Abstract
Clinical diagnosis guidelines aim at specifying the steps that may lead to adiagnosis. Inspired by guidelines, we aim to learn the optimal sequence ofactions to perform in order to obtain a correct diagnosis from electronichealth records. We apply various deep reinforcement learning algorithms to thistask and experiment on a synthetic but realistic dataset to differentiallydiagnose anemia and its subtypes and particularly evaluate the robustness ofvarious approaches to noise and missing data. Experimental results show thatthe deep reinforcement learning algorithms show competitive performancecompared to the state-of-the-art methods with the added advantage that theyenable the progressive generation of a pathway to the suggested diagnosis,which can both guide and explain the decision process.