Towards an Eye-Brain-Computer Interface: Combining Gaze with the Stimulus-Preceding Negativity for Target Selections in XR

The Dwell technique has been the most common gaze interaction technique [24, 25, 30, 43, 75, 76] in HCI literature. Prior works like Jacob [43] used Dwell for selections (who also first noticed the Midas touch) and other works like Starker and Bolt [109] leveraged Dwell to make the system provide more information after a fixed period, sort of an interest metric. Past research has also explored optimal parameters for the Dwell technique, such as minimum target size [120] or dwell threshold [75, 98, 116], finding that it is heavily dependent on the task, context, and cost of errors. [90].

Researchers have studied gaze as both an individual and a multimodal system, where other modalities usually confirm the user’s gaze target [107, 108], addressing the Midas touch. Hence, it is a two-step procedure where the user’s gaze acts as a cursor and the selection is triggered by an additional modality. Various modalities have been investigated, including eye blinks [71], speech [63, 96, 122], head movements [115, 121], handheld devices [29, 64], tongue-based interfaces [34], body movements [57], and hand gestures such as the pinch gesture [64, 82, 99]. Additional gestures, such as gaze-hand alignment [74] confirm target selection either by aligning a ray that emanates from the user’s hand with their gaze (Gaze&Handray) or aligning the user’s finger with their gaze (Gaze&Finger) [117]. However, gestural interfaces can lead to physical fatigue [45], making them difficult to use for extended periods. Gaze interactions may also cause eye fatigue or strain [97], but the direct measurement of their impact is not yet known, as they are assessed using devices that themselves induce fatigue symptoms [37]. Moreover, research has found less eye strain when gaze was used only for pointing, rather than being used for both pointing and selection [68].

Research has also explored using natural eye movements as confirmation triggers, such as voluntary blinks [71]. In contrast, some studies have explored half-blinks [65], removing the need to differentiate between spontaneous and voluntary blinks. Smooth Pursuits [114], offer another approach that bypasses the need for precise gaze tracking or a prior eye-calibration step, and this technique has found applications in XR environments [28, 52]. Vergence-matching [2, 112] is a method that gauges the simultaneous opposing movement of both eyes, correlated with subtle changes in a target’s depth. If the vergence movement corresponds to a specific target’s depth alteration, that target is identified as the user’s point of focus. Though this approach has proven effective in selecting small targets [106], it is often perceived as uncomfortable, difficult to execute consistently, and diverts the user’s attention [56].

Model predictions using head and eye endpoint distributions were explored by Wei et al. [121], taking into account the distributions of the head and eye positions/orientations during target selection. They developed a robust target prediction model that surpassed the baseline. Other works like Bednarik et al. [5] have explored predicting intents using gaze metrics alone such as saccade and fixation durations, saccade amplitudes, and pupillary responses. They achieved an accuracy of 76% with Area Under Curve (AUC) around 80%. David-John et al. [14] similarly explored intent prediction in VR, using gaze features such as the K coefficient [61], gaze velocity, and so on; predicting intent significantly above chance with the the Receiver Operating Characteristic(ROC)-AUC scores showing a maximum of 0.92 and a minimum of 0.71.

Gaze signals, when combined with BCIs, give rise to hybrid BCIs (hBCIs). This fusion is commonly adopted to increase the range of control commands, enhance classification accuracy, or reduce the intent detection time [39, 91]. In studies conducted between 2009 to 2016, 52% of them investigated this merger of gaze and neural signals [39], including gaze captured from both Electrooculogram (EOG) and an eye-tracker. EOG signals, known for their temporal precision, are primarily used to reduce ocular artifacts leading to lesser false positives in classification [4, 40, 81, 113]. Gaze signals have also been used to improve control commands such as the control of a wheel-chair [55], where the classifier achieved an accuracy of 97.2%. Kalika et al. [47] combined gaze with EEG in a P300-based speller system. The integrated system surpassed the EEG-only setup in accuracy and bit rate.

Eye blinks have also been incorporated with both MI and P300-based BCI paradigms for wheelchair control [119]. Pfurtscheller et al. [91] examined an MI-based hBCI solution to the Midas touch problem using eight subjects in a search-and-select task. The hybrid approach (EEG + Gaze) proved significantly more accurate in selections compared to the Dwell technique implemented using gaze alone. Nonetheless, the hBCI was the slowest activation method, while demanding conscious effort from the user. Other works have investigated MI and gaze-based BCIs, achieving higher accuracy, yet voluntarily imagining motor movements was challenging for users [20].

