RSD: A Reinforced Siamese Network with Domain Knowledge for Early Diagnosis

The availability of electronic health record data makes it possible to develop automatic disease diagnosis approaches. In this paper, we study the early diagnosis of diseases. As being a difficult task (even for experienced doctors), early diagnosis of diseases poses several challenges that are not well solved by prior studies, including insufficient training data, dynamic and complex signs of complications and trade-off between earliness and accuracy. To address these challenges, we propose a <u>R</u>einforced <u>S</u>iamese network with <u>D</u>omain knowledge regularization approach, namely RSD, to achieve high performance for early diagnosis. The RSD approach consists of a diagnosis module and a control module. The diagnosis module adopts any EHR Encoder as a basic framework to extract representations, and introduces two improved training strategies. To overcome the insufficient sample problem, we design a Siamese network architecture to enhance the model learning. Furthermore, we propose a domain knowledge regularization strategy to guide the model learning with domain knowledge. Based on the diagnosis module, our control module learns to automatically determine whether making a disease alert to the patients based on the diagnosis results. Through carefully designed architecture, rewards and policies, it is able to effectively balance earliness and accuracy for diagnosis. Experimental results have demonstrated the effectiveness of our approach on both diagnosis prediction and early diagnosis. We also perform extensive analysis experiments to verify the robustness of the proposed approach.