PainWaive 便携式 EEG 头戴设备用于慢性疼痛研究
Evaluation of PainWaive: a consumer-grade EEG headset for remote EEG monitoring and neurofeedback research in chronic pain
研究评估了 PainWaive,一种专为慢性疼痛患者设计的便携式 EEG 头戴设备,用于神经反馈(NFB)应用。该设备在中央头皮区域记录脑电活动,并在相同条件下同时测试了与研究级系统的测试-重测可靠性和跨设备一致性,结果显示在闭眼状态下 α 波和 PAF 的测试-重测可靠性良好,而 β 和 θ 波则表现一般。研究还发现,与研究级系统相比,该设备在闭眼条件下的跨设备一致性较好,但在睁眼条件下则较差。
Introduction
Electroencephalography (EEG) is widely used in research and clinical practice to study brain function and to diagnose conditions such as epilepsy and sleep disorders1. It is also used therapeutically in the form of neurofeedback (NFB), where individuals learn to modulate their brain activity in real time, providing a non-pharmacological option for conditions characterised by dysregulated oscillations, such as chronic pain2,3. However, traditional NFB applications have relied on research-grade EEG systems, which require lengthy setup, precise electrode placement and trained staff. This creates substantial barriers to wider clinical use4,5,6,7, particularly for people with impaired mobility, chronic illness, or those living in rural and remote regions6,7,8.
Portable, wireless, home-based EEG headsets offer a scalable alternative for NFB applications. However, most available consumer-grade systems target frontal or temporal regions for applications such as meditation, attention or sleep9,10,11,12,13,14,15,16,17, with relatively little focus on central scalp activity, which has been shown to be altered in chronic pain populations18,19,20. As such, there is currently no available consumer-grade EEG headset targeting central scalp sites that has undergone rigorous test–retest and cross-device evaluation. Despite the challenges posed by condition-dependent variability in EEG spectral measures (e.g. eyes-closed [EC] vs eyes-open [EO])15,16,17,21,22,23,24,25,26,27,28 and the greater inherent stability of some metrics compared to others (e.g. peak alpha frequency (PAF) is more stable than beta or theta power29,30,31,32,33), such evaluation is essential before widespread clinical and home-based implementation. To address this gap, we developed and evaluated PainWaive, an in-house 2-channel headset purpose-built to record central scalp activity from C1 and C2, designed for NFB applications in people with chronic pain. The name PainWaive was derived from collaboration between the research team and people with lived experience of chronic pain.
Since the PainWaive headset is intended for NFB, its evaluation must build on and improve the existing evidence base for testing consumer-grade EEG headsets. Indeed, several studies have examined test–retest reliability and cross-device consistency in consumer-grade EEG systems. However, findings remain inconclusive due to the following methodological and technical limitations. Firstly, agreement with a research-grade system and test–retest reliability of the consumer device are often assessed in separate samples or under a single recording condition13,15,16,17,23,25,26,27,28,34,35,36,37,38,39. Cannard et al.12 reported strong correlations between the consumer-grade MUSE EEG headset and a research-grade system for alpha (r = 0.87) and theta power (r = 0.73) during EC resting-state recordings. In contrast, Mikhaylov et al.40 found negligible correlations during EO recordings (alpha r = –0.06, theta r = 0.12). It is unclear the extent to which these substantially lower correlations reflect the use of EO conditions, device performance or an interaction between the two. Second, many studies lack statistical rigour, with few reporting a priori power or sample-size calculations9,16,25,27,39,41,42 and most summarising reliability only with relative indices (e.g. intraclass correlation coefficient, ICC), without absolute indices such as the standard error of measurement (SEM)9,16,25,27,39,41,42. As such, it is unclear whether between-session variability is small enough to conclude clinically meaningful changes43. Finally, device-level technological limitations constrain achievable reliability: many low-cost headsets lack impedance monitoring9,10,11,12,13,14,15,16,17 and rely on dry electrodes, which reduce signal-to-noise ratios and increase susceptibility to motion artefacts compared with gel- or saline-based sensors44,45,46,47,48,49,50,51,52.
In designing the present evaluation of PainWaive, we therefore explicitly addressed these limitations by (a) assessing both test–retest reliability and cross-device consistency, (b) including both EC and EO resting conditions in the same powered sample, and (c) incorporating impedance monitoring and saline-soaked electrodes to mitigate common technical constraints. Resting-state EEG under EC and EO conditions was recorded using both PainWaive and a research-grade system (Brain Products LiveAmp) in separate blocks across two sessions on the same day, with alpha, beta and theta power and PAF as primary outcomes. To evaluate intended real-world use, we also examined reliability in a sample of adults with chronic neuropathic pain who completed a 20-session PainWaive NFB programme at home, using repeated recordings per session to assess the consistency of these metrics over time.
Based on expected differences in stability across metrics (PAF and alpha power typically more stable than beta and theta power)53,54,55,56,57,58,59,60,61 and recording conditions (EC typically more stable than EO), we hypothesised that EC recordings would show excellent test–retest reliability (ICC ≥ 0.80) for alpha power and PAF, and good-to-excellent reliability (ICC = 0.60–0.90) for beta and theta power. Under EO conditions, we predicted good-to-excellent reliability (ICC = 0.60–0.8) for alpha power and PAF and moderate-to-good reliability (ICC = 0.40–0.7) for beta and theta power. We further predicted that cross-device consistency with the LiveAmp would be constrained by inherent reliability profiles. For EC, we expected alpha power and PAF to show good cross-device consistency (ICC = 0.60–0.8), with beta and theta power showing only moderate-to-good consistency (ICC = 0.40–0.70). For EO, we anticipated moderate-to-good cross-device consistency (ICC = 0.50–0.70) for alpha power and PAF, and fair-to-moderate consistency (ICC = 0.30–0.60) for beta and theta power.
Results
Descriptives
Tables 1 and 2 present descriptive statistics for each headset, session and condition. Most measures met normality criteria based on skewness and kurtosis. Exceptions included theta power in both LiveAmp sessions and beta and theta power in the PainWaive, which exceeded kurtosis thresholds. In the EO condition (Table 2), several LiveAmp measures, notably alpha power and PAF, also showed elevated skewness and kurtosis, with PainWaive alpha power likewise deviating from normality. As the majority of metrics satisfied normality assumptions, data were not transformed and non-parametric tests were not applied, in order to preserve interpretability and allow direct comparisons across devices and conditions.
Absolute reliability
Table 3 reports SEMeas, indexing between-session fluctuations for each metric. SEM values were broadly similar across headsets, indicating that overall, between-session variability was comparable for the LiveAmp and PainWaive.
Relative test–retest reliability
Findings of the planned analyses are shown in Fig. 1 (left) and Fig. 2 (left) for EC and EO testing respectively. During EC testing (Fig. 1 left), relative reliability was excellent for alpha power, theta power and PAF across both devices, whereas beta power showed moderate reliability for the LiveAmp and good reliability for the PainWaive. During EO testing (Fig. 2 left), the test-retest reliability of alpha power remained in the excellent range across devices, and good range for beta and PAF. Reliability of theta power was in the moderate range for the PainWaive and good range for the LiveAmp.
Left panels show the primary/planned test–retest reliability estimates for A PainWaive and B LiveAmp, with intraclass correlation coefficients (ICCs) and 95% confidence intervals for beta power, peak alpha frequency (PAF), alpha power and theta power. Right panels summarise sensitivity analyses examining alternative spectral estimation and preprocessing choices, EMG-related contamination, exclusion of recordings with flagged C1/C2 channels, individualised PAF estimation and low- and high-beta sub-bands where applicable. Grey cells indicate analyses that were not applicable to the corresponding metric or device. Cell colours denote reliability category: excellent (ICC ≥ 0.80), good (0.60–0.79), moderate (0.40–0.59), fair (0.21–0.39) and poor (≤0.20). Full details of the sensitivity analyses are provided in the Supplementary Material.
Left panels show the primary/planned test–retest reliability estimates for A PainWaive and B LiveAmp for beta power, peak alpha frequency (PAF), alpha power and theta power. Values are intraclass correlation coefficients (ICCs) with 95% confidence intervals. Right panels summarise sensitivity analyses examining alternative preprocessing and spectral-estimation approaches, an EMG-aware exclusion pipeline, exclusion of recordings with flagged C1/C2 channels and low- and high-beta sub-bands where applicable. Grey cells indicate analyses that were not applicable or not performed. Cell colours indicate coefficient magnitude: excellent (≥0.80), good (0.60–0.79), moderate (0.40–0.59), fair (0.21–0.39) and poor (≤0.20). Full details of the sensitivity analyses are provided in the Supplementary Material.
