Due to the development of information technology, data in various fields often contain a large number of redundant features, which leads to the degradation of classification model performance. Feature selection is a data preprocessing technique that effectively removes redundant features, however, existing feature selection methods are unable to select effective features while ensuring high classification accuracy of the model. In order to solve this problem, a graph convolutional network-based reinforcement learning for feature selection is proposed. A deep Q-network is used as the basic framework to map the feature selection problem into a Markov decision process. Firstly, a state representation method based on graph convolutional network is designed to convert feature subsets into graph structures to capture inter-feature relationships efficiently. Then, a reward function considering feature importance, feature-to-feature correlation and classification performance is designed to guide the agent to select high Q-value features so that the obtained feature subset combines several aspects of performance. Experiments are performed on 14 public datasets and stroke screening datasets, and the results show that compared with the existing deep Q-network feature selection method, the accuracy of the proposed method on Congress dataset and the stroke screening dataset is 99% and 85%, which is an improvement of 11% and 3.6%, respectively, and verifies the validity and feasibility of the method.