According to ithome.com.tw, Changi General Hospital and IBM built an MR and agentic-AI training model for pediatric mass-casualty drills, launched in January 2026, aiming to train 100 medical staff within nine months.
Why does Singapore need this pediatric emergency training innovation?
In April 2025, a shopping-street fire in Singapore sent a large number of casualties to hospitals: as many as 16 children and 6 adults were injured and rushed for emergency treatment, according to ithome.com.tw. The following month, in May 2025, Singapore's Ministry of Health convened 1,700 personnel for a joint pediatric mass-casualty response drill, the same report states. Those two events — a real fire with 16 injured children, followed a month later by a 1,700-person joint exercise — form the backdrop against which Changi General Hospital's new training program was developed.
How did the Changi General Hospital–IBM collaboration begin?
Changhi General Hospital launched a new project in January 2026, teaming up with IBM to combine Mixed Reality (MR) technology with Agentic AI, building what the hospital describes as a highly immersive and scalable next-generation model for mass-casualty emergency training, per ithome.com.tw. IBM first made the training program public at its Think Singapore event on July 21, 2025. The project is led by Wu Chih-Min (吳志敏), Senior Consultant and Adjunct Assistant Professor in Changi General Hospital's Department of Emergency Medicine, who said that integrating AI, MR technology, and clinical medical expertise offers a new generation of training for preparedness in pediatric mass-casualty events, the report quotes.
How does the MR scenario recreate a pediatric emergency scene?
At the public demonstration, IBM showed a scenario in which as many as 6 pediatric patients required simultaneous treatment, according to ithome.com.tw. To build the underlying 3D environment, Changi General Hospital collected multiple sets of de-identified pediatric emergency treatment data and photographs, and combined them with IBM's existing medical-scene 3D model assets — a process that took 6 months to complete the 3D emergency scenario inside the MR environment, the report states.
How are AI agents used to assess trainees' emergency response?
The evaluation system runs on IBM watsonx Orchestrate, the new-generation AI-agent collaboration platform IBM released in May 2025, using a multi-agent architecture, per ithome.com.tw. The system comprises three specialized agents — a Team Assessment Agent, a Clinical Assessment Agent, and a Summary Agent — plus an Orchestrator Agent that coordinates all of them, the report details. The Clinical Assessment Agent is built with knowledge of multiple pediatric clinical evaluation frameworks, including the ATLS/ABCDE emergency assessment method, the Pediatric Trauma Score (PTS), the Pediatric Assessment Triangle (PAT), and the JumpSTART and SALT mass-casualty triage systems. The platform's underlying large language model is OpenAI's open-source GPT-Oss model, which Changi General Hospital fine-tuned for this project, though the platform can support other third-party LLMs if needed, according to the report.
What is Changi General Hospital's training target and current progress?
The project's stated goal is to train at least 100 medical staff within nine months, building emergency response capability for pediatric mass-casualty events, per ithome.com.tw. As of the report, more than 30 medical staff had already taken part in testing and training, with full-scale training to begin once the system completes its final round of tuning.
Four advantages IBM cites for MR and agentic-AI training over traditional drills
IBM stated that, compared with traditional mass-casualty incident (MCI) drills using mannequins or picture cards, the MR-plus-agentic-AI training approach offers four advantages, according to ithome.com.tw:
| Advantage | What IBM says |
|---|
| Lower resource dependence | Reduces reliance on physical simulation equipment and venue resources, cutting environment setup and preparation time while increasing reusability |
| More immersive environment | Creates a more realistic, immersive clinical training environment |
| Stronger post-exercise review | Strengthens after-action debrief and review of trainee performance |
| More diverse scenario design | Enables a wider variety of training scenarios to be built |
How urgent are real pediatric emergencies on the ground?
A separate case reported in Taiwan illustrates the real-time pressure emergency responders face outside any training environment. In Tanzi District, Taichung, a 2-year-old girl was seriously injured after apparently falling from the 8th floor, according to udn.com. The Taichung City Fire Bureau received the call at 9:11 p.m. and dispatched one vehicle and two personnel from its Tanzi squad, the report states. When paramedics arrived, they found the roughly 2-year-old girl had blood around her nose and mouth, no visible external injuries, and was unconscious; she was taken to Taichung Tzu Chi Hospital. Separately, the Dajia Police Precinct confirmed it had received a 119 emergency notification about the incident, dispatched officers to the scene, and that the victim had already been sent to hospital by ambulance, with the cause of the fall still under investigation, per udn.com.
What this means
Taken together, the evidence cited above shows two very different scales of pediatric emergency preparedness sitting side by side. On one end, Singapore's response infrastructure has been tested at scale: a single fire injured 16 children and 6 adults, prompting a 1,700-person joint drill the following month, and Changi General Hospital's new MR-and-agentic-AI system is now built around a 6-patient simulated scenario with a target of training at least 100 staff within nine months. On the other end, the Taichung case shows what a single pediatric emergency actually looks like in real time — a 9:11 p.m. call, a two-person crew dispatched within minutes, and a child whose condition had to be assessed and acted on immediately, with no time for the kind of structured, multi-agent debrief the Changi-IBM platform is designed to provide after a training session. The scale of institutional preparation described in the Singapore reporting and the single-incident urgency described in the Taiwan reporting are documented in separate sources and are not presented here as connected events, but read together they underline why the JumpSTART and SALT triage frameworks embedded in the Clinical Assessment Agent exist in the first place: real pediatric emergencies, whether involving one child or a dozen, leave responders with minutes, not months, to act.