Findings of sensitivity analyses are shown in Fig. 1 (right) and Fig. 2 (right) for EC and EO testing respectively. To assess whether these findings were dependent on preprocessing choices, we repeated the test–retest analyses across multiple plausible analytical pipelines varying: (a) the use of log-transformed absolute power rather than z-transformed power, (b) DPSS rather than Hanning tapers and (c) relaxed rather than rigorous cleaning (Detailed results shown in Figs. S1–S2). Alpha power and PAF were comparatively robust to these analytic choices across both devices, whereas beta and theta showed greater variation. Within-device reliability nevertheless remained broadly moderate-to-excellent across alternative pipelines.
Because residual muscle activity may particularly affect beta-band estimates, we additionally examined low beta (13–20 Hz) and high beta (20–30 Hz) separately and implemented a more stringent pipeline in which epochs with unusually high (potentially muscle related) 30–45 Hz power were additionally rejected. Test–retest reliability was generally good-to-excellent for both high and low beta bands, and the EMG-aware pipeline produced a pattern similar to the primary analysis (Detailed results shown in Figs. S6–S7).
To assess whether retaining potentially noisy C1/C2 channels influenced LiveAmp reliability, we quantified how often either channel exceeded the variance criterion and repeated the analysis after excluding affected recordings. Only 3 of 320 valid LiveAmp recordings (0.94%), arising from 2 of 80 participants (2.5%), were flagged. Excluding these recordings had negligible impact on LiveAmp test–retest reliability (Further details in Table S1 and Fig. S8).
We also examined whether the LiveAmp reliability estimates were influenced by spatially diffuse, volume-conducted scalp activity using a supplementary current source density (CSD) analysis. The overall eyes-closed pattern was preserved after CSD transformation: alpha power, theta power and PAF showed little change relative to the primary pipeline, whereas beta reliability increased from moderate to good (ICC = 0.56–0.70, further details in Fig. S5).
Finally, because fixed-window estimation of PAF may be sensitive to the predefined 8–12 Hz range, we repeated the eyes-closed PAF analysis using the Corcoran restingIAF toolbox62. Valid individualised PAF estimates were available for 122/160 LiveAmp recordings (76.3%) and 116/160 PainWaive recordings (72.5%). Among recordings with a detectable alpha peak, reliability was good for LiveAmp (ICC = 0.76) and excellent for PainWaive (ICC = 0.94), providing convergent support for the primary fixed-window PAF findings while highlighting the reduced data yield of individualised peak estimation (further details in Fig. S9).
Relative test-retest reliability in other channels
Under EC conditions (Fig. 3, top), alpha power and PAF exhibited uniformly high ICC values across the scalp, with most channels in the good-to-excellent range. Theta power also showed broadly good reliability, with only small clusters of slightly lower ICCs over lateral regions. In contrast, beta power displayed a more heterogeneous pattern, with good reliability over posterior sites but only fair-to-moderate reliability over midline frontal–central channels.
Topographical maps show intraclass correlation coefficients (ICCs) across all LiveAmp channels for alpha power, beta power, theta power and peak alpha frequency (PAF) during eyes-closed recordings (top row) and eyes-open recordings (bottom row). The colour scale indicates ICC magnitude.
Under EO conditions (Fig. 3, bottom), alpha power remained highly reliable over central and posterior electrodes but showed reduced reliability at frontal sites. Theta and beta power showed the lowest and most spatially variable ICCs, with predominantly fair-to-moderate values over central and posterior regions and only isolated areas of good reliability. For PAF, EO reliability was still good-to-excellent over central and posterior channels but dropped sharply at frontal electrodes.
Relative test-retest reliability (clinical sample)
In the planned analysis of the clinical NFB sample (Fig. 4 left), across all 160 sessions completed by the eight participants, the consistency of alpha power, PAF, beta power and theta power across the five NFB games within each session was in the good-to-excellent range (ICC > 0.72), indicating that a single recording provided a stable estimate of the session-level EEG value under nominally identical conditions.
The left panel shows the primary/planned pooled ICC(3,1) estimates obtained using the standard 5-s non-overlapping analysis for beta power, peak alpha frequency (PAF), alpha power and theta power. The right panel summarises sensitivity analyses examining a shorter 2-s non-overlapping window, a 2-s window with 80% overlap, mean participant-level ICC(3,1), cohort-specific ICC(3,1) estimates for the neuropathic eye pain and chronic low back pain cohorts, and reliability of the protocol-specific 10–15 Hz reward band and high-beta band where applicable. Grey cells indicate analyses that were not applicable to the corresponding metric. Cell colours indicate reliability magnitude: excellent (≥0.80), good (0.60–0.79), moderate (0.40–0.59), fair (0.21–0.39) and poor (≤0.20). Full details of the sensitivity analyses are provided in the Supplementary Material.
Findings of sensitivity analyses are shown in Fig. 4 (Right). Because the clinical protocol targeted frequency bands that differed slightly from the conventional bands examined in the experimental sample, we additionally assessed the protocol-specific 10–15 Hz reward band and the 4–7 Hz and 20–30 Hz suppression bands. Reliability was excellent for 10–15 Hz power and good for both theta (4–7 Hz) and high-beta (20–30 Hz) power (further details in Fig. S12).
We also examined whether reliability was maintained over analysis windows more representative of real-time NFB. Band-power measures were therefore recomputed using non-overlapping 1-, 2-, 3-, 4- and 5-s windows. Fixed-window PAF was evaluated only from 2–5 s because 1-s epochs provide insufficient frequency resolution for stable centre-of-gravity estimation within the 8–12 Hz range. Reliability remained moderate-to-excellent across alpha, beta, theta, high beta, 10–15 Hz reward-band power and PAF. A feedback-relevant configuration using 2-s windows with 80% overlap, corresponding to an update interval of approximately 400 ms, likewise produced good-to-excellent reliability across metrics, including PAF (further details in Fig. S13).
Because the pooled clinical analysis combines repeated observations nested within participants and across two clinical cohorts, we additionally examined reliability at the individual-participant and cohort levels (further details in Tables S3–S4). Participant-level single-recording ICC(3,1) estimates varied across individuals and metrics, indicating that reliability was not uniform when participants were considered separately. Reliability was consistently higher when the five recordings/games were averaged, with ICC(3,k) values generally in the good-to-excellent range. Cohort-stratified ICC(3,1) estimates were also good-to-excellent, ranging from 0.74 to 0.88 in the neuropathic eye pain cohort and from 0.70 to 0.89 in the chronic low back pain cohort. Thus, the pooled clinical estimates provide useful summaries of measurement stability but should not be interpreted as indicating uniform reliability across all participants or sessions.
Cross-device consistency
PSD plots during EC testing (Fig. 5, top) show highly similar spectral profiles for the PainWaive and LiveAmp across sessions, with excellent consistency in spectral values across frequency bins (r > 90) between devices. During EO testing (Fig. 5, bottom), spectral consistency remained excellent (r > 90).
Mean z-transformed spectral power is shown across frequencies for PainWaive (blue) and LiveAmp (red) during (A) EC and (B) EO recordings. Shaded areas indicate 95% confidence intervals, and r denotes the mean participant-level Pearson correlation between devices.
For band-limited metrics (Fig. 6 left), consistency during EC was in the good range for alpha power, theta power and PAF, and fair range for beta power. In EO, consistency was in the good range for alpha power, moderate range for PAF, poor range for theta and fair range beta power.
Left panels show the primary/planned cross-device estimates under A eyes-closed and B eyes-open conditions. Overall spectral-shape correspondence is quantified using the mean participant-level Pearson correlation (r), while beta power, peak alpha frequency (PAF), alpha power and theta power are quantified using cross-device intraclass correlation coefficients (ICCs); 95% confidence intervals are shown for the primary estimates. Right panels summarise sensitivity analyses examining Spearman spectral-shape correlations, alternative spectral estimation and preprocessing approaches, reference matching, an EMG-aware exclusion pipeline, exclusion of recordings with flagged C1/C2 channels, individualised PAF estimation and low- and high-beta sub-bands where applicable. Grey cells indicate analyses that were not applicable or not performed. For visual consistency, coefficients are colour-coded as excellent (≥0.80), good (0.60–0.79), moderate (0.40–0.59), fair (0.21–0.39) and poor (≤0.20). Full details of the sensitivity analyses are provided in the Supplementary Material.
Findings of sensitivity analyses are shown in Fig. 6 (Right). To assess the extent to which cross-device agreement depended on preprocessing choices, we repeated the analyses across pipelines varying power transformation, taper and cleaning stringency: log-transformed absolute versus z-transformed power, DPSS versus Hanning tapers and relaxed versus rigorous cleaning (further details in Fig. S3). Alpha power and PAF were relatively robust to analytic choice, whereas beta and theta agreement was more pipeline-dependent. The primary z-scored Hanning pipeline yielded some of the lowest cross-device ICCs for beta and eyes-open theta. Replacing Hanning with DPSS within the z-scored pipelines produced modest improvements, particularly for eyes-closed beta and theta, whereas the largest increases for beta and eyes-open theta occurred under log-transformed pipelines. In particular, beta agreement increased from fair under the primary pipeline to moderate-to-good under log-transformed pipelines, while eyes-open theta increased from poor/fair to moderate.
