Reading a 24-hour or multi-day EEG, or scoring a full-night sleep study in 30-second epochs, is slow, specialised work with notable inter-rater variability. Deep-learning models now pre-read recordings: they flag epileptiform discharges and seizures, stage sleep and score respiratory events, so that clinicians confirm and sign instead of marking from scratch. The tools are regulated medical-device software (MDR in the EU, FDA 510(k) in the US) and are used as assistive or triage aids.
Problem & value
- Today: EEG technologists and neurologists scroll through hours to days of multichannel signal; sleep technologists score polysomnography (PSG) manually per AASM rules, typically 1-2 hours per study. Neurophysiology and sleep medicine are capacity-constrained, so waiting lists are long and ICU continuous EEG (cEEG) is often not read in real time.
- AI value: shorter reading time, faster triage of urgent findings (non-convulsive seizures in ICU), more consistent scoring across sites, and scaling of home sleep testing. Studies of autoSCORE report expert-level sensitivity for triaging routine EEG as normal vs. abnormal, and automated sleep staging reaches agreement with human scorers close to inter-scorer agreement [1][2][3].
- Not a replacement: all regulated products position the output as a suggestion that a qualified physician reviews; false negatives on rare patterns remain the key concern.
Variants
- Routine and ambulatory EEG: spike/sharp-wave detection, normal-vs-abnormal triage, standardised SCORE-style report generation.
- ICU / long-term video-EEG monitoring: seizure detection, trend displays (quantitative EEG), alerts.
- Neonatal EEG: seizure detection in NICU (specialised models, separate validation).
- Polysomnography in sleep labs: sleep staging, arousals, limb movements, apnea/hypopnea scoring.
- Home sleep apnea testing: simplified channel sets; consumer wearables (PERS-002) are not PSG-grade.
How it works
- acquisition (EEG amplifier / PSG recorder, EDF export or vendor API)
- pre-processing (filtering, montage, resampling, artefact rejection)
- model inference (CNN / transformer on windows or 30 s epochs)
- event / epoch labels with confidence (spikes, seizures; sleep stage, apnea, arousal)
- review UI: technologist or physician accepts, edits, rejects markers
- draft report (structured findings, optional LLM text drafting)
- physician sign-off
- EHR / neurophysiology or sleep information system
- Models: convolutional and recurrent networks, increasingly transformers and self-supervised EEG foundation models, on raw multichannel time series. Sleep staging models take C3/C4/O EEG plus EOG/EMG; respiratory scoring uses airflow, effort and SpO2 channels.
- Human-in-the-loop: mandatory physician review and signature; technologist edits are logged and should feed quality monitoring. Urgent ICU alerts need defined escalation paths.
- LLM layer (optional): drafts narrative reports from validated structured findings; it must not generate findings itself (see HLTH-011).
Data preparation
The data are EDF/EDF+ or vendor-format EEG and PSG recordings, technologist annotations, prior reports and patient metadata. They live in the EEG/PSG acquisition systems, on departmental file shares or PACS-like archives, in the hospital information system, and sometimes on paper reports. Typical gaps: proprietary formats, inconsistent channel names and montages, differing sampling rates, missing or free-text annotations, scorer disagreement, recordings without outcome labels, and unclear access rights between departments.
Preparation steps: inventory recordings and formats (PREP-001), convert to EDF with a harmonised channel map using tools such as MNE-Python or vendor exporters (PREP-007), digitise paper reports (PREP-006), define annotation categories per AASM rules (PREP-019), build a local validation set scored by two experts (PREP-018), and pseudonymise exports (PREP-022). For a bought product, most work is integration and local validation, about 20 to 30 percent of effort; for training your own model, data work reaches 60 percent or more. Use GUIDE-003 for the readiness check.
Implementation paths
Which path for whom
- Enterprise (hospitals, private sleep labs): Path A, because CE-marked EEG and PSG scoring products integrate with existing EEG systems and carry the required MDR conformity.
- Developer: Path C, because open-source EEG toolkits and open models (e.g. MNE-Python based pipelines, sleep-staging models) run on one GPU for research and prototypes.
- Public (university and municipal hospitals): Path A with certified products hosted on-premises or in EU clouds, because procurement rules and medical-device law favour certified solutions over self-built ones.