A notable application by Kim et al. [54] showcased a unique combination: gaze signals from an eye-tracker combined with the SSVEP paradigm to direct a turtle’s movement. Directional commands (e.g., turning left or right) were controlled by SSVEP, whereas reset and idle commands were activated by simply opening or closing the eyes. McCullagh et al. [80] also examined the SSVEP with the Dwell technique, finding users lacking technical proficiency, encountered difficulties navigating the interface. In a related study, Putze et al. [95] combined gaze with SSVEP for AR-based smart home management, yielding better accuracy than SSVEP by itself. Nonetheless, the flashing visual stimuli of SSVEP targets disrupted users’ cognitive flow.

A passive hBCI combines signals from multiple modalities, and derives its output from implicit brain activity without demanding active user participation, with the goal of enhancing interaction techniques by making them more seamless and effortless. Prior works have explored such interfaces like Zander et al. [127], who used error-related potentials (ErrPs) to guide a two-dimensional cursor to its desired target based on implicit user states from the pre-frontal cortex. If the cursor’s position mismatched the user’s expectation, an ErrP was triggered, with its peak amplitude correlating with the error magnitude. In a similar vein, Krol et al. [62] designed a Tetris game where gaze and ErrPs were the primary controls: Gaze determined block movements while factors like the player’s relaxed state influenced game speed, and ErrPs eliminated undesired blocks. Others have also employed ErrPs to rectify inadvertent Dwell technique selections [46].

Research efforts on passive BCIs has also sought to classify user intent. For instance, Sharma et al. [102] achieved a 97.89% accuracy in discerning intent during free-viewing, target searching, and target presentation conditions, utilizing EEG features such as power spectral intensity and detrended fluctuation analysis. However, given the task’s emphasis on visual searches, it’s likely that the primary EEG components were the P300 wave or reward potentials [3, 32]. This suggests that the model’s accuracy might differ when selecting familiar targets, as is the case for everyday UI interactions. Previous studies have also investigated the classification of user intent to interact with virtual objects by analyzing EEG and Electromyography signals, achieving detection accuracies of pre-movement states up to 69% [84]. While not directly stated, this research likely leveraged the CNV wave, an anticipatory potential akin to SPN but linked to motor activity [8]. Kato et al. [49] also used the CNV to develop a switch for activating or deactivating a BCI system. In other works [123], the CNV was used in conjunction with a P300-based BCI to improve its accuracy from 85% to 92.4% in identifying visual targets. While the CNV appears promising as a passive mental ’click’, its reliance on imagined motor activity can distract users from their primary task.

Conversely, the SPN is tied to nonmotor activity [8]. An early study by Protzak et al. [94] used this slow negative wave, achieving a mean accuracy of 81% when distinguishing between intent-to-select and intent-to-observe interactions. This was attributed to pronounced negativity in the parietal region of interest (ROI) during intent-to-select interactions. The study involved target selections on a two-dimensional screen using a 1000ms Dwell threshold. The SPN in this study was associated with the stimuli indicating a successful selection, i.e., the task proceeding to the next trial, rather than being an explicit, separate visual or auditory stimulus. However, the study focused on visual searches where the prominent P300 component [3, 32] could influence classification results, potentially affecting intentions towards familiar targets. Moreover, the study was limited by the lack of tests for statistical significance and absence of details on ocular artifact corrections.

Shishkin et al. [104] demonstrated that the SPN, in the parietooccipital ROI, can distinguish intent-to-select from intent-to-observe fixations within 300ms segments. In their study, eight participants engaged in a gaze-controlled game, using the Dwell technique to select and maneuver game pieces. The authors explored both 500ms and 1000ms dwell thresholds, which were found not to influence the SPN. Visual feedback was provided for intent-to-select interactions only, likely to induce the SPN wave. While EOG was recorded, it was not used for ocular artifact correction. Not correcting for these artifacts can increase false positives in classification [4, 40, 81, 113]. Instead, the authors analyzed windows believed to be free from such contamination.

