EEG-based cognitive load estimation during the use of a virtual wheelchair simulator

Driving a powered wheelchair is a complex task that requires the integration of motor, visual, and cognitive skills. The development of assistive technologies without appropriate assessment methods that help bridge the gap between users and developers may lead to abandonment and reduced engagement. Most assessments rely on explicit measures, such as performance metrics, or subjective tools like interviews and questionnaires. In contrast, implicit measures allow continuous inference of mental states during task execution. This study proposes the use of blink indices derived from electroencephalographic (EEG) signals as implicit metrics to estimate cognitive load during the use of a virtual reality wheelchair training simulator. A total of 25 participants (14 females and 11 males; mean age 26.50 ± 5.7 years) completed a predefined route using a virtual wheelchair simulator. Blink parameters, including frequency, duration, and velocity, were extracted from EEG signals during task performance. After completing the simulation, participants responded to the NASA Task Load Index (NASA-TLX) to assess subjective cognitive load, as well as the System Usability Scale (SUS) and the Igroup Presence Questionnaire (IPQ). The findings showed that higher mental-visual demand was associated with decreases in blink frequency, duration, and velocity. Correlation analyses between NASA-TLX scores and blink parameters revealed weak to moderate associations. These results suggest partial convergence between subjective and physiological measures of cognitive load. Blink-based indices derived from EEG signals provide relevant information regarding cognitive demand during wheelchair simulator use. However, blink parameters alone are insufficient to reliably infer cognitive load. When combined with subjective questionnaires, implicit physiological metrics may offer a more comprehensive assessment than questionnaires alone, supporting the development and refinement of assistive training technologies.

The World Health Organization (WHO) defines disability as any restriction or lack of ability to perform an activity in a manner or within a range considered normal for a human being (World Health-Organization, 2015; World Health-Organization, 2020). According to WHO estimates, more than one billion people worldwide live with some form of disability, representing approximately 15% of the global population. In this context, there is a growing demand for assistive technologies (AT) aimed at improving autonomy, accessibility, and quality of life for people with disabilities. AT encompasses devices and systems designed to maintain or enhance functional capabilities and promote independence and wellbeing (World Health-Organization, 2020).

Despite their potential benefits, the effective use of assistive technologies often presents challenges, particularly during the learning and adaptation phases. Training with real devices in real-world environments may involve safety risks, especially for novice users. As a result, virtual training tools have been increasingly explored as safer alternatives for skill acquisition. In parallel, user-centred design (UCD) and human–computer interaction (HCI) approaches have been adopted to ensure that AT systems align with users’ capabilities, needs, and limitations (Kübler et al., 2014; Nayak et al., 2016; Tirtha, 2017; Dorrington et al., 2016).

An important factor influencing the usability and effectiveness of AT training systems is the cognitive load imposed on users during interaction. Excessive cognitive load can degrade performance, increase error rates, and negatively affect user experience (Leppink, 2017). Therefore, assessing cognitive load has become an essential component of AT evaluation and UCD processes, in line with Cognitive Load Theory, which emphasizes minimizing unnecessary cognitive demands while supporting task-relevant learning (Leppink, 2017).

Cognitive load is commonly assessed using subjective instruments, such as questionnaires, with NASA-TLX being one of the most widely used tools (Hart and Staveland, 1988). However, subjective measures alone may not fully capture cognitive demands, as they rely on post-task self-report and are influenced by individual perception and bias (Naismith et al., 2015). Consequently, several objective and physiological approaches have been proposed for mental workload assessment, including EEG spectral analysis, pupillometry, heart-rate variability, galvanic skin response, and electrooculography (Hart and Staveland, 1988; Naismith et al., 2015; Chen and Epps, 2014; Grootjen et al., 2024; Shin et al., 2012; Plass et al., 2010).

