Optimizing automated sleep stage scoring of 5-s mini-epochs: a transfer learning study

Clinical Snapshot

Key Findings

• P Value: p < 0.05 for F1-score improvement

• Effect Size: F1-score improvement of 0.07 (9.5% relative improvement)

• Primary Outcome: F1-score improvement from 0.74 to 0.81 for optimized U-Sleep model

• Nnt Or Sensitivity: Sensitivity and specificity not reported; F1-score of 0.81 indicates good overall diagnostic performance

• Confidence Interval: Not reported

Clinical Application

Requires specialized deep learning infrastructure and validation in clinical sleep laboratories Relevant to Australian sleep medicine services; would require validation against local sleep scoring practices and integration with existing polysomnography systems used in Australian sleep laboratories Sleep laboratory patients undergoing polysomnography for sleep disorder evaluation

Abstract

STUDY OBJECTIVES: Conventional sleep staging relies on 30-s epochs, potentially concealing transient sleep stage intrusion and reducing precision. Building on our previous study of mini-epochs, we investigated whether U-Sleep, an existing automatic deep learning-based sleep staging model with high performance in epochs, could be optimized to similar performance level in 5-s mini-epoch scoring, thereby enabling more detailed sleep characterization. METHODS: We created a dataset of 48 000 human-scored 5-s mini-epochs from 100 polysomnographies. We compared mini-epochs to human-scored epochs before U-Sleep was optimized using transfer learning and evaluated on a test set. Model performance was assessed using F1-scores, confusion matrices, stage distributions and transition rates comparing scorings of the original U-Sleep before, and the optimized U-Sleep after transfer learning to human-scored mini-epochs. RESULTS: Compared to human-scored epochs, human-scored mini-epochs captured significantly more transitions (1.70/min vs. 0.21/min, p < .001), and significantly more wake (8.4 per cent vs. 5.4 per cent), N1 (7.2 per cent vs. 5.4 per cent), and N2 (51.8 per cent vs. 40.9 per cent), less N3 (15.4 per cent vs. 25.2 per cent), and REM sleep (16.7 per cent vs. 23.0 per cent) (all p < .001). Optimizing U-Sleep improved its performance significantly from F1 = 0.74 to F1 = 0.81 (p < .05) and gave increased transition rates in the test set (original U-Sleep: 1.06/min, optimized U-Sleep: 1.34/min, human-scored mini-epochs: 1.70/min). Stage distributions did not differ between optimized U-Sleep's scorings and human-scored mini-epochs. CONCLUSION: After optimization, U-Sleep performance in mini-epochs matched the high performance levels previously reported in both human and automated 30-s epoch scoring. This demonstrates the feasibility of precise, automated high-resolution sleep staging. Future work should include external validation and application to full-night recordings. Statement of Significance Conventional 30-s epochs limit temporal resolution in sleep staging and may conceal transient intrusions of wake or sleep stages. However, no validated methods are available for high-resolution scoring. In this study, we trained and validated the state-of-the-art deep learning model U-Sleep for accurate automatic 5-s mini-epoch scoring using a large dataset of human-scored mini-epochs. The optimized model achieved a high performance, matching levels from previously reported automatic and human epoch scoring. Compared to epoch scoring, mini-epochs captured significantly more stage transitions, supporting their ability to uncover sleep dynamics that are otherwise lost. Our findings show the potential of high-resolution sleep staging for more detailed characterization of sleep architecture and demonstrate the feasibility of precise, automatic mini-epoch scoring.

References

1. 1.Follin, L. F., Christensen, J. A. E., Vevelstad, J., Juvodden, H. T., Viste, R., Hansen, B. H., Perslev, M., Kaufmann, T., Zahid, A. N., & Knudsen-Heier, S. (2026). Optimizing automated sleep stage scoring of 5-s mini-epochs: a transfer learning study. Sleep, zsaf393. https://doi.org/10.1093/sleep/zsaf393