World Sleep Industry ENCYCLOPEDIA

sleep EEG data analysis services

6.5.3

Definition, scope, and interpretation boundary

Sleep EEG data analysis services receive, quality-check, annotate, stage, detect events in, quantify, visualize, and report electroencephalographic and related physiological signals acquired during sleep. Inputs may come from laboratory polysomnography, research EEG, home systems, or wearables and are often combined with electrooculography, electromyography, electrocardiography, respiration, oxygen saturation, position, video, and event markers. Analysis is limited by the acquired signals and prespecified question: it cannot restore an unrecorded lead or create reliable conclusions from poor-quality or unsuitable data.

The service should distinguish technical processing, sleep-technologist scoring, research analysis, algorithm development, and clinical interpretation by an appropriately authorized professional. Annotations of sleep stages, arousals, respiratory events, or movements may support a report, but an analyst should not diagnose, change treatment, or promise regulatory acceptance outside the authorized role. Research, clinical, and product-validation outputs may require different rules, review, records, and validation. Performance on one dataset does not establish that the same algorithm works across other devices, populations, or purposes.

Data intake, provenance, and analysis plan

Intake should confirm the study or test identifier, consent and use authority, acquisition date, device and software versions, lead names and positions, reference, units, sampling rates, filters, calibration, impedance records, time zone, lights-off and lights-on markers, medications, interventions, and technologist notes. It should establish whether source files are continuous, resampled, or exported, whether clocks drift, how devices were synchronized, and whether earlier processing is reversible. Patient or participant identity should be separated from the analysis identifier, with transfer, storage, and access controlled under applicable privacy, ethics, and data agreements.

The preanalysis plan should specify the target population, recording type, scoring standard and version, primary and secondary measures, valid-record thresholds, analysis interval, lead priorities, artifact rules, missing-data approach, subgroups, repeated measures, and statistics. Routine clinical analysis, exploratory spectra, algorithm training, and validation should be labeled separately so that frequency bands, windows, or exclusions are not changed after reviewing results. When the input is unfit for the purpose, the appropriate output is a data-gap and feasibility assessment, not an assumption that interpolation, rereferencing, or automated repair can hide the deficiency.

Signal quality control and preprocessing

Quality control should first check readability, duration, channel mapping, units, dynamic range, sampling, and timestamps, then identify lead loss, saturation, clipping, bridging, drift, mains interference, electrode movement, ocular, cardiac, muscular, and motion contamination. Usability, loss, abnormality, and repair should be logged for every channel and period, with an immutable source copy retained. Quality metrics can support triage but cannot replace inspection of the recording. Treating a bad channel, artifact, or device restart as physiology can distort staging, events, and spectra at the same time.

Rereferencing, filtering, notch filtering, resampling, interpolation, independent-component procedures, and other artifact processing change the signal. Parameters should follow the question and data properties, with before-and-after comparisons preserved. Filters should not create or remove defining waveforms, resampling requires anti-aliasing, and interpolated data must not be represented as measured. Quantitative EEG analysis should address artifact before statistics and report excluded proportions by stage and group. If exclusion is related to disease, hardware, or intervention, resulting selection bias requires assessment.

Sleep staging, events, and agreement

Manual or assisted scoring should use the project-specified rule version appropriate to age, recording type, and use, defining epoch length, lead set, sleep stages, arousals, respiratory events, movements, and cardiac events. Total sleep time, efficiency, latency, stage proportions, and event indices in polysomnography depend on denominators and recording intervals. If a home test does not measure credible sleep time, an index based on recording time should not automatically be named as though it were the identically labeled polysomnographic index. Rule revisions and alternative desaturation thresholds can change results and must be reported.

Scorers should have relevant training, difficult segments should follow a prespecified review path, and agreement for stages, arousals, respiratory events, and movement events should be assessed periodically. Overall accuracy alone can conceal systematic errors in uncommon stages or events, so class-level confusion, count differences, bias, and agreement measures should also be considered. Reference scoring itself may differ among experts. Consensus creation should retain each scorer’s labels, adjudication, and final label rather than overwrite the original judgment history.

Quantitative EEG and statistical analysis

Quantitative work may cover spectra, slow-wave activity, spindles, K complexes, microarousals, connectivity, phase–amplitude coupling, or time-frequency features. Every measure should state leads, reference, frequency bands, window, overlap, normalization, baseline, and detection threshold. Architecture and signal features should be organized around a physiological question, stage, and time rather than selected from many features solely because they are significant. Scalp EEG records potential differences influenced by volume conduction, reference, and noise; a sensor-level association is not by itself a localized brain source or causal mechanism.

The statistical unit should match the design and distinguish participant, night, epoch, event, and channel, avoiding treatment of thousands of epochs from one person as independent participants. Models may need to address repeated measures, overnight sequence, site and device batches, age, sex, medication, and prespecified confounders and should report effect sizes, intervals, missingness, multiplicity, and sensitivity analyses. Small or exploratory studies require explicit uncertainty. A group-average difference cannot be used directly as an individual diagnosis, and association, classification performance, and prediction are not interchangeable.

Automated scoring and algorithm validation

An automated staging or event detector should define intended use, input channels, preprocessing, model and weight versions, training and validation sources, labeling standard, participant-level splitting, and the role of human review. Training, tuning, and independent test sets should be isolated to prevent the same person, adjacent epochs, or the same night from leaking across partitions and inflating performance. Validation should cover target ages, conditions, devices, sites, and signal quality and report class-level performance, calibration, confidence intervals, failure and indeterminate rates, and bias in underrepresented groups and across hardware.

Algorithm outputs should retain confidence or quality information, with poor-quality, unfamiliar, or out-of-distribution records eligible for human review instead of receiving forced labels. The scope of review, correction rate, and final responsibility should be explicit; speed does not exempt automated output from source-signal review. Changes to model, scoring rules, firmware, or preprocessing require impact assessment, version control, and necessary revalidation. Software used for medical decisions also faces applicable product, quality, risk, cybersecurity, and market requirements; research code is not an authorized medical device merely because it runs successfully.

Delivery, reproducibility, and quality oversight

Deliverables may include a source inventory, quality report, channel and event dictionary, processing log, stage and event annotations, architecture tables, quantitative features, figures, analysis code, software environment, and results report. Source data, cleaned data, labels, derived measures, and final tables should be layered and connected by participant, session, channel, time, and version keys. Open formats and structured metadata support exchange, but conversion loss must be recorded. The report should identify standards, limitations, anomalies, exclusions, deviations, human review, and the valid scope of use.

The quality system should manage staff qualifications, inter-scorer agreement, code review, test data, device and software versions, access logs, backups, defects, corrective action, and change control. Before delivery, tables and figures should be reproduced from frozen inputs and controlled code and the sample flow reconciled. At service closeout, data should be transferred or destroyed under agreement, access revoked, and an audit trail retained. Quality is demonstrated by clear provenance, traceable raw signals, explicit rules, reproducible processing, use-specific validation, and restrained interpretation—not by automation level, feature count, or a single accuracy score.