Researchers built 4,400 synthetic EMS conversations to train AI that diagnoses patients faster

A team created a dataset of realistic emergency medical service dialogues by having AI agents generate conversations from real patient records, then checking them for accuracy. This lets hospitals train AI systems to diagnose patients during live calls instead of after the fact.

Emergency medical diagnosis is a real-time problem: paramedics need to know what they're treating while they're treating it. Until now, there weren't enough real multi-person EMS conversations recorded and annotated to train AI on this specific task. This dataset is synthetic but grounded in actual patient data, which means hospitals can now start building and testing diagnostic AI without waiting years for enough real conversations to accumulate. The payoff is concrete: the paper shows that models trained on this synthetic data get faster and more accurate at committing to diagnoses during active calls, which matters because paramedics have seconds, not minutes.