EEG Usability Evaluation
- EEGUsability is a method that integrates EEG-based metrics with standard usability tools to assess interface demand and data quality.
- It applies precise temporal measures and diverse signal-processing techniques to evaluate cognitive load, attention, and error recognition in various HCI contexts.
- The approach leverages machine learning and standardized preprocessing pipelines to improve the objectivity of usability assessments across consumer and research settings.
Searching arXiv for recent and foundational papers on EEG-based usability evaluation and the specific term "eegUsability". eegUsability denotes the use of electroencephalography as an objective, temporally precise complement to questionnaires, logs, and performance measures in usability and user-experience evaluation. In the HCI literature, EEG has been applied to workload, attention, vigilance, fatigue, error recognition, emotions, engagement, flow, and immersion, with workload, attention, and emotions identified as especially promising targets for evaluation (Frey et al., 2013). In a distinct but related usage, the eegFloss ecosystem uses eegUsability as the name of a LightGBM model that screens sleep-EEG recordings for usable versus artifact-contaminated segments before downstream analysis (Sikder et al., 8 Jul 2025). Taken together, these usages suggest a common concern: whether EEG can support rigorous judgments about either the usability of an interface or the usability of the signal itself.
1. Conceptual scope and evaluation logic
EEG-based usability evaluation has been positioned as an “exocentric” complement to “egocentric” methods such as questionnaires and think-aloud protocols. The central claim is not that EEG replaces behavioral or subjective data, but that it adds millisecond-level temporal resolution, continuous monitoring, and sensitivity to states that are only partially visible in post-hoc reports (Frey et al., 2013). This logic is explicit in frameworks that combine EEG with standard HCI measures rather than treating it as a standalone ground truth.
A canonical operationalization appears in a framework that estimates three constructs continuously: mental workload, attention, and recognition of interaction errors. In that framework, short calibration tasks are used to obtain per-user labels, shrinkage LDA models are trained, and the resulting outputs are expressed as a continuous workload index , an attention index , and an error-recognition rate . Calibration performance was reported as mean AUROC for mental workload, $0.86$ for attention, and $0.82$ for error recognition, after which the indices were used to compare a keyboard and a touch-based interface in a controlled 3D-maze task (Frey et al., 2016).
This architecture establishes a general evaluation pattern that recurs across later studies. A task is instrumented with EEG, a construct of interest is represented through spectral or ERP markers, and the resulting measure is interpreted jointly with NASA-TLX, SUS, UEQ-S, reaction time, accuracy, or error rate. A plausible implication is that eegUsability is less a single metric than a family of construct-specific inference pipelines.
2. Instrumentation and signal-processing infrastructure
The methodological substrate of eegUsability is heterogeneous. Studies have used a single dry electrode at with an ear-clip reference and proprietary “eSense” outputs (Bose et al., 2016), 12-channel dry EEG with reference and ground (Gaspar-Figueiredo et al., 2023), 14-channel Emotiv EPOC or EPOC X montages (Bilalpur et al., 2018, Fallahi et al., 2024, Coutray et al., 9 Sep 2025), 28-channel ERP-oriented laboratory acquisition (Frey et al., 2015), and 32-channel research-grade caps in interactive tasks and driving simulators (Frey et al., 2016, Liu et al., 2024). The literature therefore spans consumer, wearable, and laboratory systems rather than a single hardware standard.
Despite this diversity, preprocessing conventions are comparatively stable. Representative pipelines include band-pass filtering in ranges such as Hz, 0 Hz, or 1 Hz; 50 Hz notch or line-noise suppression; re-referencing to common average or electrode average; ICA-based removal of ocular and muscle artifacts; baseline correction; bad-channel interpolation; and epoching around task events (Liu et al., 2024, Sengupta et al., 2017, Frey et al., 2015, Chiossi et al., 2024). A particularly rigorous driving-simulator pipeline combined downsampling, 2 Hz band-pass filtering, ZapLine, bad-channel interpolation using a correlation threshold of 3 for more than 4 of the recording, common-average reference, AMICA for 10 iterations, ICLabel-based component rejection, a final second-order Butterworth 5 Hz filter, and epoching from 6 s to 7 s around SAE Level 2 activation (Liu et al., 2024).
