PSG-Audio: Polysomnography with Simultaneous Audio Recordings

A partial mirror of the PSG-Audio dataset (Korompili et al., 2021) — full-night polysomnography recordings from subjects with obstructive sleep apnea, including a bedside microphone channel suitable for sound-based sleep staging research.

Note: This copy contains approximately 59% of the original dataset by file count (~60% by size). See Completeness below for details on what is included and what is missing.

Dataset Description

Each subject has:

• EDF files (European Data Format): Multi-channel PSG recordings, typically split into ~5 one-hour parts. 20 channels including EEG, EOG, EMG, ECG, airflow, SpO2, snore sensor, and a 48kHz bedside microphone.
• RML files: XML-based expert sleep stage annotations. Each 30-second epoch is labeled as Wake, N1, N2, N3, or REM.

Key Statistics

Attribute	Value
Total size	~586 GB
File count	1,682 (of 2,851 in original)
Audio channel	Mic (channel 18), 48kHz sample rate
Annotation format	30-second epochs, 5-class (Wake/N1/N2/N3/REM)
Population	Obstructive sleep apnea patients

Completeness

The original PSG-Audio dataset (~986 GB, 2,851 files across V1/V2/V3) is hosted on scidb.cn behind time-limited FTP credentials. This mirror was created from a partial transfer before those credentials expired.

What's included

Directory	Files here	Notes
V2/APNEA_RML	175	Near-complete
V2/APNEA_RML_clean	182	Near-complete
V3/APNEA_RML	287	Complete
V3/APNEA_EDF	~750 files	Partial — 556 EDF files missing
V1/APNEA_RML	14	Partial
V1/APNEA_RML_clean	26	Partial
V3/APNEA_RML_clean	10	Partial

What's missing (1,171 files)

Directory	Missing	Notes
V3/APNEA_EDF	556	Raw multi-channel PSG recordings
V3/APNEA_RML_clean	184	Cleaned annotations
V1/APNEA_RML_clean	184	V1 cleaned annotations
V1/APNEA_RML	178	V1 raw annotations
V2/APNEA_RML	37	V2 raw annotations
V2/APNEA_RML_clean	30	V2 cleaned annotations

If you have access to the full dataset and would like to contribute the missing files, please open a discussion.

Usage

Reading EDF files (Python)

import pyedflib

edf = pyedflib.EdfReader("00001001_1.edf")
channel_labels = edf.getSignalLabels()
mic_idx = next(i for i, l in enumerate(channel_labels) if "mic" in l.lower())
audio = edf.readSignal(mic_idx)
sample_rate = edf.getSampleFrequency(mic_idx)  # 48000
edf.close()

Reading annotations (Python)

from lxml import etree

tree = etree.parse("00001001.rml")
events = tree.xpath("//ScoredEvent[Type='Stages|Stages']")
for event in events:
    start = float(event.find("Start").text)
    duration = float(event.find("Duration").text)
    stage = event.find("Type").text  # e.g., "NREM2", "Wake", "REM"

Audio channel warning

The EDF files contain a Snore channel (channel 10, 500Hz) that is not raw audio — it's a derived sensor signal. The correct audio channel is Mic (channel 18, 48kHz). Any channel below 8kHz sample rate is not a microphone.

Source

Originally hosted at scidb.cn with time-limited FTP access. This is a partial permanent mirror for reproducibility.

Paper: PSG-Audio, a scored polysomnography dataset with simultaneous audio recordings for sleep apnea studies (Scientific Data, 2021)

Citation

@article{korompili2021psgaudio,
  title={PSG-Audio, a scored polysomnography dataset with simultaneous audio
         recordings for sleep apnea studies},
  author={Korompili, Georgia and others},
  journal={Scientific Data},
  volume={8},
  year={2021},
  doi={10.11922/sciencedb.00345}
}

License

CC BY 4.0