Health status prediction of live-line workers based on body state perception is an important prediction task in live-line work. It is easily affected by insufficient data. In addition, live working domain knowledge has shown an important role in improving the performance of health status prediction models, but existing methods cannot make full use of these domain knowledge. Firstly, the correlation between the health status of live-line workers is modeled, and the health status information of workers is constructed as a graph. Then, through the graph convolution module, the spatial features between the health codes of the live operators in each monitoring are obtained on the graph structure. Finally, it is used to model the relationship between the health status characteristics and the multi-level live working monitoring data. The experimental results on two status data sets which can more effectively guarantee the health status of live workers and improve the safety of live work.