Reinforcement learning
Reinforcement learning has recently delivered breakthroughs in the
biomedical field. This strategy is characterized by an iterative process
that aims to take actions that deliver maximum reward based on a defined
objective function. This is comparable to nature, where animals have
learned to interpret signs such as hunger as negative, whereas satiety
after food intake is seen as positive reinforcement. When animals learn
how to behave to gain optimal positive reinforcement, they show
reinforcement learning. Applications within medicine have indicated its
potential. For example, an “AI Clinician” algorithm has been developed
to improve the treatment of sepsis by suggesting the personalized
treatment of intravenous fluids and vasopressors. While still requiring
prospective validation, an independent validation cohort was used to
assess this algorithm, showing that mortality was lowest when
clinicians’ treatment policy was close to AI recommended policy and
higher when deviating from it.