 This research proposes an end-to-end, multi-scale, subject-adaptive network to improve the accuracy of sleep-stage classification using single-lead electrocardiogram data. The proposed model takes into account the domain shift caused by individual differences and acquisition conditions as well as class imbalance to achieve better results than existing models. Additionally, the model uses a multi-class focal loss to reduce the negative impact of class imbalance and a loss of sequence prediction to help the model judge sleep-stages. This research demonstrates that the proposed model can provide more accurate sleep-stage classification than existing methods with main accuracies of 0.849, 0.827, and 0.868 on three public data sets. Furthermore, the model performs well in cross-data set testing, indicating its generalizability. This article was authored by Ji Wei Zhang and Min Feng Tang.