AnySleep can score many sleep setups. Validate the clinic you actually have

THE SIGNAL IN ONE SENTENCE

A sleep laboratory is not a tidy row of identical beds. One clinic records a full set of brain and eye signals. Another has fewer channels, older hardware or a montage shaped by local practice. A patient turns over and loosens an electrode. A night technician marks an artifact. A scorer sees a transition that does not fit neatly inside a 30-second box. AnySleep is built for that mess. Researchers at FH Aachen University of Applied Sciences in Germany and Utrecht University in the Netherlands published the version of record on October 30 in npj Digital Medicine. The model is designed to score sleep from flexible combinations of electroencephalography, or EEG, and electrooculography, or EOG, channels. Instead of requiring one fixed electrode arrangement, it can dynamically combine the signals available for a recording. The paper reports training and validation on 20,589 overnight recordings from 28 datasets across multiple centers, representing more than 200,000 hours of EEG and EOG. That is a large and deliberately varied base. The authors report that AnySleep equals or exceeds established baselines at the conventional 30-second epoch, improves as more channels are available and remains useful when EOG is missing, when only EOG is available or when a single frontal, central or occipital EEG derivation is supplied. The system can also produce predictions at shorter time scales. The paper reports that those high-resolution outputs capture short wake intrusions consistent with arousals and provide additional information for distinguishing healthy controls from people with obstructive sleep apnea, narcolepsy type 1 or insomnia. The code, configurations and model resources are public under an MIT license. The repository includes inference examples for EDF files, experiment configurations and links to a Zenodo archive for results that are too large for ordinary Git storage. That makes AnySleep unusually inspectable for a medical machine-learning release. It does not make the model a universal sleep clinician. The paper describes a sleep-staging and research system. It does not authorize autonomous diagnosis, treatment, billing, triage or patient communication. The right next artifact is a clinic transfer validation sheet. Start with the recording system. The phrase any EEG or EOG channels is easy to read as any device, any setup, any night. That is not what channel flexibility proves. An electrode derivation is only one part of the pipeline. The signal also depends on amplifier behavior, sampling rate, filters, reference choice, sensor placement, impedance, room noise, patient movement, technician practice and preprocessing. The public inference example resamples input to 128 hertz, clips noise and applies robust scaling before producing five sleep-stage logits. A local validation sheet should record whether the clinic's pipeline matches those assumptions. Name the acquisition system, software version, sampling rate, filter settings, channel labels, physical montage, reference electrodes, preprocessing steps and artifact rules. Then record every transformation between the raw signal and the model input. If a channel is renamed, re-referenced or dropped, say so. If a device applies a filter before export, say so. If a technician manually repairs or excludes a segment, keep that intervention beside the prediction. The model should never receive a polished tensor while the validation report pretends it saw the clinic's raw night. Next comes channel coverage. AnySleep is valuable because it can work with different combinations. The clinic still needs to test the combinations it actually uses. Create a matrix with rows for the real configurations in local records. Full EEG and EOG. EEG without EOG. Single frontal EEG. Single central EEG. Single occipital EEG. EOG only if that configuration is genuinely considered. Then add the ugly versions. One noisy channel. One flat channel. A channel that disappears halfway through the night. An incorrect label. An unusual reference. The transfer test should measure not only accuracy but whether the system recognizes that the input is degraded. A confident answer from a broken electrode is not resilience. It is a quieter failure. Patient mix belongs on the same sheet. The model was trained across many datasets and centers, which is a strong design choice for generalization. A large multi-center sample still cannot stand in for every clinic's population. Age distribution matters. So do sleep disorders, medications, neurological conditions, psychiatric conditions, respiratory support, shift work, pregnancy, body position and the reason a person was referred. The clinic should compare its population with the study cohorts and list groups that are rare, absent or poorly described. Do not hide that gap inside one overall score. Report performance by clinically relevant group when the sample is large enough to support responsible analysis. When it is not, mark the group as insufficiently tested and require human review. The same goes for the night itself. Some recordings are clean research nights. Others contain respiratory events, limb movement, awakenings, medication effects, sensor failure and long periods that two experienced scorers would debate. The validation set should include ordinary local complexity, not only the files that survived a quality filter. Manual scoring is not a perfect gold standard. Sleep staging is labor-intensive partly because trained humans can disagree, especially around transitions and ambiguous segments. A clinic transfer sheet should include at least two independent scorers for a representative subset, the scoring rules they followed and how disagreements were resolved. Compare the model with each scorer and with the adjudicated label. If the model agrees with one scorer and not the other, that is different from a clear error. The disagreement itself is evidence about the task. It can also reveal where a high-resolution model is seeing a short transition that the 30-second convention forces a human