Journal of Reliable Intelligent Environments (2019) 5:105–113 https://doi.org/10.1007/s40860-019-00084-z ORIGINAL ARTICLE Risk management for nuclear medical department using reinforcement learning algorithms Giovanni Paragliola 1 · Muddasar Naeem 1,2 Received: 28 September 2018 / Accepted: 3 May 2019 / Published online: 14 May 2019 © Springer Nature Switzerland AG 2019 Abstract Modern medical software systems are often classified as medical devices and governed by regulations which require stringent risk safety activities to be implemented to minimize the occurrence of risky events. This paper proposes a reinforcement learning (RL) based approaches for training a software agent for risk management of medical software systems. The goal of RL agent is to avoid that a patient enters in dangerous and undesirable states. At the same time agent must be able to reach on a safe state or an exit in a minimum interval of time. RL based system is also able to guide a patient to a safe path if he/she mistakenly enter into risk or undesirable states. Keywords Reinforcement learning · Risk management · Software agent · Machine learning · e-Health 1 Introduction The standards of medical software systems or devices depend on a number of factors like vendor management, risk man- agement, design control, etc. Companies have been facing challenges about the safe and effective use of medical devices for humans and they have to comply with International Orga- nization of Standardization (ISO) quality system rules and regulation for design and development of medical devices. Similar to other risk management domains such as: oper- ational; financial; hazard; strategic; human capital; legal and regulatory; environmental and infrastructure based hazard; technological, risk management is an important part of medi- cal software system development. The complexity of medical examination systems causes adverse events if left uncon- trolled [12]. The major objective of risk management is to mitigate or minimize device failure and risks environment. It ensures the reliability of product and device cause no hazards to operators, patients or environments. B Giovanni Paragliola giovanni.paragliola@icar.cnr.it Muddasar Naeem muddasar.naeem@icar.cnr.it; muddasar.naeem@uniparthenope.it 1 ICAR-CNR, Naples, Italy 2 Department of Engineering, University of Parthenope, ICAR-CNR, Naples, Italy To address risk management tasks, ICT has been propos- ing many solutions along the last decades ranging from agent-based solutions [2,4,11], formal methods approach [1,3,6,8] to machine learning-based methodologies [5,7]. The focus of the present work is on the risk management for medical software systems and steps to safely implement it during its application in any healthcare setup using rein- forcement learning (RL) techniques. RL is a type of machine learning which targets to learn by repeatedly interacting with the environment and contin- uously improve its learning based on experience and finally start to perform optimally in a given situation. The scenario is that a patient entering in a healthcare envi- ronment, more precisely a patient interacts with a medical environment for medical examination may not know every- thing about its surrounding and there are chances to enter in a dangerous situation. Similarly, an RL agent knows nothing about the world before interaction with it. Therefore, we can train a RL agent to interact with the medical environment and learns how to act safely and effec- tively. The use of RL techniques for risk management in the medical environment would improve the performance of such systems and devices. A medical system consists of many components and a patient seeking for medical examination may have to pass through many process and stages. Some stages or states are useful while some are dangerous or undesirable. 123