Data Standards and Benchmarking: Toward Reproducible Artificial Intelligence in Sleep Medicine

This review outlines the rapid evolution of artificial intelligence in sleep medicine, particularly highlighting the role of machine learning in analyzing inherently digital polysomnography data. While machine learning offers time-saving automated scoring, the field struggles with reproducibility and generalizability across varying clinical settings due to domain shift. To establish a standardized framework, this text emphasizes the utilization of public datasets, open models, and common benchmarks. In addition, it details advanced algorithmic approaches, such as domain adaptation and generalization, alongside foundational best practices in signal preprocessing and data splitting to ensure robust performance and prevent data leakage.