Because PainWaive and LiveAmp used different reference configurations, we examined whether reference placement contributed to cross-device discrepancies by offline re-referencing LiveAmp to TP7, the closest available analogue to PainWaive’s ear-adjacent reference, and repeating the comparisons against the PainWaive primary pipeline (further details in Fig. S4). The overall pattern was preserved, although agreement improved for selected measures: eyes-closed PAF increased from good to excellent and eyes-closed beta from fair to moderate. Eyes-closed alpha decreased from good to moderate, while the remaining metrics were materially unchanged.
Because higher beta frequencies are particularly susceptible to EMG contamination, we examined whether the weaker beta agreement was driven predominantly by the upper beta range. Cross-device agreement was consistently stronger for low beta (13–20 Hz) than high beta (20–30 Hz), both during eyes-closed recordings (ICC = 0.59 vs 0.21) and eyes-open recordings (ICC = 0.61 vs 0.37; Fig. S6). We also applied a more stringent EMG-aware pipeline incorporating additional rejection of epochs with unusually high 30–45 Hz power. This produced a pattern similar to the primary analysis, indicating that additional exclusion of epochs with elevated high-frequency activity did not materially alter the principal cross-device findings (further details in Fig. S7).
To determine whether the a priori retention of C1/C2 in the LiveAmp preprocessing affected cross-device agreement, we repeated the analyses after excluding recordings in which either channel exceeded the variance criterion. As noted above, only 3 of 320 LiveAmp recordings (0.94%) were flagged, and their exclusion had negligible impact on the cross-device estimates (further details in Table S1 and Fig. S8).
The sensitivity of cross-device PAF agreement to the fixed 8–12 Hz estimation window was also examined using the Corcoran restingIAF approach62. Among recordings with a detectable individualised alpha peak, cross-device agreement was good-to-excellent (ICC = 0.83), providing convergent support for the primary fixed-window result despite reduced data availability (further details in Fig. S9).
To evaluate the robustness of the overall spectral-shape findings, we examined participant-level agreement and repeated the analysis using rank-based Spearman correlations. Most participants clustered within the high-agreement range, with only 7 eyes-closed and 8 eyes-open cases showing comparatively reduced correlations (further details in Fig. S10). Spearman correlations remained excellent in both conditions (mean ρ = 0.87 eyes closed; ρ = 0.84 eyes open), and all spectral-shape correlations remained significant following false-discovery-rate correction (further details in Table S2).
Finally, because high relative consistency does not necessarily imply agreement in absolute measurement values, we conducted Bland–Altman analyses to examine fixed and proportional cross-device bias (further details in Fig. S11). LiveAmp generally produced lower alpha power and higher beta power estimates than PainWaive, with relatively broad limits of agreement. PAF showed little average bias overall, although proportional bias was evident in eyes-closed recordings, whereas theta showed only a small negative bias and little evidence of proportional bias. These findings indicate that the devices captured broadly similar patterns but that their absolute values were not strictly interchangeable.
Impedance and hair characteristics
To further characterise the practical acquisition conditions of PainWaive, available impedance measurements and participant hair characteristics were examined (Supplementary Analysis 12). Across the available clinical recordings, cleaned C1 and C2 impedance values remained at or below the study threshold of 50 kΩ. Hair characteristics were heterogeneous across the experimental and clinical samples in terms of length, texture and density, providing descriptive evidence that acceptable electrode contact was achievable across a range of hair characteristics.
Discussion
This study aimed to evaluate the test–retest reliability and cross-device consistency of PainWaive, an in-house EEG headset recording from central scalp electrodes and developed for remote EEG monitoring and future NFB applications in chronic pain. PainWaive was developed to support repeated home-based EEG acquisition and future evaluation of NFB applications outside the laboratory, which requires evidence of measurement stability, cross-device consistency with research-grade systems and robustness to analytic variation. To address key limitations of prior consumer EEG research, we evaluated test–retest reliability and cross-device consistency under both EC and EO conditions, extended analyses to additional channels and a clinical sample, reported absolute reliability metrics and a priori sample size calculations, and developed a headset with impedance monitoring and saline-soaked electrodes rather than dry sensors. Across the experimental sample, test–retest reliability was generally good to excellent for eyes-closed recordings and moderate to excellent for eyes-open recordings, with broadly similar reliability across devices and sensitivity analyses. Reliability was also good to excellent in the clinical NFB sample and remained stable across shorter, NFB-relevant epoch windows and protocol-specific frequency bands. Cross-device consistency was strongest for spectral shape, which was excellent under both eyes-closed and eyes-open conditions, and for alpha power and PAF, which showed good agreement in eyes-closed recordings and moderate-to-good agreement in eyes-open recordings, with similar results across sensitivity analyses. By contrast, cross-device consistency was weaker for beta and theta, particularly in eyes-open recordings, although values improved under some alternative analytic choices. Overall, these findings suggest that PainWaive can provide stable estimates of EEG features, particularly spectral shape, alpha power and PAF, while also indicating that some band-limited comparisons, especially those involving beta and theta, are more sensitive to methodological decisions. Importantly, these analyses evaluate measurement reliability and cross-device consistency at the scalp level; they do not establish the anatomical or physiological origin of the recorded activity.
The achievable reliability of EEG metrics is shaped not only by device performance but also by the inherent fluctuations of different metrics and recording conditions, with EC typically exhibiting greater stability than EO. Acknowledging this, our values for test-retest reliability fell within the typical range reported for these metrics and conditions53,54,55,56,57,58,59,60,61, and they were comparable to those obtained with the research-grade LiveAmp system. This suggests that PainWaive shows an expected level of reliability for an EEG device. The findings also suggest potential reliability advantages of impedance monitoring and wet electrodes, although we did not directly compare PainWaive with a dry-electrode system, so firm conclusions about that design choice cannot be drawn. Our assessment of absolute reliability using SEM provides a useful contribution to the home-based EEG literature, as many prior studies do not report this metric9,16,25,27,39,41,42. Between-session absolute fluctuations across metrics were comparable for PainWaive and LiveAmp, indicating a similar magnitude of within-device measurement variability. Beyond reliability benchmarking, SEM estimates provide the basis for the minimal detectable difference for each spectral measure, which is essential for intervention design and sample size calculations. Collectively, these results provide valuable benchmarks for future studies using home-based EEG systems. A key shortcoming of laboratory-based validation is that it occurs under comparatively optimal conditions, including trained experimenters, controlled environments and immediate quality feedback. Real-world use is likely to introduce additional variability through less precise electrode placement, reduced impedance control, environmental noise and variable adherence. To bridge this gap, we also examined the reliability of the same metrics across NFB sessions conducted in two separate clinical studies, providing a real-world test of repeated home-based use of the device. The pooled clinical ICCs addressed whether, across the full set of clinical recordings, repeated values were consistent for a given recording/game. These pooled estimates showed good-to-excellent reliability across metrics, supporting the stability of PainWaive-derived spectral features during repeated home use in these clinical cohorts. However, this pooled approach collapses across participants and clinical cohorts, and therefore does not fully capture individual-level or cohort-level variability. Supplementary participant-level analyses showed moderate to good estimates (on average), indicating that reliability was not uniform across individuals. Reliability was generally higher when estimates were averaged across the five recordings/games, and cohort-stratified analyses showed good-to-excellent reliability in both clinical groups. Thus, the clinical analyses support the reliability of PainWaive measures in applied settings, but also indicate that pooled reliability estimates should be interpreted as descriptive summaries rather than evidence that all participants or clinical sessions are interchangeable. We also found that relevant spectral features were reliable at timescales relevant to real-time feedback. However, such stability may also reflect non-specific behavioural or physiological strategies, such as controlled breathing, reduced movement, or relaxation, rather than voluntary modulation of a targeted neural signal. Accordingly, while our findings support measurement reliability, they do not establish NFB efficacy, cortical specificity, or clinical effectiveness, all of which remain separate empirical questions for future studies.