Path A - Buy (recommended)
- EEG: autoSCORE (Holberg EEG, Norway/Europe, CE-marked, integrated into several EEG systems), encevis (AIT Austrian Institute of Technology, Austria; CE-marked seizure/spike detection), Persyst (US; widely used in ICU and EMUs), Natus and Nihon Kohden platforms with embedded detection, Ceribell (US; rapid-EEG with seizure detection at the point of care).
- Sleep: Philips Somnolyzer (Philips / Somnolyzer 24x7 lineage, scoring of PSG), EnsoSleep (EnsoData, US), Cerebra, plus Compumedics and Nox Medical scoring modules. Many are FDA-cleared; for the EU check the CE certificate and intended use.
- Choose A when: you need certified, integrated software. Insist on MDR certificate, intended use statement, local validation, EHR/EEG-system integration and EU hosting or on-premises operation.
Path B - Build on cloud APIs
- Not suitable for the core detection: no general-purpose frontier API is validated for EEG/PSG, and streaming patient waveforms to a cloud model is hard to justify under GDPR and MDR.
- Reasonable cloud uses: de-identified research pipelines on EU regions (Azure, AWS, Google Cloud), LLM drafting of report text from structured findings via an EU-region API, quality dashboards.
- Any production clinical use of a self-assembled model makes you the manufacturer (see Risks).
Path C - Self-host (open source)
- Sleep staging: YASA (Python, open source, validated against large PSG datasets), U-Sleep (open model published in npj Digital Medicine), SleepTransformer, plus SleepECG-type tools.
- EEG: MNE-Python for processing, Braindecode (PyTorch) with models such as EEGNet, and open foundation models (LaBraM, BIOT, CBraMod) for research. Seizure/spike detectors from academic groups (e.g. SeizureTransformer, SPaRCNet-style models) exist but are research-grade.
- Data: TUH EEG corpus, CHB-MIT, NSRR (National Sleep Research Resource) and SHHS/MESA cohorts.
- Hardware: HW1-HW2; inference on CPU or a single 16-24 GB GPU is sufficient; training needs HW2/HW3.
- Status: suitable for research, retrospective audit and internal tooling, not for clinical decisions unless certified.
Path D - Custom training
- Worth it for hospitals or manufacturers with large annotated archives, or special populations (neonates, ICU, paediatric). Needs thousands of expert-annotated hours, multi-rater consensus labels, cross-site external validation and a MDR-conform quality system (ISO 13485, IEC 62304, ISO 14971). Federated learning is attractive for multi-centre data.
State of the art (2026)
- Sleep staging: deep models match inter-scorer agreement on large datasets (typical Cohen's kappa around 0.7-0.8 versus consensus), with weaker performance on N1, on patients with comorbidity and on device or population shift [3].
- EEG: autoSCORE and comparable systems perform well for normal/abnormal triage and common epileptiform patterns in routine EEG; ICU seizure detection reduces review time but produces false alarms from artefact. Rare patterns, encephalopathic rhythms and neonatal EEG remain harder [1][2].
- Foundation models: self-supervised EEG models (LaBraM, BIOT and successors) improve transfer with limited labels, but are mostly research-stage.
- Limits and failure modes: artefacts (muscle, electrode pop, ECG) mistaken for discharges; distribution shift across amplifiers and montages; biased or small training cohorts; automation bias in the reader; and weak performance on rare syndromes. Local validation is essential.
Shared resources & combinations
- Same capabilities (time-series, anomaly-detection, streaming): HLTH-013 ECG analysis, HLTH-007 remote monitoring and HLTH-012 deterioration warning share waveform pipelines, edge inference and alert routing; PSG also contains ECG and SpO2 channels.
- Neighbours: HLTH-032 (neural interfaces decode signals for control, not diagnosis), HLTH-019 (vocal biomarkers), PERS-002 (consumer sleep coaching).
- Downstream: HLTH-011 report generation, HLTH-004 decision support, HLTH-005 coding.
Effort, cost & KPIs
- Path A: weeks to a few months (integration with EEG/PSG system, local validation study); licence per study, per workstation or per site.
- Path C: 3-6 months for a research-grade pipeline; staff cost dominates.
