母体敗血症の治療アプローチを再評価する

タイトル: Fair Reinforcement Learning for Maternal Sepsis Treatment

概要: ObjectivesReinforcement Learning is a branch of artificial intelligence (AI) which has the potential to support significant improvement in patient care. There is concern that such approaches may reinforce existing biases within patient groups. Understanding discrimination in AI models is important for building trust and ensuring fair and safe use. We explore the fairness of a published reinforcement learning model, used to suggest drug dosages for sepsis treatment of patients in critical care, on whether it safe to use with maternal sepsis patients. MethodsWe evaluate the current model using by a) comparing the results for a group of patients with maternal sepsis against a matched control group and b) using random forests to explore feature importance in the model. ResultsOur results show that the original clinicians decisions and model suggestions were similar across cohorts. Our feature importance ranking shows high variance for many of the features. DiscussionIn medical settings, different subgroups may have specific clinical needs and require different treatment however, in the absence of a clinical consensus on the most appropriate treatment, AI algorithms that give consistent treatment to patients regardless of subgroup could be judged as the safest and fairest option. ConclusionOur experiments showed that the evaluated model gave the same treatment to maternal and non-maternal sepsis patients. The methods developed for evaluating fair reinforcement learning may be more generally applicable to understanding how clinical AI tools can be used for safely and fairly. What is already known on this topicThe use of reinforcement learning to suggest drug dosages for sepsis patients in critical care is a well-researched area, with multiple open-source models available. It has not previously been considered whether these models can be used on maternal sepsis patients. What this study addsThe model studied behaves consistently on maternal and non-maternal sepsis patients, and appears to suggest an increased use of vasopressors compared with historical actions. How this study might affect research, practice or policyThis study shows that it is possible to design models which are consistent across maternal and non-maternal sepsis patients, suggesting that a single model may be appropriate across a variety of patients with sepsis.