Several studies have demonstrated a robust relationship between blink behavior and cognitive load, particularly in tasks requiring sustained visual attention and mental processing (Chen and Epps, 2014; Molina-Cantero et al., 2019; Martins and Carvalho, 2015; Holland and Tarlow, 1972). Early experimental investigations showed that blink frequency decreases as task difficulty and attentional demand increase, suggesting that blinking is suppressed to maintain visual information intake during cognitively demanding activities (Holland and Tarlow, 1972). Subsequent studies extended these findings by analyzing not only blink rate, but also blink duration and velocity, reporting shorter and slower blinks during tasks associated with higher mental workload (Schleicher et al., 2008; Johns, 2003). More recent research has explored blink-related parameters in applied contexts such as driving, human–computer interaction, and simulated environments (Chen et al., 2023; Chen and Epps, 2014; Biondi et al., 2023; Ayala et al., 2024). These studies typically compared baseline or low-demand conditions with tasks involving increased visual–cognitive complexity, demonstrating consistent reductions in blink frequency and duration as cognitive load increased (Chen and Epps, 2014; Schleicher et al., 2008; Martins and Carvalho, 2015). Some works also reported moderate correlations between blink metrics and subjective workload measures, such as NASA-TLX scores, reinforcing the interpretation of blink behavior as an implicit indicator of mental workload (Chen and Epps, 2014; Biondi et al., 2023). Collectively, these findings provide empirical support for the use of blink-derived features as a low-intrusion and continuous measure of cognitive load during task execution.

Nevertheless, many physiological assessment techniques require expensive equipment, extensive calibration procedures, or impose constraints that reduce ecological validity, particularly in rehabilitation and training scenarios. Blink-related parameters represent a minimally intrusive alternative that can be captured implicitly during task execution and integrated into existing sensing setups. When derived from EEG signals, blink features can be extracted with high temporal resolution while remaining compatible with broader cognitive workload analysis frameworks.

In this context, the objective of this study is to analyze both explicit and implicit indicators of cognitive load during the use of a virtual wheelchair training system. Explicit measures were obtained using standardized questionnaires assessing usability, presence, and subjective workload, while implicit measures were derived from blink-related parameters extracted from EEG signals. The study investigates whether blink-derived metrics can complement subjective workload assessment and contribute to a more comprehensive understanding of user cognitive demands during assistive technology training.

The following modules were implemented for the development of this work: A) Virtual powered wheelchair simulator; B) Blink data collection module; C) System to synchronize the data from the simulator with the EEG module; D) Data processing; E) Experimental protocol; F) Statistical analysis.

Virtual powered wheelchair simulator

We used the virtual simulator of a powered wheelchair developed by Martins et al. (2021) Figure 1. The simulator is a safe environment for inexperienced power wheelchair users to learn power wheelchair driving skills. The wheelchair simulator was controlled with joystick of the wheelchair model VR2 with haptic feedback, which was adapted following the instructions in Martins et al. (2021). The joystick sends 5 commands about the driving conditions through the serial port and controls of the virtual simulator: (i) move forward, (ii) move backward, (iii) rotate to the right, (iv) rotate to the left and (v) stop. Figure 1 also illustrates the visual presentation of the virtual environment used during the task. Panel (a) shows the 3D representation of the task execution path, while panel (b) presents the driving protocol, in which the numbered waypoints (1–15) indicate the sequence to be followed, starting at point 1 and ending at point 15.

The virtual environment was displayed on a desktop monitor (24 inches, 1,920 × 1,080 pixels resolution, 60 Hz refresh rate). Participants viewed the scene while seated in the simulator.

Blink data collection module

In this study, the term “blink” refers to eye-movement–related events inferred from ocular artifacts present in frontal EEG signals, rather than to full eyelid closures measured using video-based or electrooculography (EOG) methods. These EEG-derived blink-related events capture rapid eye-movement transients associated with blinking activity and allow the extraction of temporal features such as duration and amplitude–velocity relationships based on signal landmarks. Accordingly, all blink metrics reported in this work correspond to EEG-based temporal measurements of eye-movement dynamics and not to direct measurements of physical eyelid closure. For clarity, the term “blink” is used throughout the manuscript to denote these EEG-derived events.

Although blink detection can be performed using vision-based methods or simple webcam setups when only blink onset or frequency is of interest, the present study required the extraction of detailed blink morphology, including duration, amplitude-related features, and amplitude–velocity dynamics. These parameters demand high temporal resolution to accurately identify blink landmarks and rapid transient changes. Frontal EEG signals provide millisecond-level temporal resolution and exhibit high-amplitude ocular artifacts, enabling reliable extraction of fine-grained blink features that are difficult to estimate using standard video frame rates. In addition, EEG-based acquisition allows ocular-related information to be collected within the same sensing modality commonly used in cognitive workload research, facilitating methodological consistency and future integration with EEG-based workload measures (Schleicher et al., 2008; Brooke, 1996; Project Consortium, 2015; Delorme and Makeig, 2004; Kleifges et al., 2017; Kamaraj et al., 2016; Mahajan et al., 2013; Johns, 2003).

