Out of the lab, into real life: Evaluating at‐home EEG self‐monitoring

Abstract

Seizure forecasting models require long‐term, high‐quality data collected in real‐life settings using noninvasive or minimally invasive devices, yet, the lack of such systems remains a major barrier to their clinical translation. Here, we aimed to evaluate the signal quality of self‐applied at‐home EEG monitoring using a wearable system in patients with epilepsy to assess its reliability for future seizure forecasting applications. All EEG recordings were reviewed to identify non‐usable data and interictal epileptiform discharges (IEDs). We analyzed power spectral density and the temporal evolution of a signal‐to‐noise ratio, and applied composite quality criteria for each patient based on spectral profile and the proportion of EEG data discarded. Twelve patients with drug‐resistant epilepsy performed self‐applied resting‐state EEG recordings twice daily over a median period of 173.5 days (min. 12, max. 235). Two‐thirds of patients had data of good or moderate quality (N = 3 and 5, respectively), which remained overall stable over time with cap replacement every 2–3 months. IEDs were found in four patients and were concordant with prior in‐hospital recordings. Self‐applied at‐home EEG monitoring can yield clinically relevant insights and may support future seizure forecasting strategies in selected patients, provided patient adherence and the feasibility of regular maintenance follow‐up are addressed.

Plain Language Summary

Twelve people whose epilepsy was hard to control with medication recorded a short brain‐wave test (electroencephalography, or EEG) at home twice a day for several months. In 8 of the 12 people, most recordings were clear enough to use and stayed steady over time, although some EEG caps needed replacement. In four people, the home EEG showed abnormal spikes between seizures that matched earlier hospital EEGs. This suggests that long‐term, self‐recorded EEG at home is possible for some patients and may help clinicians track brain activity outside the hospital.

Key points.

• Patients self‐recorded 5‐min resting EEG at home twice daily over a median 173.5 days using an 8‐electrode dry cap.

• Adherence to the intended twice‐daily schedule averaged 63% of expected recordings.

• Two‐thirds of participants had good or moderate data quality based on composite criteria.

• Signal‐to‐noise ratio was overall stable; six patients replaced caps after a median 86.5 days.

• IEDs were captured in 4/12 patients and matched prior in‐hospital EEG in three cases.

1. INTRODUCTION

Seizure prediction models depend on detecting subtle and dynamic changes that may precede the onset of a seizure. To identify such clinically relevant patterns, these models require high‐quality data collected over extended periods. ¹ A major challenge is to obtain such data in real‐life conditions using noninvasive or minimally invasive devices. ² The development of such systems remains one of the key bottlenecks for translating seizure forecasting from research to clinical application.

In recent years, wearable devices have emerged as promising tools for seizure forecasting, enabling the long‐term monitoring of physiological signals, such as EEG, heart rate variability, sleep, and electrodermal activity. ³ Several studies showed that signal fluctuations may precede seizures, with some wearable‐based models performing above chance^{4, 5, 6} and enabling ambulatory seizure risk estimation.

While EEG recordings remain the gold standard for detecting preictal changes, EEG‐based seizure prediction models have long relied on intracranial or in‐hospital scalp recordings. ⁷ Portable and wearable EEG systems have made ambulatory, long‐term monitoring technically feasible. Noninvasive wearable systems enable patients to self‐record EEG activity with minimal setup and without conductive gel, opening new perspectives for prediction models in real‐life conditions.^{8, 9, 10}

In this study, we aimed to evaluate the signal quality of at‐home EEG monitoring performed independently by patients with epilepsy using a dry‐electrode wearable system. Our goal was to assess whether such a setup could yield reliable EEG data in real‐life, minimally supervised conditions, as a foundational step toward integrating self‐administered EEG into future seizure forecasting models.

