Lumo Sleep Open Database — publication plan and safety conditions

The Lumo Sleep Open Database is a plan to collect Apple Watch data and a reference sleep measurement during the same night so that researchers can independently re-evaluate the results.

However, as of 25 August 2026, it has 0 participants, 0 recorded nights, and has not been released. It exists as a research plan, not as a dataset that can currently be downloaded. We will not describe it as “already public” or as containing “about 1,000 nights.”

Item	Current status
Provisional target	Apple Watch + PSG under a common protocol, provisionally at least 200 people. This remains unconfirmed because the primary endpoint and power design have not yet been set
Participants and records	0 people, 0 nights
Ethics review	Reviewing body and approval both undecided
Study site	Undecided
Participant consent form	Not yet drafted
Public data	None
DOI and permanent archive	None

Why is a new dataset needed?

Existing public sleep datasets differ in site, sensors, sleep-stage scoring, and participant populations. In Paper 17’s internal analysis, adding heterogeneous cohorts did not improve performance; heterogeneity in sensors or stage distributions may have contributed. This does not causally prove that performance would improve if data were collected under one protocol.

The planned dataset would align the following measurements within the same night:

• Timestamped Apple Watch heart rate, movement, and available sleep outputs
• PSG—including EEG, eye movements, and chin muscle activity—as the main reference measurement, with sleep stages assigned using prespecified scoring rules
• Device, operating-system version, wrist side, missingness, and time-synchronisation information
• The minimum background information required for analysis, such as age group
• A development population and a separate external-validation population left untouched until final evaluation

If a research-grade simplified EEG system is used, it will be treated as a pilot or secondary analysis and not mixed with PSG as though it were the same reference measurement. Scorers will be blinded to Apple Watch results, and a subset of records will be scored independently by multiple scorers so that inter-scorer agreement can be published. EEG derivations, EOG and chin EMG, scoring rules, blinding, and agreement evaluation will be fixed at Stage 1.

What must happen before “free public release”?

Sleep time series and HRV do not necessarily become safe when names are removed. Sleep timing, rare diagnoses, device information, and person-specific cardiac patterns could be combined in ways that increase re-identification risk.

At minimum, the following must be completed before data collection begins:

1. Obtain ethics approval and define the study purpose, secondary use, retention period, and withdrawal process
2. Obtain informed consent that clearly covers the release scope; do not broaden consent retrospectively
3. Have an independent party assess re-identification risk from time, attributes, rare diagnoses, and related fields
4. Separate an open public layer from a controlled-access layer that requires review
5. Prohibit re-identification, matching, and discriminatory use in the data-use terms
6. Publish the data dictionary, exclusion flow, missingness, code, hashes, and version identifiers at the same time
7. Define how deletion or withdrawal requests from participants will be handled

Planned two-layer release

Open layer available to anyone

• Aggregate tables made sufficiently coarse after independent re-identification-risk assessment
• Sufficiently coarse derived features
• Data dictionary, analysis code, and synthetic data
• Research protocol, exclusion flow, and model cards

Controlled layer for approved researchers

• Fine-grained timestamped heart rate and movement
• PSG / EEG waveforms
• Identifiers that link multiple nights from the same person
• Attributes with higher re-identification risk

Even with “Open” in the name, unrestricted release of every raw waveform is not necessarily responsible. The release design must balance reproducibility with participant safety.

Gates before publication

• Stage 0: current — zero data and no ethics approval; only the publication plan exists
• Stage 1 — timestamp and publish the protocol, primary endpoint, power design, data dictionary, synthetic data, and analysis plan; determine the required sample size
• Stage 2 — use a small pilot to check time synchronisation, missingness, wearing procedures, and safety
• Stage 3 — collect the sample required by the Stage 1 power design under one protocol (the current provisional target is at least 200 people) and freeze the validation population before model development
• Stage 4 — pass independent privacy and reproducibility audits
• Stage 5 — make the first release with a version identifier, DOI, change log, and terms of use

How this page will be updated

If the participant count, number of nights, ethics status, repository, primary endpoint, required sample, protocol, consent language, or privacy assessment changes, the table will be updated with supporting links and a revision date. We will not advertise a “public database” before the data actually exist.

Design references

1. American Academy of Sleep Medicine. The AASM Manual for the Scoring of Sleep and Associated Events, Version 3. Summary of updates
2. Rosenberg RS, Van Hout S. The American Academy of Sleep Medicine inter-scorer reliability program: sleep stage scoring. Journal of Clinical Sleep Medicine. 2013;9(1):81–87. doi:10.5664/jcsm.2350

※ This publication plan does not establish that the Feelmo app is effective.