Blink parameters were derived from electroencephalography (EEG) signals acquired using the MindWave Mobile 2 headset (NeuroSky Inc.), a single-channel dry-electrode EEG device with a bandwidth of 3–100 Hz, 12-bit resolution, and an internal sampling rate of 512 Hz, with wireless data transmission via Bluetooth. EEG signals were recorded using one active frontal electrode positioned at an Fp1-equivalent location according to the international 10–20 system. Two auxiliary electrodes, serving as reference and ground, were placed on the earclip and temporal region, respectively, and were used exclusively for signal stabilization and noise rejection; they were not treated as additional EEG channels.

A single-channel frontal configuration was adopted to minimize intrusiveness, reduce setup time, and preserve ecological validity, which are critical requirements in rehabilitation and assistive-technology training scenarios. Because blink detection relies primarily on high-amplitude ocular artifacts that are strongest at frontal sites, particularly near Fp1, this configuration is sufficient for reliable blink extraction without the need for multichannel EEG or dedicated EOG setups, as demonstrated in previous blink-based workload studies (Schleicher et al., 2008; Brooke, 1996; Project Consortium, 2015; Delorme and Makeig, 2004; Kleifges et al., 2017; Kamaraj et al., 2016; Mahajan et al., 2013; Johns, 2003).

EEG signals were visualized and stored using the OpenViBE platform. Blink extraction was performed through a multi-stage processing pipeline that included filtering to remove line noise and slow drifts, automated detection of blink candidates based on amplitude and zero-crossing criteria, exclusion of events with low signal-to-noise ratio (SNR), and removal of outliers that did not satisfy physiological and morphological constraints. Kinematic variables from the simulator were not analyzed, as the objective of this study was to investigate blink-based estimation of cognitive load rather than motor-performance characterization.

To ensure reliable blink feature extraction, an SNR criterion was applied to all detected blink candidates. For each event, the signal component was defined as the peak-to-baseline amplitude of the blink waveform, measured relative to the local baseline estimated from pre- and post-blink segments (leftBase and rightBase). The noise component was estimated as the standard deviation of EEG signal segments adjacent to the blink event, excluding the blink interval itself. Blink candidates with SNR values below a predefined quality-control threshold were excluded from further analysis. This criterion was applied consistently across all participants and experimental conditions to ensure reliable identification of blink landmarks while excluding events dominated by background EEG noise or poorly defined morphology.

All detected blink candidates were initially classified as “possible blinks.” From these, “good blinks” were defined as events that satisfied all quality-control requirements, including sufficient signal-to-noise ratio, well-defined blink landmarks, and physiologically plausible values for duration and amplitude–velocity parameters. Only blinks classified as “good blinks” were retained for subsequent feature extraction and statistical analysis. Events classified as “possible blinks” satisfied amplitude criteria but exhibited incomplete landmark detection, which is expected given the distance between frontal EEG electrodes and the ocular sources of blink-related activity.

Blink duration was defined based on EEG signal landmarks rather than full eyelid closure time. Specifically, duration was computed as the interval between the leftZero and rightZero points of the blink waveform, corresponding to the zero-crossings surrounding the blink peak and capturing the high-amplitude transient component of the ocular artifact present in frontal EEG channels. This EEG-derived blink duration differs from blink durations measured using video-based or EOG approaches, which typically include the full eyelid closure and reopening phases and therefore yield longer duration values. Consequently, the absolute duration values reported in this study should not be directly compared with blink durations obtained using video or EOG methods.

Synchronisation system

The EEG and simulator data were synchronized using Lab Streaming Layer (LSL), which is open source software for data acquisition and real-time synchronization and access. During the virtual simulator experiment, we collected the participant’s EEG signals while controlling the virtual wheelchair using the joystick. The performance metrics captured by the simulator and the joystick events were simultaneously recorded with the EEG data by the LSL in the same file. The saved data was subsequently processed and analyzed using MATLAB.

Data processing

The data are of implicit objective and explicit subjective type. The implicit objective metrics are the EEG signal, used for detection of blink features, and performance measures (task time, number of commands and collisions). The explicit subjective data come from the questionnaires answered by the participants after the experiment with the virtual simulator. The NASA Task Load Index (NASA-TLX) (Hart and Staveland, 1988) questionnaire was used to infer the cognitive load of the task (mental effort). In addition, the participants answered questions related to usability, these questions were based on the System Usability Scale (SUS) (Brooke, 1996) and Igroup Presence Questionnaire (IPQ) (Project Consortium, 2015).

