Abstract
The medical automatic diagnosis system aims to imitate human doctors in thereal diagnostic process. This task is formulated as a sequentialdecision-making problem with symptom inquiring and disease diagnosis. In recentyears, many researchers have used reinforcement learning methods to handle thistask. However, most recent works neglected to distinguish the symptom inquiringand disease diagnosing actions and mixed them into one action space. Thisresults in the unsatisfactory performance of reinforcement learning methods onthis task. Moreover, there is a lack of a public evaluation dataset thatcontains various diseases and corresponding information. To address theseissues, we first propose a novel method for medical automatic diagnosis withsymptom inquiring and disease diagnosing formulated as a reinforcement learningtask and a classification task, respectively. We also propose a robust andadaptive method to align the two tasks using distribution entropies as media.Then, we create a new dataset extracted from the MedlinePlus knowledge base.The dataset contains more diseases and more complete symptom information. Thesimulated patients for experiments are more realistic. Experimental evaluationresults show that our method outperforms three recent state-of-the-art methodson different datasets by achieving higher medical diagnosis accuracies with fewinquiring turns.