Feature extraction likewise varies with the construct being measured. Spectral studies use FFT, Welch PSD, or STFT; ERP studies use time-locked windows and sometimes spatial filtering into “virtual channels”; workload studies frequently add CSP or SSCSP to reduce dimensionality before LDA; and adaptive-interface studies use frontal asymmetry, theta/alpha ratios, or P300-related measures (Wobrock et al., 2015, Frey et al., 2016, Gaspar-Figueiredo et al., 2023). This diversity indicates that eegUsability is methodologically modular: the construct dictates the feature space.
3. Canonical EEG markers and analytic formulations
Mental workload is the most recurrent target. A standard review summarizes the dominant spectral pattern as frontal theta increase with rising working-memory demand and parietal alpha decrease, with a common index
8
Variants such as 9 and normalized subject-specific scores are also described (Frey et al., 2013). Adaptive-menu studies used
0
and additionally defined attraction as frontal alpha asymmetry,
1
plus a memorisation measure derived from the subsequent-memory P300 effect (Gaspar-Figueiredo et al., 2023).
Driving-simulator work adopted relative, rather than absolute, band powers. Total power was defined as 2, with
3
where 4 and 5. Parietal sites 6 were used for 7, frontal sites 8 for 9, and four 20 s epochs per HMI were concatenated into one 80 s block before PSD estimation (Liu et al., 2024).
ERP-based eegUsability targets different constructs. Attention is commonly indexed through P300 amplitude in oddball paradigms, while interaction-error recognition uses an ErrP signature with a 0 ms negativity (Frey et al., 2016). Sonification work reported N100 and P200 at 1, with a larger N100 in compared recall, and also observed N400/P600 in a sonification context (Bilalpur et al., 2018). Visual-discomfort work in HMDs focused on occipito-parietal P1, N2, and P3, treating them as indicators of early visual processing, conflict monitoring, and resource allocation (Chiossi et al., 2024). The resulting picture is not of a single EEG usability marker, but of a layered marker set aligned to task demands.
4. Empirical domains and comparative findings
Driving-simulator evaluation shows both the promise and the limits of spectral eegUsability. In a static simulator with three HMI designs—Fog, Trans, and Trans-fog—subjective measures differentiated the interfaces, with repeated-measures ANOVA reporting NASA-TLX 2 and Subjective Transparency 3. The Trans HMI produced the lowest perceived workload, 4, and the highest transparency, 5. By contrast, EEG spectral measures were not significant: 6, 7; 8, 9. The expected pattern—higher $0.86$0 and lower $0.86$1 for the lower-workload interface—was present, but did not reach significance (Liu et al., 2024).
In sonification, EEG aligned more closely with explicit usability judgments. Low cognitive load-inducing acoustic parameters corresponded to higher mapping accuracies, self-reported low-load blocks correlated with accuracy in immediate recall $0.86$2 and compared recall $0.86$3, and low-load parameters exhibited higher $0.86$4-band power. A CNN operating on raw 14-channel epochs reached a peak $0.86$5-score of $0.86$6, supporting the claim that reliable workload estimation is achievable with wearable EEG (Bilalpur et al., 2018).
Gaze-based typing provides a contrasting case in which conventional performance metrics and EEG diverged. The study reports that words per minute and keystrokes per character did not distinguish different keyboard designs, whereas STFT-based EEG analysis revealed workload differences across designs and across typing phases. A theta/alpha workload index was higher during correction than selection, with correction inducing approximately $0.86$7 higher theta-to-alpha ratio $0.86$8, and suggestion-bar placement at the top yielded significantly lower workload than placement at the bottom despite nearly identical text-entry rates (Sengupta et al., 2017).