Processing the EEG signals for blink detection

The EEG signal was processed offline using MATLAB (MathWorks, Natick, MA, United States) in combination with EEGLAB (Delorme and Makeig, 2004) and the Blinker toolbox (Kleifges et al., 2017). Raw EEG data were first visually inspected to identify segments with excessive noise or signal loss. As a preprocessing step, the EEG signal was filtered to remove power-line interference and slow drifts, including application of a 60 Hz notch filter to suppress electrical noise from the power network and high-pass filtering to ensure signal stability. Baseline correction was applied by subtracting the mean signal value computed over non-blink segments, allowing consistent amplitude comparison across time.

Automated blink detection was performed using the Blinker library, which identifies candidate blink events in EEG signals based on amplitude thresholding, zero-crossing detection, and polynomial fitting of the blink waveform. Prior to blink detection, preprocessing steps included filtering, artifact attenuation, and segmentation aligned to blink-related landmarks. For each detected blink, reference points were extracted, including leftBase, leftZero, peak amplitude, rightZero, and rightBase, defining the onset, peak, and offset of each blink event. Figure 2 illustrates an example of an EEG signal segment corresponding to a blink, highlighting these landmarks as well as the ascending (upstroke) and descending (downstroke) phases of the waveform.

To improve robustness, blink candidates with low signal-to-noise ratio (SNR) were excluded from further analysis. For each blink event, the signal component was defined as the peak-to-baseline amplitude of the blink waveform, while the noise component was estimated from EEG segments adjacent to the blink event, excluding the blink interval itself. Blink candidates whose SNR values fell below a predefined quality-control threshold were rejected. Additional artifact rejection was performed by removing outliers whose amplitude, duration, or slope values fell outside physiologically plausible ranges, following the criteria recommended in the Blinker documentation.

All detected blink candidates were initially classified as “possible blinks.” From these, “good blinks” were defined as events that satisfied all quality-control requirements, including sufficient signal-to-noise ratio, well-defined blink landmarks, and physiologically plausible feature values. Only blinks classified as “good blinks” were retained for subsequent feature extraction and statistical analysis.

From the retained blink events, the following features were computed: blink frequency (blinks per minute), blink duration (defined as the time interval between the leftZero and rightZero landmarks), and amplitude–velocity ratio, derived from the slopes of the ascending and descending phases of the blink waveform. These features were computed separately for baseline and task periods, enabling within-subject comparisons. This processing pipeline provides a reproducible and robust framework for extracting blink-related parameters suitable for cognitive load analysis.

Processing of performance metrics

As observed in related works (Kamaraj et al., 2016; Mahajan et al., 2013; Devigne et al., 2017; Moya et al., 2017), the main parameters for performance evaluation in driving a wheelchair in the literature are: time spent to complete the task, number of collisions and number of commands with the joystick. Thus, these three parameters were calculated during the route proposed in the experimental protocol. Each experimental session lasted approximately 35–45 min, following the recommendations of the ITU-T P.913 standard, which states that sessions should not exceed 60 min. This duration corresponds to the full execution of the experimental phases and ensured participant comfort and consistency across sessions.

Processing the subjective metrics of cognitive effort (NASA-TLX)

To analyse the amount of cognitive load required to control the virtual simulator the NASA “Total Load Task” (NASA-TLX) questionnaire was applied, which subjectively measures this load (Hart and Staveland, 1988). The questionnaire has a multidimensional rating that derives an overall workload score based on an average of six factors (sub-scales), as depicted in Table 1.

According to the literature, NASA-TLX is applied in several human-machine interface systems. Therefore, it can display correlations with several variables, such as performance in simulators (Hart and Staveland, 1988). For example, high cognitive load can have negative effect on task completion.

Processing subjective usability and presence metrics

In addition to the NASA-TLX questionnaire, participants answered questions about presence and the usability of the system - Table 2. These questions were based on the IGroup Presence Questionnaire (IPQ) and the System Usability Scale (SUS) (Brooke, 1996; Project Consortium, 2015; Borges et al., 2016). We asked participants to rate each question using the absolute category rating (ACR) system described in the ITU-TP.913 standard (ITU, 2022). The rating system used a five-point Likert scale (Joshi et al., 2015) to determine whether the user agreed or disagreed with the statements.