In a later study, Zhao et al. [129] corroborated the SPN’s presence in voluntary selections of moving objects and successfully distinguished between intent-to-select and intent-to-observe smooth pursuits in 8 participants among 20. They employed a two-dimensional task where participants explicitly selected 15 targets sequentially, while intent-to-observe interactions involved a counting task. Notably, heightened negativity was recorded at the Oz and POz electrodes during intent-to-select pursuits compared to intent-to-observe ones. Only intent-to-select interactions received visual feedback, where the target turned green upon a successful selection.

The University of Colorado recruited 22 participants through online bulletins, but rejected 5 due to technical issues, difficulties with the counting task, or the eye-tracker’s failure to accurately track gaze. 17 participants (9 male and 8 female), averaging 28.8 years (SD = 10.7) are analyzed in this work. One participant had moderate experience with virtual and augmented reality, 10 participants had minimal experience, and 6 had no experience; 5 participants had experience with gaze-based control. Participants who required corrective eye-wear were excluded due to gaze tracking concerns, although clear contact lenses were allowed and 4 participants wore them; 9 participants had naturally normal vision and 8 had corrected-to-normal vision. All participants reported free of any neurological disorders. They were paid 15 USD per hour for the two-hour study.

The experimental setup, as seen in Figure 2, was comprised of an (i) HP Reverb G2 Omnicept Edition VR headset with an integrated Tobii eye-tracker ¹, and (ii) a 64-channel BioSemi ActiveTwo EEG system² with four additional EOG electrodes and two reference electrodes that were recorded on the same amplifier. 64 EEG electrodes were placed in a 10-10 montage across the scalp [87]. The reference electrodes were placed on the left and right mastoids, and the EOG on the outer canthus and below the orbital fossa of each eye. We made four slits in the HMD’s face cushion to reduce recording interference and pressure points on the EOG electrodes. All electrode impedances were kept below 50 kΩ and referenced to the Common Mode Sense active electrode during the recording. EEG and EOG data were recorded at 512 Hz. The Tobii eye-tracker, sampling at 120Hz, identified the user’s gaze point in real time, while also logging pupillometry at the same rate.

The VR headset was tethered to a computer with an Intel i7-9700K @ 3.6GHz CPU, RAM of 16GB, and an NVIDIA GeForce RTX 2080 Super GPU. Our tasks, which were developed and run on Unity³, maintained a 90fps frame rate, fully leveraging the HMD’s display capacity. The HMD used inside-out tracking for six-degrees-of-freedom tracking, had a resolution of 2160 x 2160 per eye, and a 114° field of view. A separate computer was connected to the BioSemi amplifier, on which all data was logged. The EEG, EOG, eye-tracker, and an experimental marker stream (which highlighted interaction events and other key moments in our task) were synchronized and logged using the Lab Streaming Layer⁴, facilitated by wired LAN connections between the systems.

Participants engaged in VR-based three-dimensional object selection tasks modeled after Zhao et al. [129], though the objects remained static in our study. The objects comprised of 15 targets, i.e., numbered blue spheres spread throughout a three-dimensional space as depicted in the left image of Figure 3. The nearest target occupied a visual angle of 19.6°, and the farthest target had a visual angle of 7.8°. The targets were distributed across an area that spanned a horizontal visual angle of 60.4° and a vertical visual angle of 47°.

Dwell-based selection mechanism. The HMD’s eye-tracker provided real-time coordinates of the gaze origin and the normalized gaze direction. Using Unity’s Raycast function, we projected a virtual ray from the gaze origin in the direction of this normalized vector. The point at which this ray intersected an object in the environment indicated the user’s current visual focus. When the user’s gaze landed on a target, the ray intersected the target’s collider, changing its color from blue to red, as seen in the right image of Figure 3.

Prior research found the presence of the SPN at dwell thresholds of both 1000ms and 500ms [104]. We chose a 750ms threshold for our study as a balanced approach, considering the novel challenge of detecting an SPN in immersive environments. The dwell countdown started when the user’s gaze first landed on the target (or when it turned red). If the user’s gaze remained fixed on that target for the full 750ms duration, the target illuminated in bright green (as seen in the middle image of Figure 3) accompanied by the audio cue, signaling a successful selection. It is this bright green visual stimuli that we expect the user to anticipate in intent-to-select interactions. The target turned back to blue when the user’s gaze shifted away from it.

