4rooms/tinysleepnet-supratak-embeddings

TinySleepNet-Supratak Embeddings

Pre-extracted contextualized per-epoch embeddings from TinySleepNet (Supratak & Guo 2020), trained on Sleep-EDF via PhysioEx.

Each subject directory contains:

• embeddings.npy — (n_epochs, 128) contextualized epoch embeddings (bfloat16)
• labels.npy — (n_epochs,) AASM sleep stage labels (W=0, N1=1, N2=2, N3=3, REM=4, unscored=-1)

Usage

from physioex.models import load_embeddings

path = load_embeddings("tinysleepnet-supratak", "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
sleepedf	153	0.8754	0.8330	0.8278	0.95	0.60	0.89	0.84	0.89
mass_ss05	26	0.8620	0.7873	0.7986	0.82	0.48	0.90	0.87	0.87
mass_ss02	19	0.8545	0.7745	0.7862	0.82	0.44	0.90	0.87	0.83
mass_ss04	40	0.8432	0.7768	0.7746	0.82	0.48	0.90	0.84	0.84
mass_ss03	62	0.8397	0.7751	0.7616	0.84	0.49	0.89	0.82	0.84
dcsm	255	0.8332	0.7212	0.7552	0.91	0.32	0.81	0.82	0.74
shhs_visit2	2651	0.8159	0.7095	0.7375	0.86	0.27	0.85	0.83	0.75
shhs_visit1	5793	0.8044	0.7083	0.7198	0.84	0.31	0.83	0.80	0.75
mass_ss01	53	0.7966	0.7423	0.7124	0.87	0.49	0.86	0.69	0.81
stages_GSDV	232	0.7706	0.6121	0.6261	0.77	0.18	0.84	0.56	0.70
hpap	247	0.7567	0.7040	0.6663	0.81	0.38	0.81	0.79	0.73
mesa	2056	0.7556	0.6648	0.6499	0.82	0.34	0.81	0.62	0.74
stages_MSMI	63	0.7549	0.6706	0.6303	0.78	0.31	0.82	0.69	0.75
stages_GSSW	105	0.7534	0.5568	0.5964	0.75	0.11	0.84	0.38	0.71
stages_GSLH	45	0.7518	0.6254	0.6077	0.79	0.26	0.83	0.61	0.64
stages_MSQW	153	0.7468	0.6564	0.6265	0.77	0.42	0.84	0.54	0.71
stages_GSBB	30	0.7369	0.6286	0.6038	0.84	0.27	0.80	0.62	0.61
stages_GSSA	26	0.7348	0.5183	0.5279	0.65	0.02	0.82	0.46	0.64
hmc	151	0.7271	0.6837	0.6387	0.77	0.36	0.76	0.79	0.75
stages_STLK	158	0.7240	0.5987	0.5814	0.75	0.24	0.80	0.50	0.71
stages_STNF	460	0.6055	0.5369	0.4687	0.68	0.13	0.63	0.70	0.55
stages_MSNF	38	0.5200	0.3414	0.2392	0.50	0.02	0.65	0.35	0.19
stages_MSTR	285	0.5181	0.3448	0.2159	0.43	0.00	0.65	0.35	0.29
stages_BOGN	85	0.5080	0.3543	0.2355	0.44	0.04	0.63	0.29	0.36
stages_MSTH	31	0.4946	0.1712	0.0219	0.17	0.00	0.66	0.01	0.02

Model Details

• Architecture: TinySleepNet (Supratak & Guo 2020) — CNN feature extractor + LSTM sequence encoder
• Training data: Sleep-EDF (153 subjects, single EEG channel)
• Pipeline: raw (bandpass 0.3-40 Hz, resample 100 Hz)
• Sequence length: L=20 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},
}

@inproceedings{supratak2020tinysleepnet,
    title={TinySleepNet: An Efficient Deep Learning Model for Sleep Stage Scoring based on Raw Single-Channel EEG},
    author={Supratak, Akara and Guo, Yike},
    booktitle={IEEE EMBC},
    year={2020},
}