The usability questions aimed to collect participants’ opinions about the virtual simulator. For instance, whether the simulator was easy to use, whether they believe that other people can learn to use a wheelchair using the simulator, about the consistency of the simulator, and whether the simulator can be used without any prior knowledge. The presence questions assessed four aspects, these being: I. General presence “the feeling of being in the virtual environment (VE)”, II. Spatial presence “the feeling of being physically present in the VE”, III. Engagement “measuring the attention devoted to the experience in the VE” and IV. Realism “measuring the subjective experience of realism in the VE”.

Experimental protocol

The study was approved by the Athlone Institute of Technology Ethics Committee and conducted with healthy participants. A total of 25 participants (14 females and 11 males), with a mean age of 26.5 ± 5.7 years (mean ± SD), were included in the analysis. Each experimental session lasted approximately 35–45 min and followed the ITU-T P.913 recommendation for subjective evaluation experiments (ITU, 2022).

The experimental protocol consisted of the following stages. First, participants received a detailed explanation of the study objectives and procedures and provided written informed consent. Next, a screening phase was conducted to ensure eligibility. Participants were asked about epilepsy, sleep deprivation (less than 6 hours of sleep in the previous 24 h), suspected pregnancy, and alcohol consumption in the previous 24 h. Visual acuity was assessed using the Snellen test (Sue, 2007), and color perception was evaluated using the Ishihara test (Hoffmann and Menozzi, 1999). Only participants with normal or corrected-to-normal visual acuity (20/20–20/40) and normal color perception (no more than one incorrect Ishihara plate) were included.

Following screening, a 5-min baseline EEG recording was collected while participants remained seated quietly in the test room without performing any task. After the baseline period, participants underwent a brief training session in which the virtual wheelchair simulator was presented, and joystick control commands were explained. Under supervision, participants practiced basic navigation maneuvers to familiarize themselves with the system.

In the test phase, participants were required to complete the predefined route in the virtual wheelchair simulator (Figure 1). During task execution, EEG signals and performance metrics were continuously recorded for subsequent analysis. Upon completion of the task, participants completed the NASA-TLX questionnaire to assess subjective cognitive workload, as well as questionnaires related to usability (SUS) and presence (IPQ).

To reduce inter-individual variability, blink parameters extracted during task execution were baseline-corrected by comparison with each participant’s own baseline values, allowing each participant to serve as their own reference.

Statistical analysis

Firstly, the Shapiro-Wilk normality test was performed (Surhone et al., 2010), which indicated that the data concerning the blink signal came from a normal distribution. However, the performance values and the questionnaires were not considered normal. For this reason, a parametric analysis approach (t-test for comparison of repeated measures) was performed for the data concerning blinking and a non-parametric analysis for the others (Spearman correlation test) (Spearman, 2010). All analyses were performed using the IBM SPSS and Matlab platforms.

All participants completed the proposed task. Results will be presented for: A) Performance metrics; B) Usability and presence metrics; C) Subjective Metrics of Cognitive Effort (NASA-TLX); D) Blink metrics; E) Correlation between NASA-TLX and Blink Index; F) Correlation between NASA-TLX and Performance Metrics; G) Correlation between NASA-TLX and IPQ questionnaire.

Performance metrics

The mean values of the task executed by the participants were: number of commands (81.32), number of collisions (6.64) and time spent to complete the task (192.76 s). The median of the number of commands was 75.5, the number of collisions was 6 and the time to complete the task was 179 s. The mode for the numbers of commands, errors and time were: 66, 6 and 166 s, respectively.

Usability and presence metrics

Table 3 shows the mean values of the questionnaire results on usability and presence. A Likert scale of 5 (strongly disagree to strongly agree) was applied (Joshi et al., 2015). The maximum value it can reach is 5 for each question and the minimum is 1. Scoring in the middle of the scale, referring to 3, represents a neutral result. Usability and presence metrics are reported as mean ± standard deviation across all participants.

Subjective metrics of cognitive effort (NASA-TLX)

The workload was assessed using NASA-TLX. The results show that both raw and adjusted final values have averages of 32.6 ± 2.81 and 40.56 ± 3.71 respectively. Adjusted NASA-TLX values correspond to weighted scores based on participant-specific pairwise weighting. The ±values represent the standard deviation across participants. Mental demand had the highest adjusted and raw value, this shows that the participants considered that the task demands more mental load than the other factors.