2. METHODS

2.1. EEG@HOME protocol

2.1.1. Recruitment and training of participants

Adult patients with drug‐resistant epilepsy were recruited from King's College Hospital, London, between October 2020 and August 2022. Eligible participants provided informed consent and received either in‐person or remote training depending on COVID‐19 restrictions at the time. Participants were equipped with a portable, noninvasive EEG system connected to an amplifier and trained to independently set up the EEG system at home, including cap placement, amplifier connection, and attachment of reference/ground electrodes using dedicated adhesive patches. They also had to visually check the signal quality on the software interface. Remote technical support and additional training were provided when needed. ¹¹

Study participants were instructed to perform self‐applied, resting‐state EEG recordings twice daily—once in the morning and once in the evening—for a minimum of 5 min per session, as a compromise between patient adherence and the need to capture stable resting‐state activity. During each recording, participants were asked to sit quietly with their eyes closed to minimize artifacts.

This study was conducted according to the recommendations outlined in the Declaration of Helsinki and approved by the Bromley Research Ethics Committee (REC reference: 19/LO/0554).

2.1.2. EEG acquisition equipment

The EEG system consisted of a cap (waveguard™ touch by ANT Neuro) including eight dry electrodes positioned according to the international 10–20 system: Fpz, Fz, Cz, Pz, F3, F4, C3, and C4, with reference and ground electrodes on the left and right mastoids.

EEG signals were acquired using an eego™ amplifier (ANT Neuro) and recorded through the eego™ software interface installed on a dedicated laptop provided to each participant. Recordings were automatically saved and stored locally for later retrieval and analysis. EEG caps or laptops were replaced if technical issues occurred.

2.2. Reviewing EEG recordings

All EEG recordings were visually reviewed by three trained raters (LC, MM, PV) to identify and exclude corrupted or noisy channels from further analysis. In cases where the overall signal quality was considered unexploitable, the entire recording was discarded. During the review process, the presence of interictal epileptiform discharges (IEDs) or other pathological EEG abnormalities was also noted.

2.3. Computation of power spectra

EEG signals, sampled at 500 Hz, were preprocessed by detrending (mean removal) and applying a Hann window to reduce spectral leakage. Power spectral density (PSD) estimation was performed using the periodogram method on 20‐s segments extracted from each recording, with a frequency resolution of 0.05 Hz over the 0–250 Hz range.^{12, 13} Channels marked as invalid based on quality control were excluded before analysis, and PSD values were averaged across the retained channels.

For each segment, relative power was calculated in the delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), and beta (13–25 Hz) bands, normalized to the total spectral power within the 1–25 Hz range. To facilitate group‐level comparisons, recordings were categorized as “morning” or “evening” based on the time of acquisition, and average PSD curves were computed accordingly. All analyses were performed using MATLAB (MathWorks R2021a).

2.4. Signal quality over time

We used a simple signal‐to‐noise ratio (SNR) proxy to estimate recording quality over time. For each EEG segment, the SNR was computed per channel using a sliding‐window (1 s) approach, where the signal was smoothed using a moving average filter, and noise was defined as the deviation from this smoothed baseline. The power ratio between signal and noise was expressed in decibels (dB) and averaged across channels.

2.5. Quality composite criteria

To assess overall EEG recording quality, we applied composite criteria based on: (i) the proportion of EEG data (i.e., channels) discarded and (ii) the spectral profile of retained EEG segments. Participants were categorized into three quality groups:

• “good”: less than 20% of EEG data discarded AND a 1/f spectral profile with a visible alpha peak in the PSD plot;

• “moderate”: 20%–49% of EEG data discarded AND a 1/f spectral profile with a visible alpha peak in the PSD plot;

• “poor”: ≥50% of EEG data discarded OR PSD features indicated predominantly artifactual signals (e.g., flattened or noisy spectra lacking physiological features).

3. RESULTS

3.1. Study population

Twelve patients with drug‐resistant epilepsy participated in the study (Table S1). The median age was 37.5 years (min. 32, max. 64), and 41.7% (5/12) were female. Most participants (83.3%, 10/12) had focal epilepsy, while two had generalized epilepsy. Participants were taking a median of three anti‐seizure medications (min. 1, max. 4), and two individuals had previously undergone vagus nerve stimulation implantation.

