SLEEPYLAND: trust begins with fair evaluation of automatic sleep staging models

Abstract

Automatic sleep staging with deep learning has advanced considerably, yet clinical adoption remains hindered by limited generalization, model bias, and inconsistent evaluation practices. We present SLEEPYLAND, an open-source framework comprising ~ 220,000 h of in-domain and ~ 84,000 h of out-of-domain polysomnographic recordings, spanning diverse ages, disorders, and hardware configurations. We release pre-trained state-of-the-art models, evaluating them across single- and multi-channel EEG/EOG setups. We introduce SOMNUS, an ensemble that integrates models via soft-voting, achieving robust performance across 24 datasets (macro-F1, 68.7–87.2%), outperforming individual models in 94.9% of cases and exceeding prior state-of-the-art. Exploiting the Bern-Sleep-Wake-Registry (N = 6633), we show that while SOMNUS improves generalization, no model architecture consistently minimizes model demographic/clinical bias. On multi-annotated datasets, SOMNUS surpasses the best human scorer (macro-F1, 85.2% vs 80.8% on DOD-H, and 80.2% vs 75.9% on DOD-O), more closely reproducing consensus. Finally, ensemble disagreement metrics predict scorer ambiguity (ROC-AUC 82.8%), providing reliable proxies for human uncertainty.