In general, we may state that the participants’ perception of the factors assessed by NASA-TLX on a decreasing scale were: mental workload, effort, performance, temporal, physical and frustration. Twelve participants had a raw cognitive workload in the range between 20–40 and for the adjusted values the highest incidence was in the range between 40–60.

Table 4 summarizes the subjective cognitive workload results obtained using the NASA-TLX questionnaire. Both raw and adjusted scores indicate moderate perceived workload during task execution. Mental demand was identified as the dominant contributor to workload, followed by effort and performance, reflecting the visual–cognitive requirements of the simulated wheelchair task.

Blink metrics

The mean blink frequency was 18.37 ± 1.42 blinks per minute during the baseline period and 13.13 ± 1.16 blinks per minute during the test period. These values fall within the range of spontaneous blink frequencies reported in the literature for alert conditions (Schleicher et al., 2008; Bentivoglio et al., 1997; Caffier et al., 2003; Cruz et al., 2011). When comparing the distributions of blink frequency between baseline and test periods, the median blink frequency during the baseline was higher than that observed during the test. This reduction in blink frequency during task execution is consistent with previous findings indicating that blink frequency tends to decrease during activities involving increased mental–visual demand (Cruz et al., 2011).

Blink duration analysis showed that durations between 50 and 60 ms were the most frequent in the sample. Overall, blink durations during the test period were shorter than those observed during the baseline. In the literature, blink durations measured using video-based or electrooculography (EOG) approaches typically range between 100 and 300 ms during alert states, whereas longer durations, often exceeding 500 ms, are associated with drowsiness or fatigue due to prolonged eyelid closure (Schleicher et al., 2008; Johns, 2003). The shorter durations observed in this study are expected, as EEG-derived blink duration captures only the high-amplitude transient component of ocular artifacts rather than full eyelid closure.

Across both baseline and test periods, the observed blink frequency and duration values remained within ranges associated with alert states. The reduction in blink frequency and duration during the test period relative to baseline is in agreement with previous studies reporting suppression of blinking during tasks requiring sustained visual attention and cognitive processing (Schleicher et al., 2008; Brooke, 1996; Project Consortium, 2015; Delorme and Makeig, 2004; Kleifges et al., 2017; Kamaraj et al., 2016; Mahajan et al., 2013; Cruz et al., 2011; Johns, 2003).

Correlation between NASA-TLX and blink index

We applied the non-parametric Spearman linear correlation test to analyse the correlation coefficients between the blink rates and the NASA-TLX scores. The reason for this analysis is to assess whether the characteristics of the blinks can be used as an analysis of cognitive load during the use of the virtual wheelchair simulator. Spearman’s correlation coefficient was applied to examine the strength of the linear relationship between NASA- TLX and the blink parameters. The main results obtained in this analysis are shown in Table 4. Previous studies have already examined some blink parameters at two levels of cognitive load and correlations were found with the level of mental load (Martins and Carvalho, 2015; Holland and Tarlow, 1972). However, these correlations with mental load were not found with the results obtained in the present study. The correlations found were with the metrics of amplitude-velocity ratio and blink duration with temporal and effort demand.

Before presenting the correlation results, it is important to clarify the meaning of the blink parameters. PAVRZ and NAVRZ represent the positive and negative amplitude–velocity ratios of the blink waveform. DurationBase and DurationZero indicate blink duration measured respectively at the baseline and zero-crossing intervals. Higher absolute Spearman correlation coefficients (|r|) indicate stronger monotonic relationships between the variables, whereas “Sig” refers to the p-value, indicating whether the correlation is statistically significant (p < 0.05).

Table 5 displays the Spearman correlation coefficients between NASA-TLX dimensions and blink metrics. Negative coefficients indicate that higher blink values are associated with lower perceived workload, whereas positive coefficients indicate the opposite. Significant correlations are highlighted (p < 0.05). Values between |0.1–0.3| represent weak correlations, |0.3–0.5| moderate, and >|0.5| strong associations.

Correlation between NASA-TLX and performance metrics

The literature states that performance metrics can be used as a method of inferring cognitive load (Ayres, 2006; Hollender et al., 2010; Brünken and Plass, 2023). However, some authors infer cognitive load using subjective rating scales by applying questionnaires such as NASA-TLX. Moreover, performance metrics are reported along with this information to describe the relative efficiency of the task’s instructional conditions (Paas and Van Merriënboer, 1994; Paas and Van Merrienboer, 1993).

