Fig. 4

Relationship between in-hospital survival rate (z-axis) and differences in IV fluid (x-axis) and VP (y-axis) dosing administered by the models and human clinicians. The model is trained at a step size of 16.5 w, and the model outputs the drug dosing combination. If the number of results corresponding to the combination of drug doses is less than 50, such results will be removed because they are insufficient to explain the survival rate. The smaller the difference is between drug dosing given by the models and human clinicians, the better the survival rate of the sepsis patients. It can also be seen that the distribution map of the survival rate based on the DDPG algorithm is more concentrated than that based on the DQN algorithm. That is, the model based on the DDPG algorithm makes more treatment decisions similar to those of the human clinician than the model based on the DQN algorithm