to flatten. That possibility is scientifically interesting. It is also where caution matters most. The paper argues that 30-second epochs are pragmatic rather than physiological. AnySleep can produce finer-grained predictions and the authors report links between short wake intrusions, arousals and several conditions. Those outputs should not be turned into a new clinical measurement by enthusiasm alone. A high-resolution pattern needs a stable definition, repeatability across nights and devices, evidence that it measures the intended phenomenon, and proof that using it improves a decision people care about. The clinic should separate three questions. Does the output correlate with an established signal or condition? Does it add information beyond existing scoring? Does acting on it improve care? The paper supports research on the first two. It does not establish the third for every clinic or disease. That is why the transfer sheet needs an intended-use box. Research cohort labeling. Quality-control assistance. Second-reader support. Prioritizing segments for review. Producing a draft hypnogram for a trained scorer. Exploratory biomarker research. Each use has a different error cost. A mislabeled minute in a retrospective research set is inconvenient. A wrong summary that steers a diagnosis or treatment is more serious. A system that shortens scorer time can still be useful without pretending to replace the scorer. Write the use narrowly enough that a reviewer can tell when the model is outside it. Then define the human handoff. Who sees the prediction? What signal quality information appears beside it? Can the reviewer inspect the underlying traces? Are low-confidence and unusual channel combinations highlighted? Can the reviewer change a stage without fighting the interface? Is the correction stored for audit and later evaluation? Does the final report identify that automated staging was used? Human review is not a decorative signature at the bottom. It is the point where the evidence can still change the outcome. The validation sheet also needs failure thresholds. Set minimum acceptable performance for each intended use, not only one global accuracy target. Include macro-averaged measures across sleep stages, class-specific confusion, transition errors, recording-level failure, calibration and the proportion of the night requiring manual correction. N1 sleep often behaves differently from deep sleep or REM in both human and automated scoring. A reassuring average can hide a stage that the local workflow needs. Measure the correction burden. If a scorer has to inspect every segment because the model's confidence is not useful, the system may save no time. If the model is accurate on common stages but consistently misses the short events that matter to a research question, it may be the wrong tool for that study. If the model fails on one device family, the clinic should block that configuration rather than blend it into the mean. Prospective testing comes next. Retrospective records are useful because they allow comparison with historical scoring. They do not reproduce the full workflow change created by a live system. Run the model silently on consecutive new nights before letting it influence a report. Record upload failures, channel-mapping errors, processing time, technician interventions, missing metadata, reviewer disagreements and cases sent to manual fallback. Keep the human process unchanged during the silent phase so the clinic can measure what the system would have changed. Then introduce it gradually for the narrow intended use. The model version belongs on every result. So do the code commit, checkpoint hash, preprocessing configuration, hardware or runtime, channel list and any local thresholds. The public repository is a strength because teams can inspect and reproduce the pipeline. Public code can still change. A local deployment should pin the exact artifacts that passed validation. Do not pull the latest checkpoint into a clinical workflow on a quiet Friday and call that maintenance. Treat every material update as a new candidate. Compare it on the frozen transfer set. Review changed errors. Document the decision. Keep the prior version available for rollback. Post-deployment drift needs its own row. Clinics replace amplifiers, change software, add channels, train new staff and see new referral patterns. Performance can move even when the model file does not. Monitor signal quality, channel combinations, patient mix, confidence, scorer corrections and failure rates over time. Investigate changes by device, site, technician, referral group and stage. The model should stop being trusted before a yearly audit if the inputs have already changed. Privacy is not optional because the code is open. Sleep recordings are health data. Raw traces, event annotations, demographics and linked clinical information should remain inside approved systems with access controls, retention rules and incident handling. The clinic should document whether files leave its environment, what logs are created and whether model outputs can be linked back to a person. Open weights make local processing possible. They do not automatically configure the local deployment safely. AnySleep offers researchers a practical bridge across heterogeneous sleep datasets. That matters beyond one lab. Multi-center studies often spend enormous effort aligning channel layouts and scoring conventions before asking a scientific question. A model that handles flexible inputs can make older or differently recorded cohorts more usable and can help researchers study sleep at shorter time scales. The useful claim is not that every setup now works. It is that one model was designed and tested across far more setups than a fixed-montage system. The clinic transfer sheet completes that idea. It names the local device. The actual channels. The patient mix. The scorer disagreement. The intended use. The failure threshold. The human handoff. The model version. The drift check. AnySleep can bring many recordings into one pipeline. Validation is what keeps one pipeline from becoming one assumption.