In this study we have reported the values regarding the cognitive load using NASA-TLX and the participants’ performance using three types of metrics (time to complete the task, number of commands and collisions (errors)). The result of the correlation analysis between these metrics is shown in Table 6, where two weak negative correlations were found. In the first row of Table 6 it is displayed that the number of collisions has weak negative correlation with the raw measure of physical demand (r = −0.444). In the second row of Table 6 it is shown the weak negative correlation between the time to complete the task with the adjusted measure of effort (r = −0.426).

Table 6 presents the Spearman correlations between NASA-TLX dimensions and performance metrics. These correlations help determine whether objective task performance is associated with subjective workload ratings. As displayed, only weak negative correlations were significant (p < 0.05), meaning that as the physical demand or effort rating increased, task time or number of collisions tended to decrease slightly.

Correlation between NASA-TLX and IPQ questionnaire

We found a correlation between two questions about presence in the virtual environment (question 9 and 10) with the variables of the NASA-TLX questionnaire - Table 7. Question 9 is a question about engagement (attention devoted to the experience in the virtual environment). Question 10 is a question about realism (how consistent the experience in the virtual environment seemed with the experience in the real world).

Table 7 presents the Spearman correlation coefficients between presence-related IPQ questions (questions 9 and 10) and selected NASA-TLX workload measures. As shown in the table, question 9, which assesses engagement and attention devoted to the virtual environment, exhibited weak negative correlations with perceived workload. Question 10, related to the perceived realism of the virtual environment, showed moderate to weak negative correlations with NASA-TLX measures. Negative correlation coefficients indicate that higher perceived presence in the virtual environment was associated with lower subjective workload. Only the correlations explicitly reported in Table 7 were considered in this analysis.

In summary, all statistically significant correlations were weak or moderate, indicating that blink-derived parameters, performance metrics, and presence scores relate to perceived workload in consistent but not strongly predictive ways. This reinforces the need for multimodal approaches when estimating cognitive load.

This study aimed to examine whether blink-derived metrics extracted from a single-channel frontal EEG recording could serve as indicators of mental load during a virtual wheelchair-driving task. Overall, the results show that blink activity is modulated by task demands; however, blinking alone does not constitute a sufficiently reliable standalone estimator of mental load. Below we discuss the findings in relation to existing literature.

The observed reduction in blink frequency during task execution should be interpreted with caution. Although blink suppression is directly associated with sustained visual attention, this phenomenon is not independent of cognitive load. Tasks that require continuous visual monitoring also impose increased cognitive processing demands related to information integration, decision-making, and motor planning. Previous studies have shown that blink frequency decreases under conditions of increased mental workload, particularly in visually demanding and goal-directed tasks (Chen and Epps, 2014; Schleicher et al., 2008; Martins and Carvalho, 2015; Holland and Tarlow, 1972).

In the context of the virtual wheelchair task, participants were required to simultaneously process visual information, plan navigation actions, and avoid obstacles. These combined requirements lead to an increase in mental workload that is intrinsically linked to visual attention demands. Therefore, the changes in blink behavior observed in this study are consistent with the literature and reflect the interaction between visual attention and cognitive load rather than visual demand alone.

The magnitude of the blink-derived parameters observed in this study should be interpreted in relative rather than absolute terms. The analysis focused on within-subject variations between baseline and task conditions using a consistent signal-derived definition of blink metrics. As such, the interpretation of the results is based on observed trends and correlations with subjective workload measures, rather than on direct comparison with absolute physiological reference ranges.

Regarding subjective workload, correlations between NASA-TLX dimensions and blink indices were weak to moderate, consistent with previous research reporting modest associations between ocular metrics and perceived effort. Importantly, no significant correlation was found between blink indices and the NASA-TLX overall workload score, reinforcing the view that blink metrics alone cannot robustly capture the multifaceted construct of mental load. Performance measures—such as number of collisions and task completion time—showed weak negative correlations with perceived physical demand and effort, indicating that as task performance improved, perceived exertion tended to decrease.

Immersion measures from the IPQ also revealed weak negative correlations with performance demand and overall cognitive load. Participants reported moderate levels of immersion, suggesting that the virtual environment was sufficiently engaging without causing sensory overload. However, the weak correlations indicate that the relationship between immersion and cognitive load is complex and requires further investigation in more controlled studies.

The absolute blink durations reported in this study are shorter than values typically reported in video-based or EOG-based literature. This difference arises from the definition of blink duration adopted here, which is based on EEG signal landmarks rather than full eyelid closure. EEG-derived blink duration reflects the temporal extent of the high-amplitude ocular artifact recorded at frontal electrodes and does not include slower eyelid movements with lower signal amplitude.

