A closed-loop healthcare processing approach based on deep reinforcement learning

In healthcare, the human body is a controlled input-output system, which generates different observations with the variations of external interventions. The intervention acts as the input, and the output is the phenotype observation that reflects the latent health state of the body system. The objective of healthcare is to determine effective intervention strategies that can nurse an unhealthy human body to a healthy state. With the advances of Internet-of-Things (IoT) and body sensor networks, it becomes convenient to observe the multimedia data of the human body anywhere and anytime. To aid healthcare decision making, we put forward to construct the human body simulators based on deep neural networks (DNNs) for healthcare research. At first, we formulate the model of the human body system based on DNNs. During our analysis, we realize that DNN-based models could simulate practical situations, e.g. some health states are unreachable. Then, we combine deep reinforcement learning (DRL) with conceptual embedding techniques to explore effective healthcare strategies for simulated human bodies. We implement a virtual human body simulator, which can take interventions and represent its hidden states by high-dimensional images, and a DRL-based treatment module, which can diagnose latent health state through the image observations and choose interventions to nurse the simulated body to a target state. By combining the body simulator and treatment module, we create a dynamic closed-loop for healthcare information processing. Experimental simulations are performed to validate the feasibility of the offered approach.