Learning a Cost-Effective Strategy on Incomplete Medical Data.

Deep learning techniques have shown remarkable success in classification tasks. However, the prerequisite for high accuracy is that data is easily accessible, which is unrealistic since most features come at a cost. Under a medical scenario, each feature is associated with a medical test that costs a certain amount of money. And doctors would ask patients to do consecutive tests until they are confident enough to make a final diagnosis, whereas the overall cost incurred during the process is often ignored. In this paper, we propose to learn a cost-effective strategy which at the same time hastens the decision process. As is often the case where both medical records and initial feature values of a new patient are incomplete, we design a framework consisting of two components, the oracle classifier and the feature selector, to tackle the challenges. The classifier incorporates a sequence encoder that can handle any set of feature values in various sizes. And the selector efficiently learns the cost-effective strategy based on the state-of-art reinforcement learning techniques. Experimental results have shown that under the same classification accuracy, our strategy is superior to other related approaches in terms of the overall cost.