4rooms/sleeptransformer-phan-embeddings

SleepTransformer-Phan Embeddings

Pre-extracted contextualized per-epoch embeddings from SleepTransformer (Phan et al. 2022), trained on SHHS via PhysioEx.

Each subject directory contains:

• embeddings.npy — (n_epochs, 128) contextualized epoch embeddings (bfloat16)
• labels.npy — (n_epochs,) AASM sleep stage labels

Usage

from physioex.models import load_embeddings
path = load_embeddings("sleeptransformer-phan", "hmc", verbose=True)

Linear Probe Results (5-fold subject-wise CV)

Dataset	Subjects	ACC	MF1	κ	F1-W	F1-N1	F1-N2	F1-N3	F1-REM
shhs_visit1	5793	0.8861	0.8197	0.8380	0.92	0.51	0.90	0.85	0.91
shhs_visit2	2651	0.8824	0.8068	0.8336	0.92	0.47	0.90	0.85	0.90
dcsm	255	0.8767	0.7855	0.8202	0.95	0.45	0.84	0.83	0.85
mass_ss05	26	0.8750	0.8138	0.8183	0.85	0.55	0.90	0.87	0.89
mass_ss02	19	0.8714	0.8065	0.8111	0.83	0.55	0.92	0.87	0.87
mass_ss04	40	0.8638	0.8124	0.8048	0.85	0.58	0.91	0.85	0.88
mass_ss03	62	0.8596	0.8063	0.7918	0.86	0.56	0.90	0.82	0.89
sleepedf	153	0.8394	0.7828	0.7770	0.93	0.50	0.86	0.81	0.82
mass_ss01	53	0.8234	0.7741	0.7503	0.91	0.53	0.87	0.70	0.87
stages_GSDV	232	0.8167	0.6701	0.7015	0.84	0.24	0.87	0.58	0.82
hpap	247	0.8150	0.7755	0.7476	0.88	0.48	0.84	0.80	0.87
stages_GSBB	30	0.8124	0.7168	0.7149	0.90	0.37	0.85	0.66	0.81
stages_MSMI	63	0.7974	0.7276	0.6967	0.83	0.37	0.85	0.75	0.84
stages_GSSW	105	0.7908	0.6312	0.6617	0.79	0.18	0.86	0.50	0.83
stages_GSLH	45	0.7902	0.6714	0.6688	0.85	0.36	0.85	0.54	0.77
stages_MSQW	153	0.7839	0.7064	0.6817	0.82	0.47	0.85	0.56	0.83
hmc	151	0.7836	0.7532	0.7141	0.84	0.46	0.80	0.83	0.83
mesa	2056	0.7772	0.6932	0.6810	0.84	0.40	0.82	0.61	0.79
stages_GSSA	26	0.7767	0.5691	0.6179	0.75	0.10	0.84	0.38	0.77
stages_STLK	158	0.7737	0.6601	0.6563	0.80	0.33	0.83	0.52	0.82
stages_STNF	460	0.6409	0.5803	0.5182	0.71	0.18	0.65	0.74	0.61
stages_MSTR	285	0.5862	0.4573	0.3601	0.51	0.05	0.69	0.52	0.51
stages_MSNF	38	0.5448	0.4035	0.2983	0.55	0.10	0.66	0.38	0.33
stages_BOGN	85	0.5311	0.4027	0.2819	0.49	0.14	0.64	0.34	0.39
stages_MSTH	31	0.4917	0.2057	0.0492	0.22	0.00	0.66	0.09	0.06

Model Details

• Architecture: SleepTransformer (Phan et al. 2022) — epoch transformer + sequence transformer
• Training data: SHHS visit 1 (5793 subjects, single EEG channel)
• Pipeline: seqsleepnet (STFT spectrogram, T=29, F=129)
• Sequence length: L=21 epochs
• Embedding dim: 128

Datasets

Dataset	Source	URL
Sleep-EDF	PhysioNet	https://physionet.org/content/sleep-edfx/1.0.0/
HMC	PhysioNet	https://physionet.org/content/hmc-sleep-staging/1.1/
DCSM	ERDA/KU	https://erda.ku.dk/public/archives/db553715ecbe1f3ac66c1dc569826eef/published-archive.html
SHHS	NSRR	https://sleepdata.org/datasets/shhs
MESA	NSRR	https://sleepdata.org/datasets/mesa
HomePAP	NSRR	https://sleepdata.org/datasets/homepap
STAGES	NSRR	https://sleepdata.org/datasets/stages
MASS	CEAMS	http://ceams-carsm.ca/mass/

Citations

@article{gagliardi2025physioex,
    author={Gagliardi, Guido and Alfeo, Luca and Cimino, Mario G C A and Valenza, Gaetano and De Vos, Maarten},
    title={PhysioEx, a new Python library for explainable sleep staging through deep learning},
    journal={Physiological Measurement},
    url={http://iopscience.iop.org/article/10.1088/1361-6579/adaf73},
    year={2025},
}

@article{phan2022sleeptransformer,
    title={SleepTransformer: Automatic Sleep Staging with Interpretability and Uncertainty Quantification},
    author={Phan, Huy and Mikkelsen, Kaare and Ch\'en, Oliver Y and Koch, Philipp and Mertins, Alfred and De Vos, Maarten},
    journal={IEEE Transactions on Biomedical Engineering},
    year={2022},
}