Across participants, the median number of EEG recordings was 262.5 (min. 13, max. 439), collected over a median period of 173.5 days (min. 12, max. 235).

3.2. Signal quality

3.2.1. Data of interest from visual inspection

After excluding channels with artifacts, noise, or flat signal, a median of 67.1% (min. 17.2, max. 95.9) of channels and 90.1% (min. 49.0, max. 97.8) of total recordings per patient were available for analysis.

3.2.2. Power spectrum analysis

Eleven out of 12 patients showed a decrease in PSD at higher frequencies, corresponding to a 1/f profile, along with an alpha peak (Figure 1). One patient (#12) exhibited nonphysiological and predominantly artifactual PSD features. No difference was found between morning and evening recordings.

FIGURE 1.

Average power spectral density for each patient. For each patient, average power spectral density (PSD) curves are shown for morning (orange) and evening (blue) recordings. PSD was computed from 20‐s segments and expressed as relative power in the 1–25 Hz range. Most patients (#1–11) showed a physiological 1/f profile with a visible alpha peak (8–13 Hz), whereas patient #12 exhibited predominantly artifactual, nonphysiological spectra.

3.2.3. Data stability over time

The SNR during the initial days of recording was generally lower, with some data showing levels below 20 dB (Figure 2). Subsequently, the SNR improved and remained relatively stable over time. For one patient (#1), we noticed a pronounced drop below 20 dB around day 100, followed by a rapid recovery. This reversible decline coincided with cap damage and subsequent replacement.

FIGURE 2.

Signal‐to‐noise ratio over time. Signal‐to‐noise ratio (SNR) values from individual recordings are plotted for each patient. The red line marks the 20 dB threshold, commonly regarded as the minimum for acceptable data quality. SNR was lower during the initial training period, with values occasionally dropping below 20 dB, but improved and stabilized thereafter. In patient #1, a marked decline below 20 dB occurred around day 100 due to a damaged EEG cap, followed by rapid recovery after replacement.

Six patients changed their EEG caps during the protocol after a median of 86.5 days (minimum 30, maximum 179) due to wear or breakage.

3.2.4. Classification of data quality

Based on the composite criteria, recordings were classified as good in 3 patients (#1,5,7), moderate in five patients (#2–4,9,11), and poor in four patients (#6,8,10,12).

3.3. Clinical significance

We identified IEDs: generalized spike‐and‐wave discharges (GSWDs) in two patients (#3,5) and focal IEDs in two others (#7,9; Figure 3). These findings matched prior in‐hospital recordings for three patients, while one patient only showed IEDs on at‐home EEG monitoring. One patient (#3) also exhibited sustained GSWDs lasting over 3 seconds, interpreted as electrographic absence seizures observed in 20 home recordings (Table S1).

FIGURE 3.

EEG findings and correlation with in‐hospital recordings. Each panel (A–D) illustrates epileptiform discharges displayed in three configurations: The left trace corresponds to the at‐home recording obtained with the portable device; the middle trace shows a similar discharge recorded during in‐hospital EEG and reconstructed using a restricted bipolar longitudinal montage matching the at‐home configuration; and the right trace displays the same discharge in the full 10–20 bipolar longitudinal montage. (A) Generalized spike‐and‐wave discharges (GSWDs) lasting ~1 s (left), concordant with in‐hospital GSWDs of <3 s without contemporaneous clinical manifestation (middle and right), in a patient with juvenile myoclonic epilepsy. (B) Sustained 4‐Hz GSWDs lasting >3 s (left), concordant with in‐hospital GSWDs associated with clinically observed absence seizures (middle and right), in a patient with juvenile myoclonic epilepsy. (C) Anterior spike (left) in a patient with right frontal spikes on in‐hospital recordings (middle and right). (D) Left anterior spike–waves (left) in a patient with left fronto‐temporal sharp waves on in‐hospital recordings (middle and right).