The tasks involved interactions across three conditions: (i) Intent-to-select, (ii) Intent-to-observe without feedback, and (iii) Intent-to-observe with feedback.

We employed a within-subjects design, with participants completing three blocks, each containing five phases as shown in Figure 4. In each phase, the four tasks (depicted by the blue boxes) were randomized. Participants performed a total of 450 intent-to-select interactions across all three blocks. Each block consisted of 150 interactions, further broken down into 30 interactions for each phase, with each task comprising 15 interactions. For intent-to-observe interactions, there was no fixed number of interactions as the task was designed to be exploratory. Nonetheless, each task required participants to identify at least 5 targets marked with the specified number of dots, resulting in a minimum of 10 interactions for each phase. This translates to at least 50 interactions in each block and a total of 150 interactions minimum across all three blocks. The actual number of interactions varied from task to task for every participant.

At the start of every intent-to-observe interaction task, the number of dots to find, the number of dots present on each target, and the position of the dots on each target were randomized, with the number of dots to find falling within a range of 5 to 8. To aid familiarity with target locations, the target numbers remained in the same spatial locations within each block (phases 1-5). This was done to specifically probe H3. However, these numbers were randomized at the start of each block.

Upon arriving at the lab, participants signed consent forms per the University’s IRB (IRB# 23-0156) and completed a demographics questionnaire. Next, the EOG, reference, and EEG electrodes were attached to the participants, with electrode gel applied to ensure optimal signal conductance. Participants received task instructions, with specific guidelines such as making slow head movements. They were advised to approach the tasks at a comfortable pace and to focus exclusively on the tasks. They were also cautioned against speaking during the study and were reminded to maintain relaxed facial muscles to prevent interference with EEG signals. Before starting the tasks, participants wore the HMD and underwent an 8-point native eye-calibration.

Before the main task, participants completed a tutorial encompassing one phase of all four tasks. They then proceeded to complete the three blocks, each lasting approximately 20 minutes and each followed by a 5-minute break where they could remove the HMD. The eye-tracker was recalibrated at the beginning of each block and any time the participant put the HMD back on after removal.

Data processing and analysis were conducted offline through custom MATLAB scripts leveraging the EEGLAB toolbox [16], along with ERPLAB [69] and the EYE-EEG plugins [19]. The eye-tracking data was used for the detection of eye events. First, blinks were identified by recognizing typical gaps in gaze behavior ranging from 50 to 700ms [38]. Any gaps shorter than 50ms were interpolated and those larger than 700ms were deemed data dropouts. Subsequently, using a velocity-based algorithm [26, 27], adapted from the EYE-EEG plugin [19], we identified saccade and fixation events. Specifically, a velocity factor of six (or six times the median velocity per subject) combined with a minimum saccade duration of 12ms was applied. Fixation events were recognized as eye data with velocities below this velocity factor. To remove biologically implausible eye movements and outliers obtained from the previous velocity factor classification, we focused on fixations that lasted between 100 to 2000ms and saccades meeting the following criteria: (i) a 12 to 100ms duration, (ii) a 1 to 30° amplitude, and (iii) a velocity ranging from 30 to 1000° /sec. We then merged the eye-tracker data with the EEG data (upsampling the eye data to match EEG), including the identified blink, saccade, and fixation events.

The preproccessing for the ERPs followed an approach akin to the ICA preprocessing. We resampled the EEG signals from 512 Hz to 256 Hz, eliminated 60 Hz line noise using the Zapline-plus extension for EEGLAB [15, 58], and re-referenced the data using the average of the left and right mastoids. Bad channels were interpolated as mentioned earlier, followed by application of a 4th order 0.1 to 40 Hz Butterworth bandpass filter. ICs marked for rejection during ICA decomposion were removed and the remaining ICs were backprojected to this new pre-processed file. In this way, the ocular artifacts could be substantially reduced in our EEG signal. We used ERPLAB to epoch the data, extracting ERPs time-locked to the selection / interaction event within a -2000ms to 1500ms window. For intent-to-select interactions, only the events pertaining to correct selections (sequentially triggered) were used, and inadvertent selections were discarded. Moreover, intent-to-select interactions were reduced to the mean of the ascending and descending tasks for further analysis. For intent-to-observe interactions, all trials that led to the interaction trigger (750ms) are used, given that the accuracy of the counting task is not pertinent to our analysis.

