KurisutinaEpilepsia 66(11):4476–4489, doi 10.1111/epi.18566; open access (CC BY 4.0), PMC12661286. Preprint on medRxiv, November 2024. Read for research direction R1, 4 October 2026. Provenance: papers/amadeus/viana2025_sqeeg_15month.provenance.json.
What was read
- Read in full: the Europe PMC full-text XML converted to text: abstract, key points, introduction, Methods, Results, Tables 1–3, all figure captions, Discussion, statements, conflicts of interest and references.
- Not read: the Supporting Information (Data S1: inclusion criteria; procedure schematic; device photos; adverse-event and device-deficiency lists; individual adherence plots; participants' comments on acceptability; detector performance). The main text summarises these. Figures were not inspected as images.
Design
- Study: SUBER, a prospective observational study at King's College London and King's College Hospital. Ethics 19/LO/0354; NCT04061707. Funded by the Epilepsy Foundation's My Seizure Gauge project.
- Participants:
- Inclusion: adults with treatment-resistant focal epilepsy and more than 20 seizures a year, with seizures detectable by two-electrode scalp EEG.
- Flow: 15 consented and 12 were implanted. Two left before usable data: one implant was misplaced and protruded through the skin, and one person had immediate headache and scalp pain. One dropped out after 45 days through device malfunction, and one was re-implanted.
- Final cohort: 11 datasets from 10 people, aged 29–64, half men.
- Device: the UNEEG 24/7 EEG SubQ.
- Implant: a lead placed under the scalp (subgaleal) through a 25 mm incision behind the ear under local anaesthesia, aimed at the expected seizure focus.
- Signal: two channels at 207 Hz, band-passed 0.5–48 Hz.
- Logger: an external data logger worn by the patient powers the implant wirelessly and stores the data. It is disconnected only when it could get wet.
- Explant: half a day in hospital under local anaesthesia.
- Other streams:
- Seizure diary: an electronic app with three seizure categories.
- Fitness tracker: a consumer Fitbit for heart rate, steps, sleep duration and staging.
- Questionnaires: acceptability (7 items) and the System Usability Scale, at 3 months and at the end.
- Visits: up to monthly, to offload data.
- Seizure annotation: a high-sensitivity automated detector, plus review of ±2 h around every diary entry, plus a random 10% sample; full review where the detector's sensitivity was below 80%. Annotated by a board-certified electroencephalographer and a technologist; inter-rater agreement was not assessed.
Main results (verified)
- Volume and adherence:
- Duration: the median recording span was 433 days by Results and Table 2 (the abstract and key points say 438). Range 45–538 days.
- Hours per day: a median of 18.8 h/day (78.4% of the time). Five people recorded more than 20 h a day and three fewer than 12; one recorded only 4.1 h a day after becoming seizure-free with a vagus nerve stimulator. 71,984 hours in total.
- No attrition: a group model found no significant decline in adherence over time (intercept equivalent to 81%). Adherence patterns were individual: some people had fixed hours off, some recorded mainly by day or by night, and four showed weekday effects.
- Acceptability: high, and it stayed high at the end. Participants felt the system did not limit their daily lives; about half felt it made their illness "more exposed".
- Safety and reliability:
- Adverse events: 12, of which 7 were possibly or probably related, mostly mild and transient pain or headache. Two were serious: unrelated pneumonia, and the misplaced implant needing urgent removal. All recovered without sequelae.
- Device deficiencies: 32, of which 25 affected data collection. Most involved the external logger and were fixed by rebooting or replacement.
- The brain record against the person's own report:
- Under-reporting: 754 seizures on sqEEG, and 52% were not in the diary, including 3 of 10 convulsive seizures.
- Over-reporting: 140 diary events (27–28%) had no sqEEG seizure; 68% of those were reported as seizures with preserved awareness, which even full scalp EEG detects poorly.
- Agreement: F1 of 0.58 overall, within-person median 0.56, range 0.06 to 0.98 (the abstract says 0.97).
- Distorted rhythms: in one person, diary-only data suggested a strong 24-hour cycle that the EEG did not show. The person had not been reporting nocturnal seizures.
- Personal rhythms: clustered seizures ranged from 2% to 87% across people, with apparently stable patterns within each person. Most people had significant circadian cycles, and Fano factors were more often significant on sqEEG than in diaries.
- Analysis burden: detector sensitivity was variable. Results give a median of 70.5% (IQR 48–94.6%) with 3.4 false detections per day; the Discussion says "median of 88.8%", an internal inconsistency. Several datasets needed full manual review.
- The authors on non-invasive alternatives: "scalp EEG is limited to a few weeks at most, due to the potential for skin injury, inconvenience of visible electrode wires, and signal quality degradation with time." Dry electrodes also degrade. Behind-the-ear and in-ear EEG have been studied "in small numbers of patients, but their long-term signal quality is unknown."
- Data: selected recordings "will be made available in a future open-source seizure prediction challenge". Not available now.
Limits
- Ten patients with frequent seizures and a clinical reason to tolerate an implant. Healthy people are not represented, and motivation fell when the clinical benefit disappeared (the person recording 4.1 h a day).
- Two channels over one region. It records seizures and sleep-scale rhythms, not cognition.
- Conflicts of interest: one author is a UNEEG employee; others have advisory, consultancy, equity or research-support ties to UNEEG, Seer Medical or Medtronic.
- Small internal inconsistencies between abstract, results and discussion: 433 against 438 days, F1 maximum 0.98 against 0.97, detector sensitivity 70.5% against 88.8%.
- No simultaneous video-EEG validation, and no inter-rater agreement for annotations.
What it means for Amadeus
- Near-continuous EEG for months is proven (verified), at a cost. People wore a recorder about 19 hours a day for over a year without attrition and found it acceptable. But this needed a minimally invasive implant, and gave two channels at about 200 Hz. A month of continuous coarse EEG is therefore technically trivial in duration. What is uncertain is whether a healthy volunteer can be given a subscalp implant for research. This paper says nothing about that, and it is the ethics question (inference).
- Non-invasive continuous EEG for a month is not established (verified as the authors' statement). Scalp EEG lasts "a few weeks at most" (skin injury, wires, signal decay), and ear-EEG long-term quality is unknown. A healthy-person design should plan non-invasive EEG as daily sessions or nightly wear, not 24/7 for 30 days (inference).
- Self-report and the brain record diverge on timing, measurably (verified for seizures). Half the events the brain record caught were never reported. A quarter of reports had no brain correlate, and the misreporting biased the inferred rhythms. That is a concrete precedent for P6 (triangulate what the person says with what the brain does), with the caveat that impaired-awareness seizures are unreportable by nature. For ordinary thought, the size of the divergence is unknown (inference).
- Analysis is the hidden cost (verified). About 72,000 hours needed automated detection plus expert review, and detection failed for some people. A month of one person is about 720 hours. Labelling it by inspection is impossible, so the pipeline must be automatic, with experience-sampling anchors.
- E3 (verified): no public data yet; seizures, not recall. Not an E3 candidate.