Cross-device consistency is constrained by the reliability of the underlying spectral metrics ⁵, differences in headset setup (e.g. montage, reference, channel count and the influence of electrical signals from other channels), and variability introduced through reliability assessments. It is also important to note that research-grade systems, while often treated as the gold standard, remain subject to impedance variation, environmental noise and other artefacts30,33. Accordingly, estimates of cross-device consistency reflect broader physiological and methodological variability as well as device performance. Acknowledging these constraints, PainWaive showed excellent agreement in overall spectral profile (r’s > 0.9), with closely matched amplitude–frequency patterns across devices. This suggests that the PainWaive preserves the overall shape of the EEG spectrum and does not appear to selectively boost or attenuate specific frequencies resulting in highly similar relative spectral profiles, which is relevant to biomarkers that depend primarily on overall spectral shape63,64. For band-limited power metrics, our findings fit within a broader literature demonstrating varying degrees of cross-device consistency. Studies that assess cross device consistency during tasks, such as Jochumsen et al., have reported significant differences in signal-to-noise ratio across home-based systems (QuickCap, water-based MyndBand and cEEGrid), likely reflecting additional noise and device specific artefacts during movement initiation. Similarly, Mikhaylov et al.40 reported poor-to-fair correlations between MUSE 1 and ActiChamp Plus in a task-based paradigm suggesting that task demands may amplify variability and reduce comparability across systems. For studies that use resting-state recordings, Paillard et al. demonstrated that agreement between home-based dry-electrode and research-grade EEG systems is strongly frequency-dependent. Across devices (DSI-24, Q20r, zEEG), consistency was highest in the alpha band, while delta showed universally poor agreement (ICC < 0.23). Notably, the zEEG maintained strong agreement across alpha and beta bands (ICC = 0.76–0.94), whereas the DSI-24 and Q20r showed excellent alpha agreement (ICC > 0.89) but reduced consistency in higher beta ranges (ICC = 0.24–0.52). This pattern closely mirrors our findings, where cross-device consistency was similarly reduced in the higher beta band. While this pattern is consistent with our findings, it is not universal. For example, Cannard et al.12 reported excellent cross-device agreement in EC beta power between consumer and research-grade systems, with correlations of r = 0.84. Several methodological differences may explain differences in cross-device consistency for certain band metrics. First, reference scheme is likely important. In the present study, LiveAmp used FCz as the online reference, whereas PainWaive used an earlobe reference. Cannard et al. similarly demonstrated that mismatched referencing (e.g. wearable versus average reference) can reduce agreement, highlighting the reference dependence of spectral estimates. In our study, this possibility was supported by the supplementary re-referencing analysis, in which some cross-device estimates improved when LiveAmp data were re-referenced to a site more analogous to the PainWaive ear-adjacent reference. Second, sensor location may contribute. Cannard et al. tested a MUSE headset for frontal activity, whereas in our LiveAmp data, reliability across channels was generally lower at central scalp sites corresponding to the C1/C2 locations sampled by PainWaive. Consistent with this, the supplementary CSD analysis suggested that attenuation of spatially diffuse activity altered some estimates, including improving eyes-closed beta reliability, while not changing the overall pattern sufficiently to imply anatomical specificity. Third, signal scaling possibly plays a role. Our primary analyses used z-transformed relative power, which normalises each participant’s spectral profile and may compress between-band differences, reducing the apparent contribution of higher-frequency activity relative to alpha. In contrast, log-transformed absolute power avoids this relative scaling and produces stronger beta and theta agreement in our supplementary pipelines, suggesting that transformation choice influences cross-device estimates. Lastly, PAF appeared to be particularly sensitive to the estimation approach. An individualised approach using the Corcoran restingIAF toolbox produced higher ICCs for both test-retest reliability and cross-device agreement than the fixed-window PAF approach, suggesting that peak-based individualised estimation may improve the stability of PAF when a clear alpha peak is present. However, this improvement came at the cost of reduced data yield, as valid Corcoran-derived PAF estimates were only obtained for a subset of eyes-closed recordings. This indicates that individualised PAF estimation may provide a more stable estimate of dominant alpha frequency when a clear alpha peak is present, whereas fixed-window PAF offers greater data completeness and may provide a more pragmatic metric for device-level comparisons across the full sample. Overall, these findings indicate that cross-device EEG consistency is likely shaped by task context, referencing scheme, sensor placement and preprocessing pipeline. While our results are consistent with prior work showing robust alpha-band agreement and reduced reliability at higher frequencies, discrepancies across studies highlight the need for systematic evaluation of these methodological factors. Notably, our study extends the literature through the use of an a priori sample size calculation and a larger sample, providing more stable estimates of cross-device agreement12,37,40,42.
The primary contribution of this study is technical: a rigorous evaluation of the reliability and cross-device consistency of a lightweight central scalp EEG headset intended for repeated home-based use. As such, this work provides an empirical benchmark for central-scalp EEG applications that have received relatively limited validation, addressing a broader need identified in the home-based EEG literature for systematic device evaluation13,15,16,17,23,25,26,27,28,34,35,36,37,38,39,65. These findings therefore provide a technical foundation for future remote neurophysiology and NFB studies, rather than evidence of the clinical effectiveness of the system. The results are also encouraging for applications requiring stable measurement of central scalp spectral features over time. Importantly, however, within-device longitudinal reliability and cross-device interchangeability represent distinct measurement objectives. Cross-device agreement varied across measures, with spectral shape, alpha power and PAF showing the strongest cross-device correspondence, whereas beta and theta required more cautious interpretation. Supplementary Bland–Altman analyses further indicated that absolute values were not strictly interchangeable across devices, showing metric-specific fixed and, in some cases, proportional biases despite reasonable consistency in participant ranking. Taken together, these findings provide stronger support for within-device stability and longitudinal tracking than for direct substitution of PainWaive measurements with values obtained from a research-grade system.
Several limitations require acknowledgement. First, there is no true ‘ground truth’ EEG signal in our study. The combination of preserved within-device reliability and imperfect cross-device agreement for some metrics, especially under EO, makes it difficult to determine which device is closer to the underlying physiological signal. This challenge is compounded by the fact that scalp EEG reflects a spatially mixed signal shaped by volume conduction, tissue conductivity, head geometry and source orientation, rather than activity arising from a single cortical generator. Accordingly, the C1/C2 electrode positions used in PainWaive should not be interpreted as providing anatomically specific measurement of primary motor or primary somatosensory cortex. Indeed, activity recorded at C1/C2 may include contributions from spatially distributed generators, including posterior alpha sources and cannot be attributed specifically to sensorimotor cortex. While the supplementary CSD analysis showed that the overall reliability pattern at central sites was broadly preserved after attenuation of far-field activity, this does not establish anatomical source specificity. The present findings should therefore be interpreted as assessment of the reliability of central scalp potentials measured at C1/C2, rather than the anatomical source of the recorded signal or source-resolved sensorimotor cortical activity. Future work should adopt more spatially informative validation approaches, such as ICA decomposition with DIPFIT, beamforming, MRI-informed forward modelling, high-density EEG source modelling, or comparison with intracranial recordings where feasible, to determine the cortical generators contributing to PainWaive-recorded C1/C2 signals66. Second, the analysis focused on resting-state spectral power, which is a relatively forgiving test of device performance because the signals are slow-varying and averaged over multi-minute recordings. These results establish the reliability and cross-device consistency of central scalp spectral measures under resting conditions, but they do not determine whether the device can capture more demanding task-evoked or stimulus-locked responses that require greater temporal precision and functional specificity. Paradigms such as steady-state visually evoked potentials and motor imagery-related mu/beta event-related desynchronisation would provide stronger validation benchmarks and should be prioritised in future work. Accordingly, the present results support PainWaive as a resting-state central scalp EEG device, but do not yet establish its performance for task-evoked or active NFB applications. Third, the primary validation sample comprised young, healthy volunteers, which does not establish reliability in clinical pain populations. Although the sample of 80 adults provided a well-powered technical foundation for device evaluation, adults with chronic pain may differ in scalp properties, medication use and resting-state EEG dynamics, all of which could influence signal acquisition and spectral estimates. In addition, although the experimental sample included a range of hair densities, textures and lengths, young healthy volunteer samples may not fully capture the diversity of hairstyles and hair-related contact challenges likely to be encountered in other populations. Moreover, although we also included a clinical home-use sample, it was small, limiting generalisability. At the same time, the 160 NFB sessions collected from these participants provide preliminary evidence that the device can yield stable within-session EEG estimates during repeated home use in a clinical context. These repeated observations strengthen inference about within-device stability under real-world use, but they do not substitute for larger, more representative clinical validation studies. Accordingly, the present findings primarily support measurement reliability in healthy adults and provide preliminary evidence of reliability during repeated use in chronic pain, while the reliability, efficacy and spectral characteristics of the system in broader chronic pain populations remain unanswered questions for future research.
In conclusion, PainWaive demonstrated sufficient measurement stability to support repeated longitudinal central-scalp EEG monitoring outside the laboratory. The findings primarily support within-device tracking rather than direct numerical interchangeability with research-grade EEG, particularly for absolute band-limited measures. These results provide a technical foundation for future remote neurophysiology studies and prospective evaluation of home-based NFB applications in chronic pain.
Methods
Participants
For the experimental sample, eighty participants (47 female, mean age = 24, SD = 7.9) were recruited from the University of New South Wales (UNSW) Sydney via two research participation platforms. Eligible participants were aged ≥18 years and able to provide informed consent; no additional exclusion criteria were applied. Psychology undergraduates were recruited through the SONA-1 system and received course credit. Additional participants, not limited to psychology students, were recruited through the SONA-P system and received monetary ($50) compensation. All provided written informed consent, and the study was approved by the UNSW Human Research Ethics Advisory Panel (iRECS#8344). Participation was voluntary, with the option to withdraw at any time.
