Researchers' LBH framework improves sleep stage labeling with 86% accuracy

Sleep stage classification is critical for diagnosing sleep disorders, but most AI models are trained against a single expert's hypnogram—even though human scorers often disagree. To address this, researchers Seyyed Ali Hoseini, Javad Baseri, Hamid Saadatfar, Edris Hoseini Gol, and AmirHossein Eshghi introduced LBH (Learning-Based Hypnogram), a framework that builds more reliable labels by modeling the stage-specific behavior of each scorer. The method uses confusion matrices from machine-learning classifiers to estimate the probability of each true sleep stage given a scorer's label, then aggregates those probabilities across all experts to assign a final label for each 30-second epoch.

Tested on the public DOD-H and DOD-O datasets with EEG (C3-M2) and chin EMG signals, the team extracted 60 features (30 per modality) and evaluated random forest, support vector machine, and multilayer perceptron classifiers. The best results came from random forest with EEG+EMG: 86.07% accuracy and 85.29% F1-score on DOD-H, and 86.04% accuracy with 84.70% F1-score on DOD-O. LBH consistently outperformed both the dataset hypnogram (DH) and best-scorer hypnogram (BSH), proving that personalized scorer modeling can improve reference label construction without discarding information from any individual expert. This could lead to more accurate sleep-stage AI training and better clinical diagnostics.