Sleep stage identification is a critical component in the diagnosis and treatment of sleep disorders, yet traditional manual annotation is inefficient, and existing automatic staging models still suffer from limitations in long-range dependency modeling and cross-dataset generalization. To address this, this paper proposes EMBSleep—a novel sleep staging framework based on time-frequency analysis and efficient temporal modeling. The model first transforms polysomnography (PSG) signals into time-frequency representations using short-time Fourier transform (STFT) and employs frequency band energy normalization to eliminate individual differences. Building on this, an efficient multi-scale attention module (CEMA) is designed, integrating parallel local convolution and global context pathways to enable fusion and extraction of multi-band features. Leveraging the time-frequency features extracted by the CEMA module, Mamba performs global modeling along the temporal dimension, linking local time-frequency patterns (e.g., sleep spindles, K-complexes) with macro sleep cycles (REM-NREM alternation), thereby addressing the issue of long-range dependency loss in traditional CNN+RNN approaches. Sleep stage transitions often involve ambiguous boundaries, such as the shift from N1 to N2. Here, BiLSTM employs a gating mechanism to weight and fuse features from adjacent time steps, reducing classification jitter and improving staging continuity. EMBSleep demonstrates exceptional cross-dataset generalization capabilities. Through leave-one-out cross-validation, the model’s superior generalization performance across different individuals is further validated. The EMBSleep model adopts a unique architecture specifically designed for multi-channel physiological signal (PSG) classification and has been comprehensively evaluated on three widely recognized sleep datasets. Experimental results show that the model achieves accuracies of 87.2%, 85.1%, and 85.2% on the Sleep-EDF-20, Sleep-EDF-78, and ISRUC-S3 datasets, respectively, surpassing existing state-of-the-art methods. EMBSleep’s high accuracy and robust generalization capabilities highlight its superior performance in the field of sleep stage classification.