Can a reinforcement learning model identify high-risk hemodynamic treatments in septic shock patients before poor outcomes occur?
A reinforcement learning model can identify high-risk hemodynamic treatments in septic shock patients up to 24 hours before death, potentially alerting clinicians to reconsider therapy.
The resuscitation of septic shock with hemodynamic support to increase the total circulating volume and cardiac output is essential in clinical rescuing.This study aimed to raise high-risk awareness in hemodynamic treatment with reinforcement learning for septic shock patients instead of finding the best.Retrospective data from 7792 septic shock patients (mortality of 24.7%) were used.Data were coded as multivariate discrete-time series up to 72 h with 4-h time steps.The medical treatments of interest are the sum of intravenous fluids and a maximum dose of vasopressors.State spaces are constructed using an autoencoder network.Two separate double deep Q-Networks are trained to produce the value estimates of the embedded patient states administrated given treatments to assess the risk of transitioning to poor outcomes.Results reported that when we set a threshold of -0.16 for alarming the high-risk flag, 8.3% of treatments administered to nonsurvivors were alarmed 24 h before the outcome of death with 0.7% false alarms that misclassified the patients as near death.The global and individual trajectories of clinical variables around the first raised flag indicate the method's effectiveness.This could help warn possible high-risk treatments and help clinicians pay more attention to the alarmed patients.
Yang et al. (Sat,) studied this question.