Traditional relation extraction methods primarily focus on extracting relations between entities. However, in some texts, a relation can be part of other relations, leading to the creation of nested relations. Research on nested relation extraction is relatively limited, with most existing approaches employing a layer-by-layer extraction process, which often suffers from error propagation. In this paper, we discover that entities within a nested relation are directly or indirectly connected to other entities. Once these two relations of entity pairs are identified, nested relations can be directly decoded using techniques such as maximal clique identification and topological sorting, effectively avoiding error propagation. Based on the above strategy, we propose a new extraction model called EPNR. This model combines word representations into entity representations using the attention mechanism. Then, it employs a conditional layer normalization mechanism to convert these entity representations into richer entity pair representations for classification and subsequent decoding of nested relations. We constructed a Medication dataset containing nested relations and performed extensive experiments on this dataset and 2 widely used benchmark datasets, WebNLG and CMedCausal. The EPNR model achieves outperformance on all three datasets. Codes available at: https://github.com/taskNLP/EPNR