analysis revealed that not all studies contributed equally. Other studies provided useful context, but their quantitative evidence was limited [19, 23]. (Insert Figure 5 here).

The current review adhered to rigorous methodological standards, following PRISMA and PRISMA-S guidelines. A comprehensive search across several databases was conducted, and gray literature was included. Data extraction and quality assessment were conducted by multiple reviewers to analyse and assess quality, thereby minimizing error. By focusing on preventive and operational strategies (rather than descriptive incident reviews), this synthesis provides practical insights for stakeholders.

The majority of included studies were non-randomized or modeling-based, therefore limiting causal inference. Publication bias is possible, as successful interventions may be more frequently reported. Some relevant studies were omitted due to language and accessibility constraints. A relatively small number of studies met the inclusion criteria, reflecting the limited availability of high-quality empirical research in this field, which may constrain the generalizability of findings. The heterogeneity of settings and outcomes precluded meta-analysis; hence, our results are primarily a qualitative synthesis. Due to limited empirical data, simulation-based evidence alone is insufficient to determine the effectiveness of measures unless they undergo implementation and field evaluation before policy adoption.

The findings of this systematic review demonstrate that stampede at mass public gatherings are largely preventable through the implementation of integrated, evidence-based strategies. Rather than relying on isolated interventions, effective crowd safety depends on the synergistic application of engineering controls, operational planning, technology-enabled monitoring, and clear communication protocols. There are key measures consistently associated with improved outcomes, such as establishing one-way pedestrian flows, enforcing density caps, deploying trained personnel at critical bottlenecks, and using real-time surveillance systems to detect emerging hazards. Event organizers, public health authorities, and emergency response agencies should collaborate during the pre-event planning phase to develop context-specific, multimodal safety frameworks. These frameworks should also prioritize practical, scalable actions, such as removing physical bottlenecks, improving signage visibility, and standardizing emergency communication, while remaining adaptable to the unique dynamics of different gathering types.

Future research should prioritize the development of standardized measurement protocols and the rigorous field testing of preventive interventions across diverse cultural and geographic settings. By addressing these gaps and institutionalizing proactive, multi-sectoral approaches, stakeholders can move decisively toward minimizing preventable injuries and fatalities at large public gatherings worldwide.