Importantly, the objective of this study was not to estimate physiological blink duration per se, but to investigate whether EEG-derived blink temporal features vary as a function of cognitive load. By focusing on relative changes between baseline and task conditions, the analysis emphasizes within-subject effects rather than absolute duration values. This perspective supports the interpretation that the observed differences are associated with task-related cognitive demands.

In this study, the correctness of the mental-load assessment was verified through comparison with the NASA-TLX, a subjective instrument widely recognized as a valid reference in cognitive workload research and frequently applied in human–machine interaction and simulator studies (Biondi et al., 2023; Devigne et al., 2017; Moya et al., 2017; Borges et al., 2016). No additional physiological methods (such as heart-rate variability, GSR, pupillometry or multi-channel EEG spectral analysis) were employed, as the purpose of the work was to investigate whether blink-derived features could complement a validated subjective measure while maintaining a low-intrusion setup compatible with rehabilitation scenarios. The correlation results obtained here—moderate or weak depending on the blink parameter—indicate that blink metrics alone are not sufficient to fully infer mental workload, but they may provide useful additional information when combined with standardized questionnaires such as NASA-TLX.

In summary, the findings indicate that blink-derived metrics are sensitive to cognitive-load variations but do not offer sufficient reliability to be used as a standalone method for mental-load estimation. Their usefulness increases substantially when combined with subjective measures such as NASA-TLX or when integrated into multimodal assessment frameworks. Future studies should consider combining blink analysis with additional physiological markers (e.g., heart-rate variability, EOG, EEG spectral features) to enhance the robustness of mental-load estimation.

Although the present study demonstrates that blink-based indices derived from EEG signals are sensitive to variations in cognitive load, it is important to acknowledge that our analysis pipeline was executed offline. The multi-step processing—including filtering, landmark extraction, and blink-morphology computation—requires batch data and therefore does not yet enable continuous monitoring during task execution. This limitation means that the system, in its current form, cannot provide real-time cognitive-load feedback to users, despite the conceptual motivation described in the abstract.

However, recent advances in real-time EEG processing suggest that online implementation is feasible. Future versions of the system could employ low-latency digital filters, incremental blink-landmark detection, and lightweight artifact-rejection algorithms compatible with streaming frameworks such as Lab Streaming Layer (LSL). In addition, several real-time blink-detection methods (including webcam-based and EEG-based algorithms) have been demonstrated in human–computer interaction contexts and could be integrated into an adaptive assistive-technology interface. Such a system could provide timely feedback to the user—e.g., adjusting task difficulty, modifying visual load, or alerting the clinician—thereby supporting a continuous and responsive training experience. Developing this real-time capability represents a key direction for future research.

This study examined the applicability of cognitive load theory by integrating blink-derived metrics extracted from single-channel EEG signals with subjective workload assessments using the NASA-TLX. The results of this study indicate that blink-related parameters exhibit consistent trends during the execution of a visually and cognitively demanding task in a virtual wheelchair simulator. However, the observed associations between blink metrics and subjective workload measures were weak to moderate, suggesting that blink behavior alone is not sufficient to robustly infer cognitive load.

Rather than providing independent or strongly predictive information, blink-derived features should be interpreted as complementary indicators that may contribute to a broader, multimodal assessment of cognitive workload when used alongside validated subjective instruments such as NASA-TLX. These findings highlight the potential of blink metrics in low-intrusion assessment scenarios, while also underscoring the need for further studies with real-time implementations and additional objective measures to better establish their practical utility.

Although the present study relied on descriptive statistics and correlation analyses to explore relationships between blink-derived metrics and subjective workload measures, these approaches are inherently limited to linear and pairwise associations. More advanced machine learning techniques may enable the extraction of additional information by modeling non-linear relationships and interactions among multiple blink features simultaneously, as well as by capturing temporal patterns across task execution.

In particular, machine learning algorithms could integrate multiple blink parameters, performance metrics, and contextual task variables to identify complex patterns associated with cognitive load that are not evident through conventional statistical analyses. Such approaches may also improve sensitivity to subtle changes in workload and support individualized models that account for inter-subject variability. Therefore, while the current findings indicate that blink metrics alone are not sufficient to robustly infer cognitive load, machine learning-based methods represent a promising direction for future research aimed at enhancing multimodal workload estimation.