Power analysis primarily focused on ICC, as it is the most conservative and widely reported index in EEG reliability research. The required sample size was calculated using G*Power 3.1 and was based on the principle of sample required to detect a target ICC metric against a null hypothesis minimum value67. We anticipated that the PainWaive system would show ICCs spanning a broad range, from fair-to-moderate (0.30–0.60) for EO beta and theta cross-device consistency (representing the least consistent pairing), to good-to-excellent (0.8–0.99) for EC alpha power and PAF test–retest reliability (representing the most consistent pairing). For each sample size calculation, the null hypothesis was set at the lower bound of the relevant reliability category (e.g. ICC = 0.30 for beta/theta cross-device consistency), and the alternative hypothesis at the corresponding target value (e.g. ICC = 0.60 for beta/theta cross-device consistency). Using these assumptions, with a two-tailed α of 0.05 and 80% power, the largest sample size calculated was 56 participants (for detecting an ICC of 0.60 vs. a null of 0.30). To further increase power to 90% across all metrics and to safeguard against potential data loss, the final target sample size was set at 80 participants. Importantly, this study had no exclusion criteria due to the clinical safety of EEG and based on recommendations of the Ethics Advisory Panel.
For the clinical sample, eight adults with chronic neuropathic pain were recruited as part of two separate single-case experimental design (SCED) studies: four with neuropathic eye pain and four with chronic low back pain. Each cohort completed a SCED with multiple baselines, testing 20 sessions of home-based, self-directed EEG NFB using the PainWaive headset. Participants acted as their own controls with daily clinical outcomes recorded across baseline, intervention, post-intervention and follow-up phases. Participants and researchers were not blinded. This report focuses solely on the reliability of EEG metrics recorded during the intervention phase pooled across both cohorts; clinical outcomes for the eye pain cohort have been published previously,68 and clinical outcomes for the low back pain cohort will be reported separately.
Participants with neuropathic eye pain were 18–80 years old, had a clinical diagnosis of ocular neuropathic pain for ≥6 months, average pain ≥3/10 over the past week, were fluent in English and had no planned changes to treatment or ocular procedures during the 12-week trial. Participants with chronic low back pain were ≥18 years, living in Australia, had ongoing low-back pain for ≥3 months, and average pain ≥4/10 over the past week. Exclusion criteria included severe neurodegenerative or psychiatric disorders, planned pain surgery, concurrent trial participation, or implanted cardiac/neurological devices. All procedures received institutional ethics approval, and all participants provided written informed consent. Studies were approved by the UNSW Human Research Ethics Advisory Panel (#iRECS8134 and HC230351)
Protocol
For the experimental sample, a 2 × 2 × 2 repeated-measures design was used, with 2 headsets (PainWaive, LiveAmp), 2 conditions (EC, EO), and 2 sessions. Because recordings had to be completed in device blocks, condition order was constrained such that participants completed both resting-state tasks (EC and EO) on one headset before switching to the other headset. Accordingly, order was randomly determined at the level of device block and task order within each device block, rather than by freely permuting all four device-condition combinations. This yielded eight possible valid sequences: two device orders (LiveAmp→PainWaive or PainWaive→LiveAmp), each combined with two possible EC/EO orders within the first device block and two possible EC/EO orders within the second device block. Participants were assigned one of these valid sequences using a pre-generated scheduling script (code provided in the Supplementary Material).
For the clinical sample, each of the four participants from each cohort were randomly allocated (computer-generated) to staggered baseline lengths of 7, 10, 14, or 17 days, with lengths differing by ≥3 days. All participants from each cohort began baseline on the same day. The intervention comprised 20 home-based EEG NFB sessions over 4 weeks.
Data collection procedures—experimental sample
Participants were instructed to avoid hair products and caffeine before their session to optimise electrode conductivity and recording quality.
For preparation of the LiveAmp, prior to recording, the headset and electrode sponges were soaked in a saline solution. EEG was recorded with BrainVision Recorder (v1.27.0001, Brain Products, Germany) at a 250 Hz sampling rate using the 64-channel Live-Amp headset (Fig. 7a). The LiveAmp was arranged according to the international 10–20 system (Fig. 7b), with the ground electrode at FPz and the reference at FCz. Electrodes C1 and C2, from which key spectral measures were derived, were carefully monitored to meet this threshold, and ensure high signal quality, matching criteria between LiveAmp and PainWaive. Impedance was kept below 50 kΩ, a threshold supported by previous research and the LiveAmp manual as unlikely to cause signal loss69,70.
A LiveAmp EEG headset used for data collection. B 64-channel electrode layout based on the international 10–20 system, which positions electrodes at standardised intervals across the scalp to ensure consistent coverage of cortical regions. Reference electrodes (FCz, red arrow) and ground electrodes (FPz, green arrow) are indicated on the LiveAmp headset.
For preparation of the PainWaive, before recording, electrode sponges and ear clips were soaked in a saline solution. EEG was collected using the PainWaive headset (Fig. 8) and OpenBCI GUI (v5.2.0, OpenBCI, USA) at a sampling rate of 200 Hz. The headset had two recording silver chloride electrodes mounted in a 3D-printed frame positioned at C1 and C2, in accordance with the international 10-20 system, with ground and reference electrodes integrated into left and right ear clips2. Small Hydro-Link sponges are used to complete the connection between the scalp and the electrodes. Details about the PainWaive headset are provided in the supplementary materials. Impedances were maintained below 50 kΩ, to match the LiveAmp set-up using the OpenBCI GUI (v5.2.0). Written informed consent was obtained from the individual depicted in Fig. 8 for publication of the image. The image was cropped to minimise identifiability while retaining the device-placement information necessary to illustrate the PainWaive headset.
Orange arrows indicate recording electrodes over C1 and C2. Red arrow marks the reference ear clip, and green arrow marks the ground ear clip. The image was cropped to minimise identifiability while preserving the relevant device-placement information. Written informed consent was obtained from the individual depicted for publication of this image.
Before each recording session, participants completed a standardised sequence of movements and facial actions (eyes up/down, left/right, blinking, jaw clenching and 30 s of eye closure) to identify artefact and noisy channels. The recording protocol included two 3-min tasks: (a) fixation on a central cross with eyes open (EO condition) and (b) eyes closed while remaining awake (EC condition).
Data collection procedures—clinical sample
During each home session, EEG was recorded from C1 and C2 after verifying impedances68. After impedance verification, each session began with a 2-min EO resting-state recording while participants fixated on a central cross. Relative power in the protocol-defined NFB bands was then used to set the initial thresholds for the session, after which participants completed five 2.5-min NFB rounds targeting suppression of theta (4–7 Hz) and high beta (20–30 Hz) while reinforcing the reward band (10–15 Hz). In the real-time implementation, thresholds were fixed within session, and feedback was updated every 1.2 s using 1.2 s analysis windows. Four types of game scenarios provided visual feedback via background colour and character motion (Jellyfish, Plane, Bird and Rocket) with successful modulation making the flying object move and background brighten. Participants were instructed to minimise movement and use recommended strategies - positive thinking, focusing on game features and slow breathing exercises to regulate the targeted bands; raw EEG was not displayed. Optional background music supported relaxation. Adherence and quality control were managed via a secure dashboard. At the start of each session, the tablet captured a headset-placement photo; EEG and de-identified photos were automatically uploaded and linked to participant IDs. The research team reviewed uploads after each session to confirm placement and screen for artefacts, contacting participants with troubleshooting as needed68.
EEG data pre-processing pipeline
Data were pre-processed using MATLAB (vR2024b), EEGLAB (v2023_1) and FieldTrip (v20200215). LiveAmp recordings were down-sampled from 250 Hz to 200 Hz to match the PainWaive. Device comparisons focused solely on the C1 and C2 channels to match the PainWaive montage. In the primary pipeline, continuous data from both devices were first cleaned for line noise using EEGLAB’s CleanLine at 50 and 100 Hz and then band-pass filtered from 2 to 45 Hz using EEGLAB’s pop_eegfiltnew default finite impulse response implementation, with filter order selected automatically by the function. For LiveAmp recordings, channels were screened for excessive noise by computing the standard deviation of each channel time series within recording, z-scoring these values across channels and flagging channels with |z| > 3. Flagged LiveAmp channels were removed, except that C1 and C2 were always retained a priori to enable direct device comparisons. No analogous bad-channel exclusion rule was applied to PainWaive because it contained only two channels. Data were then segmented into consecutive non-overlapping 5-s epochs using FieldTrip. Epochs were rejected if any included channel exceeded ±150 µV. In the primary pipeline, no additional per-epoch detrending was applied before spectral estimation. The main device-comparison analyses were based on C1 and C2 only.
