Multi-Agent Crowd-Risk Prioritization for Large-Scale Events: A Hybrid ML–Agent Approach

Managing crowd safety during large-scale gatherings such as Hajj requires timely assessment of multiple risk factors, including crowd conditions, environmental factors, and health-related indicators. Many existing approaches address these aspects separately, making it difficult to obtain a unified operational view of risk. This study presents a multi-agent framework that combines machine learning with domain-specific agent coordination for crowd-risk prioritization. The framework includes five agents responsible for crowd monitoring, health monitoring, environmental sensing, coordination, and alert generation. To reduce data leakage, variables used in risk-label construction and post-event outcomes were excluded before model training, resulting in 17 real-time observable features. The framework was evaluated using the Hajj and Umrah Crowd Management dataset containing 10,000 simulated records. On the independent test set, XGBoost+SMOTE achieved 47.30% accuracy and 46.22% F1-score; the XGBoost baseline achieved 57.15% accuracy and 45.96% F1-score, and a majority-class baseline that always predicts the dominant class achieved 58.70%, underscoring that overall accuracy is an insufficient metric for this task. Although accuracy decreased after class rebalancing, minority-class sensitivity improved: XGBoost+SMOTE achieved partial High-risk recognition (recall = 0.14, F1 = 0.12), whereas the XGBoost baseline and Cost-Sensitive XGBoost failed to identify any High-risk cases (recall = 0.00). McNemar's test confirmed a statistically significant difference between the XGBoost baseline and the SMOTE-enhanced model (χ² = 69.218, p < 0.001). Cross-validation on the SMOTE-balanced training data produced 59.89% ± 0.20% accuracy and 58.94% ± 0.17% F1-score. A streaming evaluation on 500 held-out test records showed that the Full Framework generated 93 High-risk alerts, compared with 74 from the ML-only configuration and 23 from the Agents-only configuration, a 25.7% increase relative to ML-only operation. A precision check against the internal risk-score proxy found that none of the 19 agent-driven escalations beyond ML-only predictions corresponded to a High-risk label under this proxy; these alerts should accordingly be read as operational risk-prioritization signals rather than confirmed incident detections. All reported figures should further be interpreted in light of the dataset's synthetic origin: because features were algorithmically generated rather than collected from real sensors or incidents, the results are illustrative of the framework's behavior under these specific simulated conditions rather than predictive of real operational deployment. Average processing latency was 4.127 ms per record under sequential single-stream conditions. In addition to the proposed framework, the study presents a diagnostic evaluation approach that combines leakage control, dual association analysis, and direct precision auditing. This approach may also prove useful for evaluating similar ML–agent systems in other safety-critical alerting applications.

@article{Samrgandi2026,
  title     = {Multi-Agent Crowd-Risk Prioritization for Large-Scale Events: A Hybrid ML–Agent Approach},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {7},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Najwa Samrgandi},
  doi       = {10.14569/IJACSA.2026.0170738},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170738}
}