Abstract
Abstract
Objective
With advancements in AI, automated sleep scoring offers the potential for fast, precise, and cost-effective analysis. This study aimed to evaluate the accuracy, robustness, and clinical agreement of four algorithms compared to manual PSG in athletes to determine their suitability for real-life use.
Approach
Data from 75 athletes (204 polysomnography recordings) were retrospectively collected from studies conducted in our laboratory. Manual scoring was compared to four automated sleep staging algorithms: Luna, U-Sleep, YASA, and GSSC. Performance was evaluated based on accuracy (per stage, macro and weighted F1), sensitivity (confusion matrices), robustness (signal quality: R 2 ; individual differences: beta coefficients) and clinical relevance (prediction of total sleep time, sleep onset latency and wake after sleep onset).
Main results
All algorithms demonstrated strong agreement with manual scoring, with macro F1 scores of 0.76 (GSSC), 0.71 (U-Sleep), 0.70 (YASA), and 0.69 (Luna). GSSC was the most accurate and robust across sport, physiological, behavioral, and anthropometric variables. It also reliably predicted total sleep time, sleep onset latency and wake after sleep onset, showing minimal bias and narrow limits of agreement. GSSC and Luna were the most resilient to signal degradation (R² = 0.08), compared to YASA (R² = 0.16) and U-Sleep (R² = 0.29).
Significance
Automated sleep staging algorithms demonstrated accuracy comparable to typical inter-rater agreement, supporting their broader use in athletes. GSSC emerged as the most accurate and robust option, particularly resilient to signal quality degradation, a common challenge in ecological sport settings. By reducing scoring time and removing the need for expert raters, these algorithms could make polysomnography more accessible for both routine sleep monitoring and large-scale research, enabling more individualized recovery strategies and deeper insights into how sleep architecture influences athletic performance.