EEG data post-processing pipeline
Spectral analyses were performed in FieldTrip using ft_freqanalysis with method = ‘mtmfft’, taper = ‘hanning’, tapsmofrq = 1, keeptrials = ‘yes’, and frequencies of interest from 2 to 45 Hz in 0.2-Hz steps. Because data had already been segmented into non-overlapping 5-s epochs, epoch overlap was 0%. For each retained epoch, power spectra were extracted from C1 and C2, and power was averaged across these two channels within epoch. The resulting C1/C2 spectra were then averaged across retained epochs to obtain a single participant-level mean power spectral density (PSD) for each device, session, and task condition. Z-transformation was applied to this participant-level mean PSD across the full analysed frequency range (2–45 Hz); thus, z-scoring was not applied separately to each epoch. Note that we selected this z-transformed spectral pipeline as the primary analysis because it standardises each participant’s mean spectral profile across the analysed frequency range, thereby reducing the influence of absolute amplitude scaling differences between participants and devices while preserving the relative within-participant spectral shape used to derive the band-limited measures. Theta, alpha and beta power were then computed by summing the z-transformed spectrum across bins within 4–7 Hz, 8–12 Hz and 13–30 Hz, respectively. Fixed-window PAF was computed from the raw participant-level mean PSD using a centre-of-gravity approach within 8–12 Hz using Eq. 1:
$${CoG}=\frac{{\sum }_{i=1}^{n}\,{f}_{i}{* a}_{i}}{{\sum }_{i=1}^{n}{a}_{i}}$$
(1)
The fixed window was informed by PAF values observed in our previous studies of healthy young participants61,71,72. Where fi is the ith frequency bin within the frequency window of interest, and ai represents the spectral amplitude at fi.
Statistical analysis—experimental sample
We computed means and SDs per session, condition and headset. Normality was assessed using skewness and kurtosis. Distributions with skewness between ±2 and kurtosis below 3.29 were considered approximately normal for medium-sized samples, consistent with guidelines73,74. All analyses were conducted in JASP (v0.18.30) and MATLAB (R2024a).
Absolute reliability, which reflects measurement error and within-individual variability, was assessed using the standard error of measurement (SEMeas); lower SEMeas values indicate better measurement precision and reduced within-individual variability75,76. The following formula was used: SEMeas = √mean squared error, where the mean squared error (or ‘residual error’) was obtained from one-way repeated measures analysis of variance applied to test and retest measurements77,78,79
Relative test–retest reliability was quantified using ICC based on a two-way mixed-effects, single-measurement, consistency model, following the Shrout & Fleiss and McGraw & Wong conventions80,81. This model treats participants as random targets and measurement occasions as fixed, reflecting interest in the consistency of measurements across the specific repeated sessions used in the study. The ICC quantifies the proportion of the total variance that is attributable to true between-subject differences relative to total variance. ICCs were computed in MATLAB (R2024b) using a custom script and interpreted according to established thresholds82,83, including poor (≤0.2), fair (0.2–0.39), moderate (0.4–0.59), good (0.60–0.79) and excellent (0.8–0.99).
To determine whether poor reliability at C1/C2 might simply reflect a particularly noisy scalp region, we also calculated test–retest reliability for all 64 channels of the LiveAmp system. For each channel and EEG metric, we estimated two-way mixed-effects, single-measure, consistency ICCs, generating topographical maps of relative reliability across the scalp.
Cross-device consistency between the 2-channel device and the 64-channel LiveAmp was quantified using ICCs based on a two-way random-effects, single-measurement, absolute-agreement model, again following Shrout & Fleiss and McGraw & Wong80,81. This model assumes that both devices are sampled from a larger population of possible measurement systems and assesses absolute agreement between them. To further characterise cross-device agreement, Pearson correlation coefficients (r) were calculated between the full spectral profiles recorded with each headset to assess the similarity of the spectral shape. Together, these indices provide a comprehensive assessment of both individual rankings (ICC) and the similarity of spectral distributions (r) across devices.
Statistical analysis—clinical sample
Participants across both SCEDs completed 20 home-based NFB sessions (total 160 sessions), each comprising five recordings (‘games’), with each game separated by ~1 min. We treated each NFB session as a separate sampling point and asked: Do the five recordings within a session yield similar EEG values (alpha, beta, theta power and PAF)? To answer this, we pooled all available sessions across participants and constructed a session × recordings matrix for each metric (rows = sessions, columns = Games 1–5). We then computed a two-way mixed-effects, single-measurement, consistency ICC. In this framework, sessions are the targets and games are repeated measurements. The resulting pooled ICC for each metric, therefore, reflects the extent to which a single recording provides a stable estimate of the session’s EEG value under the same conditions and is interpreted as our primary device-performance index.
Supplementary analysis
Details of supplementary analyses are provided in the Supplementary Material. Briefly, we performed sensitivity analyses to examine the robustness of the main findings to alternative preprocessing pipelines, differences in reference configuration, attenuation of volume-conducted activity using CSD, possible EMG contamination of the beta band, exclusion of recordings with flagged C1/C2 channels, and individualised rather than fixed-window PAF estimation. We also evaluated the robustness of spectral-shape correlations using participant-level diagnostics, Spearman correlations, and false discovery rate correction, and complemented ICC-based cross-device analyses with Bland-Altman plots. Finally, in the clinical dataset, we examined the reliability of protocol-specific targeted bands, repeated the analyses across shorter epoch windows relevant to real-time NFB and present cohort level and participant level ICCs.
Data availability
De-identified data are available to qualified researchers on reasonable request, in accordance with the ethics approvals and study protocols governing this research.
References
Niedermeyer, E. & da Silva, F. L. Electroencephalography: Basic Principles, Clinical Applications, and Related Fields (Lippincott Williams & Wilkins, 2005).
Hesam-Shariati, N. et al. A home-based self-directed EEG neurofeedback intervention for people with chronic neuropathic pain following spinal cord injury (the StoPain Trial): description of the intervention. Spinal Cord. 62, 658–666 (2024).
Hesam-Shariati, N. et al. The analgesic effect of electroencephalographic neurofeedback for people with chronic pain: a systematic review and meta-analysis. Eur. J. Neurol. 29, 921–936 (2022).
McLane, H. C. et al. Availability, accessibility, and affordability of neurodiagnostic tests in 37 countries. Neurology 85, 1614–1622 (2015).
Ney, J. P. et al. The cost of after-hour electroencephalography. Neurol. Clin. Pract. 14, e200264 (2024).
Mushtaq, F. & Ibáñez, A. Electroencephalography (EEG) and the Quest for an Inclusive and Global Neuroscience. Eur. J. Neurosci. 61, e70078 (2025).
Yun, S. Advances, challenges, and prospects of electroencephalography-based biomarkers for psychiatric disorders: a narrative review. J. Yeungnam Med. Sci. 41, 261–268 (2024).
Rashid, M. et al. Current status, challenges, and possible solutions of EEG-based brain-computer interface: a comprehensive review. Front. Neurorobotics 14, 25 (2020).
Ratti, E., Waninger, S., Berka, C., Ruffini, G. & Verma, A. Comparison of medical and consumer wireless EEG systems for use in clinical trials. Front. Hum. Neurosci. 11, 398 (2017).
Mikkelsen, K. B., Tabar, Y. R., Christensen, C. B. & Kidmose, P. EEGs vary less between lab and home locations than they do between people. Front. Comput. Neurosci. 15, 2021 https://doi.org/10.3389/fncom.2021.565244 (2021).
Sugden, R. J., Pham-Kim-Nghiem-Phu, V. L., Campbell, I., Leon, A. & Diamandis, P. Remote collection of electrophysiological data with brain wearables: opportunities and challenges. Bioelectron. Med 9, 12 https://doi.org/10.1186/s42234-023-00114-5 (2023).
Cannard, C., Wahbeh, H. & Delorme, A. In 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) 3603–3610 (IEEE, 2021).
Krigolson, O. E., Williams, C. C., Norton, A., Hassall, C. D. & Colino, F. L. Choosing MUSE: validation of a low-cost, portable EEG system for ERP research. Front. Neurosci. 11, 2017, https://doi.org/10.3389/fnins.2017.00109 (2017).
Krigolson, O. E. et al. Using Muse: rapid mobile assessment of brain performance. Front. Neurosci. 15, 634147 (2021).
LaRocco, J., Le, M. D. & Paeng, D.-G. A systemic review of available low-cost EEG headsets used for drowsiness detection. Front. Neuroinformatics 14, 553352 (2020).
Biondi, A. et al. Feasibility and acceptability of an ultra-long-term at-home EEG monitoring system (EEG@ HOME) for people with epilepsy. Epilepsy Behav. 151, 109609 (2024).
Wong, A., Walishetti, A. & Yang, Y. Evaluating home-based EEG systems for public health telemonitoring: a scoping review. In 2024 IEEE Healthcare Innovations and Point of Care Technologies (HI-POCT), 5–8 (Institute of Electrical and Electronics Engineers, 2024).