The ERPs were baseline corrected from -1000 to -850ms as we were primarily interested in the -750 to 0ms window where an SPN is to be expected. We avoided the -850 to -750ms range due to the fixation-related Lambda response resulting from the first fixation on the target stimulus. ERP epochs with artifacts were detected and excluded using a 100μ V sample-to-sample voltage threshold, resulting in roughly 2.6% rejection rate per condition across all participants. The average number of trials used for the grand average ERP for each task is shown in Table 1. We focused our analysis on the occipitoparietal ROI (O1, Oz, O2, Iz, PO7, PO3, POz, PO4, PO8, P5, P6, P7, P8, P9, and P10 electrodes as depicted in Figure 5), in line with prior findings highlighting the SPN’s prominence in these regions [94, 104, 129].

In this study, we sought to extend our understanding of the neural underpinnings involved in an intent-to-select interaction within a dynamic VR environment where visual information is processed over a larger area, as opposed to constrained visual fields. We assessed this using an object selection task where users intentionally selected targets, and a control condition in which they gathered information about a target (intent-to-observe interactions). We examined ERPs that were time-locked to the interaction trigger, with each trigger based on a 750ms dwell threshold.

We hypothesized (H1) that users in the Intent-to-select condition will anticipate a stimulus feedback from the target as a result of the selection and hence elicit an SPN, in contrast to intent-to-observe interactions. This is evident in the amplitude topographical maps shown in Figure 6, where noticeable negativity emerges from the occipital ROI in the averaged -400ms to -300ms window during intent-to-select interactions only. Our second hypothesis (H2) posited an SPN’s absence in the Intent-to-observe with feedback condition, as users, despite stimulus feedback, did not anticipate it due to the lack of intent to select the target. As seen, intent-to-observe interactions—both with and without feedback—displayed minimal negativity, indicating that eliciting of the SPN wave is dependent on the user’s intention.

In Figure 7, we look more closely at the grand average (n = 17) ERPs, time-locked to the interaction trigger and averaged across the occipital ROI (O1, Oz, and O2 electrodes) because this region showed a strong emergence of negativity in the topographical maps and prior work’s focus on this region [104, 129]. The left plot contrasts Intent-to-select with the Intent-to-observe without feedback condition and the right plot contrasts Intent-to-select with the Intent-to-observe with feedback condition. For the Intent-to-select condition, the ERPs for both the ascending and descending tasks are plotted. As expected, the ERPs depict a Lambda wave positively peaking at -750ms for all three conditions. A Lambda wave is typically observed in the occipital ROI during post-saccadic eye movements, and is said to be involved in the visual processing of new information [100, 111, 125]. Another observation of interest is the Visually Evoked Potential from the target turning green, peaking at 300ms for the all but the Intent-to-observe condition where feedback was not provided (left plot).

In everyday human-computer interactions, users often know the locations of frequently-used UI elements, eliminating the need to search for a desired button. Given this, we examined how familiarity with UI placements impacts an SPN during intent-to-select interactions. Our hypothesis (H3) posits that as users become familiar with target locations, the SPN will manifest more rapidly. We intentionally kept the target positioning consistent for every block, allowing participants to familiarize themselves with their placements. Participants went through all 5 phases (10 rounds of intent-to-select tasks, totalling 150 interactions for each participant) in a block with the same target positioning. The target positioning was randomized at the start of each block.

We hypothesized that as participants learned target placements, the time to complete the intent-to-select tasks would reduce phase-by-phase, as a result of reduced target searches. This trend is evident in the left plot of Figure 9, which shows the mean task completion times for the five phases, averaged across participants and blocks. An ANOVA revealed a main effect (F(4,80) = 7.2, p < 0.001), with Tukey’s post-hoc tests confirming a significant difference between phase 1 and phase 5 times at p < 0.001. Accordingly, we compared the grand averaged (n = 17) ERPs in phase 1 (1471 trials with 1.9% rejected) with phase 5 (1429 trials with 2.8% rejected). The middle plot of Figure 9 illustrates the topographical distributions for the -400 to -300ms time window, time-locked to the interaction trigger. The top and bottom rows depict phase 1 and phase 5, respectively. Using a 0.1 Hz high-pass filter (HPF), as detailed in Section 4.6, a marked increase in negativity is observed for phase 5 compared to phase 1, suggesting a faster emergence of the SPN. However, a one-tailed within-subject MU t-test based on the cluster-level mass permutation approach did not reveal any significant effects (p > 0.05) for the difference wave (phase 1 ERP - phase 5 ERP). 2500 random permutations were tested and the test included all data points from -750 to 0ms, sampled at 256 Hz at all 15 electrodes, totaling 2895 comparisons. We used the MU ERP Toolbox [36], and our choice for MU analysis remains the same as stated in Section 5.1.1.

