Risk-Aware Safety Intervention in Construction: A Deep Reinforcement Learning Framework for Optimizing Managerial Decisions Using OSHA Incident Data

Introduction:

Construction sites are high-risk environments where inconsistent safety decisions can lead to severe injuries and fatalities. While prior studies have employed rule-based systems and traditional machine learning classifiers to predict incident outcomes, these approaches are limited to risk identification and do not prescribe optimal intervention actions. This study proposes a data-driven deep reinforcement learning framework that not only predicts injury risk but also recommends the most cost-effective managerial intervention using real-world OSHA incident records comprising over 53,550 cases.

Method:

The framework converts each incident into a feature vector combining event-type indicators, weather variables, and text embeddings from incident narratives. Logistic regression models, selected for their interpretability and suitability for binary risk classification, estimate fatality and injury severity risks using a 75:25 train-test split, integrated into a reinforcement learning environment with five actions: continue work, provide PPE or toolbox talks, increase supervision, conduct inspection or lockout, and stop work immediately. Three algorithms, Deep Q-Network (DQN), Advantage Actor-Critic (A2C), and Proximal Policy Optimization (PPO), are trained to minimize combined injury severity cost and intervention disruption cost.

Result:

The predictive layer achieved 71.31% test accuracy for fatality risk and 53.00% for injury severity. Among the tested policies, PPO achieved the lowest mean total cost of 37.26, followed by A2C (37.29) and DQN (37.36). Compared with the no-action baseline (48.45), PPO reduced the expected total cost by 23.10% and outperformed the random policy by 13.49%. The near-identical performance across algorithms validates the framework's robustness, while the significant margin over baselines confirms that data-driven intervention substantially outperforms conventional approaches.

Conclusion:

This study contributes by reframing safety intervention as a reinforcement learning problem, bridging predictive analytics and prescriptive decision support. The framework enables consistent intervention by recommending stronger actions only when predicted risk justifies operational disruption. Future research should explore real-time sensor integration for dynamic construction environments.