Jensen, M. et al. Brain EEG activity correlates of chronic pain in persons with spinal cord injury: clinical implications. Spinal cord. 51, 55–58 (2013).
Mussigmann, T., Bardel, B. & Lefaucheur, J.-P. Resting-state electroencephalography (EEG) biomarkers of chronic neuropathic pain. A systematic review. Neuroimage 258, 119351 (2022).
Boord, P. et al. Electroencephalographic slowing and reduced reactivity in neuropathic pain following spinal cord injury. Spinal cord. 46, 118–123 (2008).
Bhavnani, S. et al. The acceptability, feasibility, and utility of portable electroencephalography to study resting-state neurophysiology in rural communities. Front. Hum. Neurosci. 16, 802764 (2022).
Radüntz, T. Signal quality evaluation of emerging EEG devices. Front. Physiol. 9, 98 (2018).
Milne-Ives, M. et al. At home EEG monitoring technologies for people with epilepsy and intellectual disabilities: a scoping review. Seizure.: Eur. J. Epilepsy 110, 11–20 (2023).
Müller-Putz, G. R., Tunkowitsch, U., Minas, R. K., Dennis, A. R. & Riedl, R. in NeuroIS Retreat 90–95 (Springer, 2021).
D’Angiulli, A., Lockman-Dufour, G. & Buchanan, D. M. Promise for personalized diagnosis? Assessing the precision of wireless consumer-grade electroencephalography across mental states. Appl. Sci. 12, 6430 (2022).
Riedl, R., Minas, R. K., Dennis, A. R. & Müller-Putz, G. R. Consumer-grade EEG instruments: insights on the measurement quality based on a literature review and implications for NeuroIS research. in NeuroIS Retreat 350–361 (Springer, 2020).
Maskeliunas, R., Damasevicius, R., Martisius, I. & Vasiljevas, M. Consumer-grade EEG devices: are they usable for control tasks?. PeerJ 4, e1746 (2016).
Barbey, F. M. et al. Neuroscience from the comfort of your home: repeated, self-administered wireless dry EEG measures brain function with high fidelity. Front. Digit. Health 4, 944753 (2022).
Nuwer, M. R. Quantitative EEG: I. Techniques and problems of frequency analysis and topographic mapping. J. Clin. Neurophysiol. 5, 1–44 (1988).
Maswanganyi, R. C., Tu, C., Owolawi, P. A. & Du, S. Statistical evaluation of factors influencing inter-session and inter-subject variability in EEG-based brain computer interface. IEEE Access 10, 96821–96839 (2022).
Wei, C.-S., Lin, Y.-P., Wang, Y.-T., Lin, C.-T. & Jung, T.-P. A subject-transfer framework for obviating inter-and intra-subject variability in EEG-based drowsiness detection. NeuroImage 174, 407–419 (2018).
Ma, B.-Q., Li, H., Zheng, W.-L. & Lu, B.-L. in International Conference on Neural Information Processing 30–42 (Springer, 2019).
Saha, S. & Baumert, M. Intra-and inter-subject variability in EEG-based sensorimotor brain computer interface: a review. Front. Comput. Neurosci. 13, 87 (2020).
Craik, A. et al. Design and validation of a low-cost mobile EEG-based brain–computer interface. Sensors 23, 5930 (2023).
Arnal, P. J. et al. The Dreem Headband compared to polysomnography for electroencephalographic signal acquisition and sleep staging. Sleep 43 https://doi.org/10.1093/sleep/zsaa097 (2020).
Khng, K. H. & Mane, R. Beyond BCI—Validating a wireless, consumer-grade EEG headset against a medical-grade system for evaluating EEG effects of a test anxiety intervention in school. Adv. Eng. Inform. 45, 101106 (2020).
Jochumsen, M., Knoche, H., Kjaer, T. W., Dinesen, B. & Kidmose, P. EEG headset evaluation for detection of single-trial movement intention for brain-computer interfaces. Sensors 20, 2804 (2020).
Radüntz, T. Signal quality evaluation of emerging EEG devices. Front. Physiol. 9, 98 https://doi.org/10.3389/fphys.2018.00098 (2018).
Titgemeyer, Y. et al. Can commercially available wearable EEG devices be used for diagnostic purposes? An explorative pilot study. Epilepsy Behav. 103, 106507 https://doi.org/10.1016/j.yebeh.2019.106507 (2020).
Mikhaylov, D. et al. Comparison of EEG signal spectral characteristics obtained with consumer-and research-grade devices. Sensors 24, 8108 (2024).
Larson, M. J. & Carbine, K. A. Sample size calculations in human electrophysiology (EEG and ERP) studies: a systematic review and recommendations for increased rigor. Int. J. Psychophysiol. 111, 33–41 (2017).
Dickinson, A., Perry, S. R., Gulsrud, A. & Kasari, C. A direct comparison of laboratory and community EEG recordings for neurodevelopmental research. Sci. Rep. 15, 22671 https://doi.org/10.1038/s41598-025-03569-5 (2025).
Büchel, D., Lehmann, T., Sandbakk, Ø & Baumeister, J. EEG-derived brain graphs are reliable measures for exploring exercise-induced changes in brain networks. Sci. Rep. 11, 20803 (2021).
Ehrhardt, N. M. et al. Comparison of dry and wet electroencephalography for the assessment of cognitive evoked potentials and sensor-level connectivity. Front. Neurosci. 18, 1441799 (2024).
Lopez-Gordo, M. A., Sanchez-Morillo, D. & Valle, F. P. Dry EEG electrodes. Sensors 14, 12847–12870 (2014).
Mathewson, K. E., Harrison, T. J. L. & Kizuk, S. A. D. High and dry? Comparing active dry EEG electrodes to active and passive wet electrodes. Psychophysiology 54, 74–82 https://doi.org/10.1111/psyp.12536 (2017).
Shad, E. H. T., Molinas, M. & Ytterdal, T. Impedance and noise of passive and active dry EEG electrodes: a review. IEEE Sens. J. 20, 14565–14577 https://doi.org/10.1109/JSEN.2020.3012394 (2020).
Fiedler, P. et al. Comparison of three types of dry electrodes for electroencephalography. ACTA IMEKO 3, 33–37 https://doi.org/10.21014/acta_imeko.v3i3.94 (2014).
Faadhilah Afif, N., Harke Pratama, S., Haryanto, F., Nurul Khotimah, S. & Suprijadi Comparison of wet and dry EEG electrodes based on brain signals characterization in temporal and anterior frontal areas using audio stimulation. J. Phys. Conf. Ser. 1505, 012069 https://doi.org/10.1088/1742-6596/1505/1/012069 (2020).
Zander, T. O. et al. A Dry EEG-system for scientific research and brain–computer interfaces. Front. Neurosci. 5, 2011 https://doi.org/10.3389/fnins.2011.00053 (2011).
O’Sullivan, M. et al. Comparison of electrode technologies for dry and portable EEG acquisition. In 2017 7th IEEE International Workshop on Advances in Sensors and Interfaces (IWASI), 15–20 (Institute of Electrical and Electronics Engineers, 2017).
Yeung, A. et al. Comparison of foam-based and spring-loaded dry EEG electrodes with wet electrodes in resting and moving conditions. In 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 7131–7134 (Institute of Electrical and Electronics Engineers, 2015).
McKeown, D. J. et al. Test-retest reliability of spectral parameterization by 1/f characterization using SpecParam. Cereb. Cortex 34, bhad482 (2024).
Pollock, V., Schneider, L. & Lyness, S. Reliability of topographic quantitative EEG amplitude in healthy late-middle-aged and elderly subjects. Electroencephalogr. Clin. Neurophysiol. 79, 20–26 (1991).
Salinsky, M., Oken, B. & Morehead, L. Test-retest reliability in EEG frequency analysis. Electroencephalogr. Clin. Neurophysiol. 79, 382–392 (1991).
Kondacs, A. & Szabó, M. Long-term intra-individual variability of the background EEG in normals. Clin. Neurophysiol. 110, 1708–1716 (1999).
Edgar, J. C. et al. A comparison of resting-state eyes-closed and dark-room alpha-band activity in children. Psychophysiology 60, e14285 (2023).
McEvoy, L., Smith, M. & Gevins, A. Test–retest reliability of cognitive EEG. Clin. Neurophysiol. 111, 457–463 (2000).
Gasser, T., Bächer, P. & Steinberg, H. Test-retest reliability of spectral parameters of the EEG. Electroencephalogr. Clin. Neurophysiol. 60, 312–319 (1985).
Ip, C.-T. et al. Pre-intervention test-retest reliability of EEG and ERP over four recording intervals. Int. J. Psychophysiol. 134, 30–43 (2018).
Chowdhury, N. S. et al. The reliability of two prospective cortical biomarkers for pain: EEG peak alpha frequency and TMS corticomotor excitability. J. Neurosci. methods 385, 109766 (2023).
