Explainability of Agent Behavior in Clinical Decision Support Systems

The article considers the task of improving the explainability of agent behavior in clinical decision support systems using multi-agent reinforcement learning, taking into account the behavioral characteristics of individual agents. Special attention is paid to the application of the proximal policy optimization (PPO) algorithm, which is used to simulate the interaction of artificial and natural intelligence agents. In addition, the importance of taking into account the behavior-al characteristics of patients is considered, which is achieved through the use of the developed framework «beliefs-desires-intentions» (BDI). The framework allows you to model patient behavior more accurately, taking into account their individual preferences and needs, which contributes to improving the quality of medical care. To increase the explainability of agent behavior and ensure transparency of decisions, the use of a large language model (LLM) is pro-posed. To assess the quality of the framework as a means of ensuring the requirements of XAI (explicable artificial intelligence), a set of criteria has been developed and its representativeness has been demonstrated. The assessment was carried out by comparing the explanations generated by four popular language models for seven gameplay launches with different initial data. The proposed approach to ensuring explainability in multi-agent simulations opens up opportunities for further development and implementation of such systems in clinical practice.