Enabling risk-aware Reinforcement Learning for medical interventions through uncertainty decomposition

Abstract

Reinforcement Learning (RL) is emerging as tool for tackling complex control and decision-making problems. However, in high-risk environments such as healthcare, manufacturing, automotive or aerospace, it is often challenging to bridge the gap between an apparently optimal policy learned by an agent and its real-world deployment, due to the uncertainties and risk associated with it. Broadly speaking RL agents face two kinds of uncertainty, 1. aleatoric uncertainty, which reflects randomness or noise in the dynamics of the world, and 2. epistemic uncertainty, which reflects the bounded knowledge of the agent due to model limitations and finite amount of information/data the agent has acquired about the world. These two types of uncertainty carry fundamentally different implications for the evaluation of performance and the level of risk or trust. Yet these aleatoric and epistemic uncertainties are generally confounded as standard and even distributional RL is agnostic to this difference. Here we propose how a distributional approach (UA-DQN) can be recast to render uncertainties by decomposing the net effects of each uncertainty . We demonstrate the operation of this method in grid world examples to build intuition and then show a proof of concept application for an RL agent operating as a clinical decision support system in critical care.

Assuring the safety of AI-based clinical decision support systems: a case study of the AI Clinician for sepsis treatment

“…An emerging new avenue in the field is to augment AI models so that they can quantify their own confidence or uncertainty over their recommendations. 19 Going forward, it may be helpful to algorithmically combine the communication of uncertainty that a system has about itself, which reflects the risk of unwanted behaviour as we have shown in other domains of risk-aware control by medical devices, 20 with its safety features, that we have shown here.…”

Optimal discharge of patients from intensive care via a data-driven policy learning framework

“…, (x L−1 , u L−1 )} generated by following the clinician's policy H ∼ µ CP . Approaches for off-policy evaluation are an active topic of research in the context of related reinforcement learning frameworks (Hanna et al, 2017;Thomas et al, 2015;Festor et al, 2021). It should be noted that the factors driving µ CP may differ from the cost function chosen to derive µ OP , which is why effective calibration is important to objectively compare the two policies.…”

Reinforcement Learning For Survival: A Clinically Motivated Method For Critically Ill Patients

“…Recently, there has been an increased volume of research which try to learn optimal treatment strategies for critically ill and in particular for septic patients (Komorowski et al, 2018;Chen et al, 2019;Raghu et al, 2017;Li et al, 2019;Peng et al, 2018;Festor et al, 2021;Nanayakkara et al, 2022b), using Reinforcement Learning (RL) methods. Given the enormous mortality, morbidity and economic burden (Liu et al, 2014;Rhee et al, 2017;Paoli et al, 2018), the ambiguity regarding optimal treatment strategies and lack of accepted guidelines for treatment (Marik, 2015;Jarczak et al, 2021), such attempts are certainly justified.…”