- Path D: 12+ months and a QMS if the aim is a certified product.
- KPIs: reading time per study, sensitivity/specificity against consensus reads, false alerts per hour (ICU), staging kappa, time to report, share of AI marks edited by clinicians, backlog and waiting time, time-to-treatment for urgent findings.
People & change
Involve the chief physician and department lead of neurology or sleep medicine first, then EEG technologists and sleep physicians as daily users, IT for integration, the DPO, and data owners of the recording archives. Management decides budget and procurement; legal checks MDR status and liability.
Works council: depends. Scoring software that logs which technologist accepted or corrected which marking is objectively suited to monitor behaviour or performance, which triggers co-determination under §87(1) no. 6 BetrVG (BAG: suitability suffices). Inform the works council in good time when work processes change with AI (§90(1) no. 3); it may engage an AI expert (§80(3)). Public hospitals: Personalrat under BPersVG or state law. Agree that corrections are not used for individual performance evaluation.
DPO and DPIA: health data, so a DPIA is likely (Art. 35 GDPR) plus AVV with vendors (Art. 28). AI literacy: support staff with training on automation bias and product limits (Art. 4 AI Act; GUIDE-006). Pilot with one or two experienced readers in shadow mode, compare against their scoring, and keep the physician as final signer. Typical fears are deskilling and job loss; address with transparent communication and the role as reviewer. See GUIDE-004, GUIDE-008, GUIDE-005.
Risks, regulation & compliance
- MDR: diagnostic and triage software is generally medical-device software, usually Class IIa or higher under Rule 11 of MDR (EU) 2017/745; check each vendor's CE certificate, intended use and notified body. Software built in-house for clinical use without certification is only allowed under the Art. 5(5) health-institution exemption, with strict conditions.
- EU AI Act: high-risk (Art. 6(1) with Annex I): AI that is a medical device requiring third-party conformity assessment. High-risk obligations apply to such products from 2 August 2028 (Annex I products; deferred from August 2027 by the Digital Omnibus on AI, Reg. (EU) 2026/1744). Deployers must ensure human oversight, log retention, and AI literacy (Art. 4). No Art. 5 prohibition applies; EEG here is diagnostic, not emotion inference or biometric categorisation. Neurodata used for emotion recognition at work or school would be prohibited, which is out of scope.
- GDPR: EEG and sleep data are health data (Art. 9), potentially biometric-like and highly identifying; need legal basis (treatment, Art. 9(2)(h)), DPIA, processing agreements, EU/on-prem hosting, and pseudonymisation for research. Cross-border transfer needs safeguards. The European Health Data Space (EHDS) will shape secondary use.
- Clinical risk mitigations: local validation on own data, performance monitoring after deployment, post-market surveillance, mandatory physician review, bias checks across age and sex, and fallbacks when the system is down.
Getting started
- Define scope (routine EEG triage, ICU cEEG, PSG scoring) and measure baseline reading time and backlog.
- Shortlist 2-3 CE-marked products, request the MDR certificate and published validation studies.
- Run a retrospective local validation on a few hundred anonymised studies with consensus expert labels.
- Integrate in shadow mode, then assistive mode with logged edits; define alert escalation for ICU.
- Set up monitoring (drift, false alarms) and complete DPIA and AI Act deployer duties.
Sources
- Holberg EEG - autoSCORE - https://www.holbergeeg.com/ (accessed 2026-10; vendor page, not independently verified)
- Tveit J. et al., "Automated interpretation of clinical EEG using deep learning (autoSCORE)", JAMA Neurology 2023 - https://jamanetwork.com/journals/jamaneurology/fullarticle/2805187
- Perslev M. et al., "U-Sleep: resilient high-frequency sleep staging", npj Digital Medicine 2021 - https://www.nature.com/articles/s41746-021-00440-5
- Vallat R., Walker M., "An open-source, high-performance tool for automated sleep staging (YASA)", eLife 2021 - https://elifesciences.org/articles/70092
- MNE-Python - https://mne.tools/
- AASM Scoring Manual - https://aasm.org/clinical-resources/scoring-manual/
- Regulation (EU) 2017/745 (MDR) - https://eur-lex.europa.eu/eli/reg/2017/745/oj
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- National Sleep Research Resource - https://sleepdata.org/