Automatic　sleep　staging　is　an　essential　tool　in　clinical　research　and　medicine,　designed　to　address　people＇s　sleep　problems　or　disorders.　Nevertheless,　extracting　features　from　multiple　EEG　segments　and　achieving　precise　and　automated　classification　of　sleep　stages　remains　a　challenge.　To　address　this,　a　novel　sleep　staging　framework　is　introduced,　leveraging　a　sequence-to-sequence　architecture　tailored　for　sleep　analysis.　The　time-frequency　representations　derived　from　EEG　epochs　are　converted　into　sequential　images　via　parallel　filter　layers.　Subsequently,　a　multi-path　structure　enhances　the　handling　of　sequential　data.　A　primary　path　integrates　a　Graph　Neural　Network　(GNN)　to　better　identify　correlations　among　EEG　segments,　facilitating　faster　model　convergence　and　enhanced　predictive　performance　with　fewer　training　epochs.　On　the　auxiliary　path,　sequential　characteristics　are　captured　by　a　bidirectional　recurrent　network.　The　output　from　both　paths　is　combined　into　the　ultimate　eigenvector　through　a　fusion　mechanism.　The　final　classification　is　performed　using　a　softmax　layer,　generating　an　array　of　predicted　signs.　Empirical　evaluations　reveal　that　applied　to　the　Sleep-EDF　dataset,　the　model　proposed　in　this　paper　outperforms　other　models,　achieving　an　accuracy　of　approximately　87.6%.　©　2025　SPIE.