Adaptive graphical menus further strengthened the cross-validation argument. Across 20 menu designs, all four EEG metrics—cognitive load, engagement, attraction, and memorisation—and completion time differed significantly $0.86$9. Large effect sizes were reported for cognitive load $0.82$0, memorisation $0.82$1, and completion time $0.82$2. Correlations with subjective UX were also high: cognitive load versus NASA-TLX perceived cognitive load, $0.82$3; attraction versus UEQ-S hedonic quality, $0.82$4 (Gaspar-Figueiredo et al., 2023).
Other domains extend eegUsability beyond workload narrowly defined. In next-generation authentication, three brainwave-based mechanisms—Slideshow, Face, and Reading—obtained SUS scores of $0.82$5, $0.82$6, and $0.82$7, with an overall brainwave-authentication score of $0.82$8. Participants nevertheless characterized brainwave authentication as more secure yet more privacy-invasive and effort-intensive than eye-movement authentication (Fallahi et al., 2024). In HMD visual discomfort, blur induced significantly more negative N2 amplitudes, with $0.82$9 and 0, while aggregate neutral-versus-blurred contrasts were significant for P1, N2, and P3 (Chiossi et al., 2024). These studies broaden eegUsability from efficiency-oriented evaluation to trust, privacy, comfort, and fatigue.
5. Adaptive systems, passive BCIs, and hands-free interaction
Some eegUsability systems are explicitly adaptive or control-oriented rather than purely evaluative. Attention-sensitive web browsing used a NeuroSky MindWave sensor with proprietary eSense outputs, empirical attention threshold 1, blink threshold 2, and temporal thresholds of at least 1 s to stabilize control. The reported pipeline incurred approximately 3 ms latency plus a 4 s update interval, and the study proposed attention-driven browser control, adaptive content emphasis, and an EEG-oriented web API design (Bose et al., 2016). Here usability was tied to responsiveness, calibration, and false-positive control rather than to workload alone.
MindDesktop pursued a more explicit BCI interaction design. Using an Emotiv EPOC, the system mapped exactly three signals—Scroll, Zoom-In, and Zoom-Out—onto a hierarchical pointing device and virtual keyboard. Over three sessions, mean completion times decreased across all five tasks; for example, the point-click task declined from 5 s to 6 s, and the send-email task from 7 s to 8 s. Repeated-measures ANOVA showed a main effect of Session on task time, 9, and SUS scores improved from 0 to 1 overall (Ossmy et al., 2017). This suggests that eegUsability may include learnability and command-set design, not only state inference.
Passive adaptation has also been demonstrated in display comfort. A proof-of-concept stereoscopic-display system classified comfortable versus uncomfortable viewing conditions from 1 s of post-stimulus EEG using shrinkage LDA. Mean single-trial accuracy was 2 across 12 users, and Monte Carlo clustering over seven consecutive variations increased performance to 3. The intended use was a passive BCI loop that would reduce binocular disparity if discomfort was detected (Frey et al., 2015).
Hybrid multimodal control in VR pushes this trajectory further. NeuroGaze combined eye tracking with an EEG “pull” command in a 360° cube-selection task. The system produced fewer errors than the alternative methods but much longer completion times: task completion means were 4 s for controllers, 5 s for eye gaze plus hand gesture, and 6 s for NeuroGaze; error rates per block were 7, 8, and 9, respectively. Completion time differences were large, 0, while physical demand in NASA-TLX was lower for NeuroGaze than for controllers 1 (Coutray et al., 9 Sep 2025). The empirical pattern is a classic speed–accuracy trade-off rather than a monotonic gain in usability.
6. eegUsability as a sleep-EEG artifact-screening model
Within eegFloss, eegUsability is not an HCI evaluation paradigm but a supervised artifact detector for sleep EEG. Its role is to flag ten-second segments as either “Good Data” or one of four artifact classes: “No Data,” “High Noise,” “Spiky Noise,” or “M-shaped Noise.” The model operates on Zmax recordings with two forehead EEG channels, a PPG channel, and tri-axial accelerometer traces sampled at 2 Hz, and uses nonoverlapping 10 s windows so that an entire cycle of M-shaped Noise can be captured (Sikder et al., 8 Jul 2025).
