TLDR: A new AI-driven simulation platform called MasTER, utilizing deep reinforcement learning, significantly enhances patient transfer decisions and resource utilization during mass casualty incidents. A user study demonstrated that AI assistance drastically reduced completion times and mortality rates, while increasing resource match rates. Crucially, the system enabled non-experts to achieve expert-level performance, highlighting its potential for both training and real-world emergency response.
Mass casualty incidents (MCIs), such as natural disasters or large-scale accidents, pose immense challenges to healthcare systems, often overwhelming their capacity and demanding swift, accurate decisions under extreme pressure. These events require commanders to make critical choices regarding patient assessment, triage, and transfer to the most suitable hospitals, all while coordinating multiple teams and managing complex factors like patient injuries, hospital resources, and transportation logistics.
Addressing these critical needs, researchers have developed and validated a novel deep reinforcement learning (DRL)-based AI agent designed to optimize patient transfer decisions during simulated MCIs. This AI agent is integrated into a web-accessible command dashboard called MasTER (Mass-Casualty Trauma and Emergency Response), which serves as a simulation platform for MCI management.
How the AI System Works
The core of this innovation is an AI agent trained using Deep Reinforcement Learning, specifically the Proximal Policy Optimization (PPO) algorithm. This agent learns optimal behaviors through extensive interactions with 10,000 simulated MCI scenarios, covering diverse casualty volumes, injury patterns, and hospital resource configurations. It prioritizes patient survival probability while also considering transport duration, hospital capacity constraints, and specialized care requirements. The MasTER platform allows users to assess trauma patients, their injury severity, and potential destination hospitals, including travel times and available resources like ICU beds, operating rooms, and ventilators.
Evaluating Performance: A User Study
To evaluate the effectiveness of MasTER, a controlled user study was conducted with 30 participants, including 6 trauma experts and 24 non-experts. Participants engaged in 20-patient (Standard level) and 60-patient (Complex level) MCI scenarios set in the Greater Toronto Area. The study compared three interaction approaches: human-only decision-making, human-AI collaboration (where participants could request and accept/decline AI suggestions), and AI-only autonomous decision-making.
Key Findings and Impact
The results of the study were compelling, demonstrating that increased AI involvement significantly improved decision quality and consistency:
- Improved Efficiency: In Standard scenarios, Human+AI participants completed tasks 25.49% faster than Human-only. This improvement was even more pronounced in Complex scenarios, with a 45.35% reduction in completion time.
- Reduced Mortality: AI assistance led to an 85.71% reduction in simulated mortality rates for Standard scenarios and a 55.44% reduction for Complex scenarios.
- Better Resource Matching: Match rates for patient-to-resource allocation improved by 8.78% in Standard scenarios and 14.36% in Complex scenarios with AI assistance.
- Empowering Non-Experts: Remarkably, non-experts assisted by AI achieved performance levels comparable to, and in some cases even surpassing, unassisted trauma experts across all metrics. For instance, non-experts with AI assistance in Complex scenarios had lower mortality rates and higher match rates than experts working alone.
- Reduced Workload and High Usability: Qualitative assessments showed a 50.7% reduction in perceived workload for the Human+AI condition, and the MasTER system received an exceptional System Usability Scale (SUS) score of 87.87, placing it in the 95th percentile of evaluated systems.
The DRL model even performed better autonomously than with human intervention in some cases, suggesting the potential for algorithmic decision-making to outperform human-in-the-loop approaches in time-critical situations. This novel application of DRL to MCI management represents a significant advancement in emergency medicine, offering a scalable and cost-effective solution for training and real-world decision support.
Also Read:
- Optimizing Industrial Scheduling: A Novel DRL Environment for Flexible Job-Shops
- Guiding AI Decisions: How Stories Shape Reinforcement Learning
Future Directions
While the validation relies on simulated data, the findings establish the strong potential for AI-driven decision support to enhance both MCI preparedness training and real-world emergency response management. Future work will focus on real-world implementation, addressing algorithm transparency, and integrating with existing healthcare systems and electronic health records. For more detailed information, you can read the full research paper here.
