Personalized Scorer Modeling Improves Sleep Stage Classification Accuracy

Researchers have developed a novel learning-based framework, called Personalized Scorer Modeling (LBH), aimed at improving the accuracy and reliability of sleep stage classification. Recognizing the inherent variability among human experts in scoring sleep stages, this approach leverages multi-scored datasets to construct more robust reference labels. The LBH framework models the unique, stage-specific behavior of each expert scorer using confusion matrices derived from machine learning models. These matrices are then normalized to estimate the probability of true sleep stages given each scorer's label, with probabilities aggregated to assign a final, more reliable label for each 30-second epoch. Evaluated on public datasets (DOD-H and DOD-O) using various classifiers and signal combinations, LBH consistently outperformed traditional methods that rely on a single reference hypnogram. The best results, achieved with random forest and EEG+EMG signals, demonstrated significant improvements in accuracy, precision, and F1-score, suggesting a powerful method for handling inter-scorer variability in medical diagnostics.