Feature extraction combines two representations. Short-time Fourier transforms produce EEG and motion spectrograms over 3 time windows by 4 frequency bins, yielding 5 flattened features. TSFEL contributes 6 descriptors per signal, for 7 additional features. Each sample is therefore represented by a 8-dimensional feature vector. The classifier is a LightGBM ensemble trained by minimizing categorical cross-entropy over 9 classes with learning rate 0, 1, 2, and 3 (Sikder et al., 8 Jul 2025).
The evaluation metrics are strong at the aggregate level. After subject-wise splitting and undersampling to obtain approximately 4 usable versus 5 artifact data, eegUsability v1.0 achieved weighted Precision 6, Recall 7, 8, 9, and 00. Class-wise recall was 01 for Good Data, 02 for No Data, 03 for High Noise, 04 for Spiky Noise, and 05 for M-shaped Noise, with M-shaped Noise often mistaken for Good Data in 06 of its occurrences (Sikder et al., 8 Jul 2025).
This usage of eegUsability shifts the emphasis from the usability of an interface to the usability of the EEG stream. The model has also been applied to standard PSG systems and the Bitbrain Ikon headband without retuning, provided usable EEG amplitude remains near 07 and, if available, tri-axial accelerometer traces resemble Zmax burst profiles (Sikder et al., 8 Jul 2025). A plausible implication is that data-quality assessment and user-state assessment are converging components of a broader EEG usability stack.
7. Limits, controversies, and future directions
The literature repeatedly warns against overinterpreting EEG as a direct or sufficient usability measure. Small samples remain common, as in the driving study with 08 usable EEG datasets, where effect sizes for EEG outcomes were below 09 and epoch concatenation reduced the number of independent observations (Liu et al., 2024). Per-user calibration is described as mandatory in some frameworks, inter-subject variability creates dropout risk, and high event rates can cause ERP overlap (Frey et al., 2016). In 3D object manipulation, context shifts between calibration and use required covariance adaptation, and multimodal fusion with ECG and GSR actually degraded short-window classification relative to EEG alone (Wobrock et al., 2015).
Consumer systems introduce additional constraints. Proprietary attention metrics limit interpretability in browser applications (Bose et al., 2016). Wearability, headset comfort, and verification time were salient concerns in brainwave authentication despite favorable SUS scores (Fallahi et al., 2024). EEG+gaze VR control reduced errors yet remained markedly slower than conventional input (Coutray et al., 9 Sep 2025). These findings counter the common misconception that more direct neurophysiological access automatically yields more usable systems.
At the same time, the prospective directions are technically coherent. The driving literature recommends preserving multiple independent epochs, exploring event-related potentials, connectivity, time–frequency dynamics, and cross-frequency coupling, and combining EEG with pupilometry or heart-rate variability for multimodal workload indices (Liu et al., 2024). HMD discomfort work proposes posterior ERP-based detectors operating within 10 ms, and adaptive-menu work suggests switching to low-load interface policies when cognitive load exceeds a threshold (Chiossi et al., 2024, Gaspar-Figueiredo et al., 2023). In sleep research, artifact screening with eegUsability provides a preprocessing layer that can improve the reliability of automatic staging (Sikder et al., 8 Jul 2025).
Taken together, the research suggests that eegUsability is best understood as a technical program rather than a single method: rigorous EEG preprocessing, construct-specific feature design, within-subject validation, and multimodal interpretation are used to turn neural signals into evidence about interface demand, comfort, transparency, attention, or data quality. Where those components converge with behavioral and subjective measures, EEG becomes a high-resolution adjunct to usability science; where they do not, the divergence itself becomes informative.
