Fig. 1. Overall study concept diagram.

The development and evaluation process of the reinforcement learning models is outlined. A patient’s status at each time step (every 2 min) was modeled as the state, and the decision to continue on-scene resuscitation or initiate transport was considered the action. The on-scene ROSC hazard rate of each individual at each time step was used to derive the intermediate rewards and the actual survival to hospital discharge result of each patient was reflected in the terminal reward. The policies were evaluated using fitted Q-evaluation. s state, a action, ROSC return of spontaneous circulation, EMS emergency medical services, OHCA out-of-hospital cardiac arrest, FAMD Factor Analysis of Mixed Data.