4. DISCUSSION

In this study, we found that people with epilepsy could perform self‐applied at‐home EEG monitoring with moderate or good quality in two‐thirds of cases. Over time, the data quality remained globally stable, although EEG caps needed replacing periodically.

We chose a dry‐electrode system for its ease of application in self‐applied home settings, as it does not require conductive gel or skin preparation. Although dry electrodes can introduce higher impedance and increased sensitivity to motion artifacts, studies have shown that they can still yield data of sufficient quality for clinical application.^{14, 15} Recording quality mainly depended on patients' diligence in configuring the system and on structured remote follow‐up to detect device issues, adjust daily setup and provide retraining when needed.

After selecting the exploitable recordings, most patients exhibited a physiological PSD profile, characterized by decreasing power at higher frequencies and an alpha peak. ¹⁶ We did not observe any significant differences between the recordings made in the morning and those made in the evening. In the longitudinal analysis, the SNR remained consistent and mostly above 20 dB, which is typically considered the threshold for acceptable data. ¹⁷ EEG caps were usually changed after ~3 months, and timely replacement likely helped maintain data quality.

We recorded IEDs in four patients. Among these, two patients had generalized epilepsy and two had focal epilepsy. The locations of the IEDs matched between previous in‐hospital and at‐home recordings for three of them. Interestingly, IEDs were recorded only during at‐home sessions for one patient, as she only had a single standard EEG. The sustained GSWDs observed in a patient with juvenile myoclonic epilepsy were interpreted as electrographic absence seizures, given the absence of reported clinical manifestations. This interpretation aligns with the patient's seizure diary, which documented absence seizures occurring several times per week.

A growing body of work has shown that mobile and wearable EEG systems can provide signal quality comparable to conventional scalp EEG in both hospital and home environments.^{18, 19} Our study complements this literature by focusing on multi‐month, self‐applied recordings and detailed signal quality evaluation.

Our study has several limitations. First, the sample size was modest (n = 12), limiting generalizability. Adherence varied markedly, and recordings required periodic remote support and occasional retraining to maintain data quality, indicating that the system was not fully unsupervised. In a companion feasibility/acceptability report from the EEG@HOME cohort, adherence averaged 63% of expected twice‐daily recordings (79% for completing at least one EEG per day). ¹⁰ Quality ratings depended on manual review without inter‐rater agreement, and the composite criteria together with the SNR proxy were pragmatic and did not capture all aspects of noise. The absence of impedance monitoring further limits objective assessment. We were also unable to provide real‐time data checking, which would have helped detect any misuse or cap deterioration early. Finally, cortical coverage was constrained by the eight‐electrode dry‐cap montage without temporal electrodes, which likely reduced sensitivity to focal IEDs and could have contributed to the modest IED yield despite prolonged follow‐up.

In summary, our findings indicate that self‐applied at‐home EEG monitoring can yield clinically relevant insights and may support future seizure forecasting models in selected patients, provided patient adherence and the feasibility of regular maintenance follow‐up are addressed. Future studies should include larger and more diverse cohorts, incorporate automated and objective quality metrics, and explicitly test the forecasting performance of at‐home EEG features, ideally in combination with other wearable signals.

ETHICAL PUBLICATION STATEMENT

We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

AUTHOR CONTRIBUTIONS

LC, AB, PL, and MPR conceived and designed the study. AB and MS acquired the data. LC, AB, and JSW developed the visualization tools. LC and PL developed the statistical analysis tools. LC, AB, JSW, PFV, and MM performed the analyses. LC drafted the first version of the manuscript with contributions from AB. All authors critically revised the manuscript for important intellectual content and approved the final version for publication.

CONFLICT OF INTEREST STATEMENT

P. F. Viana has received consultancy and travel fees from UNEEG Medical A/S.

M. Schreuder is CEO of ANT Neuro GmbH and ANT Neuro UK Ltd.

L. Cousyn, A. Biondi, J. S. Winston. M. McWilliam, P. Laiou, and M. P. Richardson report no conflicts of interest.

Supporting information

Table S1. Overall characteristics of patients.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.