Background: Sleep spindles are important biomarkers of sleep physiology and neurological function. Automated spindle detection remains challenging due to inter-subject variability, class imbalance, inconsistent expert annotations, and limited cross-dataset generalization.

New Method: We propose an attention-gated dual-input one-dimensional U-Net for subject-independent sleep spindle detection from single-channel EEG. The framework formulates spindle detection as a dense temporal segmentation task by jointly learning from raw EEG and sigma-band representations, while incorporating imbalance-aware optimization and annotation-aware evaluation.

Results: Experiments on the MASS-SS2 dataset using subject-wise five-fold cross-validation achieved a strict event-level F1-score of 0.805 ± 0.031, a tolerant event-level F1-score of 0.846 ± 0.017, and a sample-level ROC-AUC of 0.933 ± 0.015. External zero-shot validation on the DREAMS dataset yielded a strict event-level F1-score of 0.476 and a sample-level ROC-AUC of 0.873 without retraining.

Comparison with Existing Methods: The proposed framework achieved competitive performance relative to recent deep learning methods while providing additional capabilities, including dual-input feature learning, annotation-aware evaluation across multiple expert references, and external cross-dataset validation, which are rarely investigated simultaneously in existing spindle detection studies.

Conclusions: The proposed framework provides robust and reproducible sleep spindle detection with competitive performance and transferable spindle representations. The findings highlight the importance of annotation-aware benchmarking and external validation for developing reliable AI-based sleep analysis systems.