Corcoran, A. W., Alday, P. M., Schlesewsky, M. & Bornkessel-Schlesewsky, I. Toward a reliable, automated method of individual alpha frequency (IAF) quantification. Psychophysiology 55, e13064 (2018).
Avila, C. G. et al. Assessing the balance between excitation and inhibition in chronic pain through the aperiodic component of EEG. Elife 13, RP101727 (2025).
González-Villar, A. J., Samartin-Veiga, N., Arias, M. & Carrillo-de-la-Peña, M. T. Increased neural noise and impaired brain synchronization in fibromyalgia patients during cognitive interference. Sci. Rep. 7, 5841 (2017).
Lau-Zhu, A., Lau, M. P. & McLoughlin, G. Mobile EEG in research on neurodevelopmental disorders: opportunities and challenges. Dev. Cogn. Neurosci. 36, 100635 (2019).
Barzegaran, E., Bosse, S., Kohler, P. J. & Norcia, A. M. EEGSourceSim: a framework for realistic simulation of EEG scalp data using MRI-based forward models and biologically plausible signals and noise. J. Neurosci. methods 328, 108377 (2019).
Borg, D. N., Bach, A. J., O’Brien, J. L. & Sainani, K. L. Calculating sample size for reliability studies. PMR 14, 1018–1025 (2022).
Hesam-Shariati, N. et al. The effect of an EEG neurofeedback intervention for corneal neuropathic pain: A single-case experimental design with multiple baselines. J. Pain 32, 105394 (2025).
Ferree, T. C., Luu, P., Russell, G. S. & Tucker, D. M. Scalp electrode impedance, infection risk, and EEG data quality. Clin. Neurophysiol. 112, 536–544 https://doi.org/10.1016/s1388-2457(00)00533-2 (2001).
Kappenman, E. S. & Luck, S. J. The effects of electrode impedance on data quality and statistical significance in ERP recordings. Psychophysiology 47, 888–904 https://doi.org/10.1111/j.1469-8986.2010.01009.x (2010).
Chowdhury, N. S. et al. Predicting individual pain sensitivity using a novel cortical biomarker signature. JAMA Neurol. 82, 237–246 (2025).
Chowdhury, N. S. et al. A 5-day course of repetitive transcranial magnetic stimulation before pain onset ameliorates future pain and increases sensorimotor peak alpha frequency. Pain 166, 1382–1394 (2025).
Ghasemi, A. & Zahediasl, S. Normality tests for statistical analysis: a guide for non-statisticians. Int. J. Endocrinol. Metab. 10, 486 (2012).
Kim, H.-Y. Statistical notes for clinical researchers: assessing normal distribution (2) using skewness and kurtosis. Restor. Dent. Endod. 38, 52 (2013).
Weir, J. P. Quantifying test-retest reliability using the intraclass correlation coefficient and the SEM. J. Strength Cond. Res. 19, 231–240 (2005).
Atkinson, G. & Nevill, A. M. Statistical methods for assessing measurement error (reliability) in variables relevant to sports medicine. Sports Med. 26, 217–238, https://doi.org/10.2165/00007256-199826040-00002 (1998).
Matamala, J. M. et al. Inter-session reliability of short-interval intracortical inhibition measured by threshold tracking TMS. Neurosci. Lett. 674, 18–23 https://doi.org/10.1016/j.neulet.2018.02.065 (2018).
Lexell, J. E. & Downham, D. Y. How to assess the reliability of measurements in rehabilitation. Am. J. Phys. Med. Rehabil. 84, 719–723 (2005).
Weir, J. P. Quantifying test-retest reliability using the intraclass correlation coefficient and the SEM. J. Strength Cond. Res. 19, 231–240 https://doi.org/10.1519/15184.1 (2005).
McGraw, K. O. & Wong, S. P. Forming inferences about some intraclass correlation coefficients. Psychol. methods 1, 30 (1996).
Shrout, P. E. & Fleiss, J. L. Intraclass correlations: uses in assessing rater reliability. Psychol. Bull. 86, 420–428 https://doi.org/10.1037/0033-2909.86.2.420 (1979).
Portney, L. G. & Watkins, M. P. Foundations of Clinical Research: Applications to Practice Vol. 892 (Pearson/Prentice Hall Upper Saddle River, 2009).
Chowdhury, N. S., Chang, W.-J., Cavaleri, R., Chiang, A. K. & Schabrun, S. M. The reliability and validity of rapid transcranial magnetic stimulation mapping for muscles under active contraction. BMC Neurosci. 25, 43 (2024).
Acknowledgements
The authors acknowledge the Traditional Owners of the land on which this work was conducted, the Bidjigal and Bedegal peoples of the Eora Nation. We pay deep respect to Elders past and present, and extend that respect to all Aboriginal and Torres Strait Islander peoples. We recognise their continuing connection to land, waters, and community, and acknowledge that sovereignty was never ceded. This always was and always will be Aboriginal land. The authors would also like to thank Prof. Lorimer Moseley for his helpful comments on early versions of the manuscript.
Funding
This work was supported by National Health and Medical Research Council Ideas Grant and Rebecca L. Cooper Medical Research Foundation Fellowship awarded to Sylvia M Gustin. The funders had no role in study design, data collection, manuscript preparation or the decision to submit the manuscript for publication.
Author information
Author notes
These authors contributed equally: Nahian S. Chowdhury and Joshua Rawsthorne.
Authors and Affiliations
Centre for Pain IMPACT, Neuroscience Research Australia, Sydney, NSW, Australia
Nahian S. Chowdhury, Negin Hesam-Shariati, Yann Quidé, Lara Alexander, Ekaterina Ermolenko, James H. McAuley & Sylvia M. Gustin
NeuroRecovery Research Hub, School of Psychology, University of New South Wales, Sydney, NSW, Australia
Nahian S. Chowdhury, Joshua Rawsthorne, Negin Hesam-Shariati, Yann Quidé, Lara Alexander, Ekaterina Ermolenko, Angus McIntyre, Sebastian Restrepo, Kevin Chen & Sylvia M. Gustin
School of Computer Science, Faculty of Engineering and Information Technology, University of Technology Sydney, Sydney, NSW, Australia
Chin-Teng Lin
Faculty of Health, University of Technology Sydney, Sydney, NSW, Australia
Toby Newton-John
John Walsh Centre for Rehabilitation Research, The Kolling Institute, Faculty of Medicine and Health, University of Sydney, Sydney, NSW, Australia
James W. Middleton & Ashley Craig
Department of Rehabilitation Medicine, University of Washington, Seattle, WA, USA
Mark P. Jensen
School of Health Science, University of New South Wales, Sydney, NSW, Australia
James H. McAuley
Authors
- Nahian S. Chowdhury
- Joshua Rawsthorne
- Negin Hesam-Shariati
- Yann Quidé
- Lara Alexander
- Ekaterina Ermolenko
- Angus McIntyre
- Sebastian Restrepo
- Kevin Chen
- Chin-Teng Lin
- Toby Newton-John
- James W. Middleton
- Ashley Craig
- Mark P. Jensen
- James H. McAuley
- Sylvia M. Gustin
Contributions
N.S.C.: Conceptualisation, data curation, formal analysis, investigation, methodology, software, validation, visualisation, writing (original draft) and writing (reviewing and editing). J.R.: Conceptualisation, data curation, formal analysis, investigation, methodology, software, validation, visualisation, writing (original draft) and writing (reviewing and editing). N.H.S.: Conceptualisation, validation, visualisation and writing (reviewing and editing). Y.Q.: Conceptualisation, validation, visualisation and writing (reviewing and editing). L.A.: Investigation, methodology, writing (reviewing and editing). E.E.: Investigation, methodology, writing (reviewing and editing). A.M.: Software/Hardware and writing (reviewing and editing). S.R.: Software/Hardware and writing (reviewing and editing). K.C.: Software/Hardware and writing (reviewing and editing). C.T.L.: Funding acquisition and writing (reviewing and editing). T.N.J.: Funding acquisition and writing (reviewing and editing). J.W.M.: Funding acquisition and writing (reviewing and editing). A.C.: Funding acquisition and writing (reviewing and editing). M.J.: Funding acquisition and writing (reviewing and editing). J.M.: Funding acquisition and writing (reviewing and editing). S.M.G.: Funding acquisition, supervision, validation, visualisation, writing (original draft) and writing (reviewing and editing).
Corresponding author
Correspondence to Nahian S. Chowdhury.
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Chowdhury, N.S., Rawsthorne, J., Hesam-Shariati, N. et al. Evaluation of PainWaive: a consumer-grade EEG headset for remote EEG monitoring and neurofeedback research in chronic pain. npj Digit. Med. 9, 864 (2026). https://doi.org/10.1038/s41746-026-03424-5
Received: 06 February 2026
Accepted: 07 October 2026
Published: 06 November 2026
Version of record: 06 November 2026
DOI: https://doi.org/10.1038/s41746-026-03424-5
来源:Nature 健康护理主题 · nature.com