In line with previous works that examined EEG negativity in intent-to-select versus intent-to-observe interactions [104, 129], we omitted the HPF. Interestingly, in Figure 9’s scalp maps, without the HPF, a more prominent emergence of negativity is seen for phase 5 than phase 1 during the same time window, specifically at O2. Figure 9’s right plot displays the grand average (n = 17) ERPs for O2 in phases 1 and 5, averaged across all blocks. Dotted lines depict ERPs with the 0.1 Hz HPF, while solid lines represent ERPs without it. We see a considerably larger difference in the ERPs without the HPF, in contrast to when it’s applied. We used the fractional area latency metric [53, 69] to compare onset timings between phases 1 and 5, identifying the latency (onset) at which 50% of the SPN’s negative AUC has been reached. With the HPF, phase 5’s SPN covered 50% of the negative AUC 23.44ms earlier, with its latency measured at -226.56ms, compared to phase 1’s latency of -203.12ms. Without the HPF, phase 5’s SPN was 35.15ms sooner, with its latency measured at -226.56ms, compared to phase 1’s -191.41ms.

Notably, when the HPF is removed, we see a significant cluster for the difference wave at O2 starting from -516ms until 0ms, as seen in the bottom plot of Figure 10, with the color gradient depicting the t score. The parameters for the test was similar as for the one with a HPF. The top plot of Figure 10 illustrates the mean amplitude distribution (top) and the t score distribution (bottom) for the difference wave with and without the HPF. Interestingly, only O2 shows a significant effect without the HPF, which could be attributable to the fact that an SPN is stronger in the right hemisphere [6].

We would like to emphasize that a HPF is highly advised to to mitigate low-frequency environmental noise, including artifacts from factors such as user perspiration-induced slow drifts [48, 72]. This is particularly crucial for real-world BCIs employing active dry electrodes with high impedances, even though it may influence slow ERP components [79]. Consequently, we retained this filter for H1 and H2. The absence of statistically significant differences in the ERPs with the HPF makes it challenging to reject our null hypothesis for H3. Nevertheless, the significance without the filter further strengthens our direction that the SPN emerges more rapidly when users become familiar with the target placement. The implications of this finding can assist real-time BCI systems to optimize the time windows used for classification depending on the user’s familiarity with the visually focused target.

We have identified limitations that future research can address. Firstly, the Dwell technique can lead to user fatigue [37, 42, 98]. While prior work have tried to identify an optimal threshold duration, a definitive one remains unestablished [75, 98, 116]. Second, we did not develop a classification model, and hence cannot compare how well (accuracy) an intent-to-select interaction can be distinguished from an intent-to-observe one for single trial ERPs. Furthermore, as this study is an initial exploration of the SPN’s potential in tackling the Midas touch problem, it is not yet feasible to compare our approach with an established technique that resolves this issue. Yet, our robust statistical findings hint at the potential to design a hBCI leveraging the SPN, and has the capability to be a faster, effortless, and fluent interaction paradigm than SSVEPs or P300-based BCIs.

Third, further research on real-time preprocessing, ocular artifact correction, and addressing strong motion artifacts, is vital to effectively mainstream this anticipatory interaction paradigm for real-time use. Gherman and Zander [35] found that different postures did not have an impact on a passive BCI’s classifier performance, strengthening our case for a real-time system. Fourth, it should also be noted that our high-resolution EEG data originated from a controlled lab environment using research-grade EEG equipment. Exploring the efficacy of this approach in real-world scenarios with dry electrodes, warrants further investigation. Fifth, intent-to-select and intent-to-observe interactions were restricted to their own conditions, further research is necessary in examining these neural signatures when these conditions are intermixed, like they would be in a real world setting.

Finally, our study did not include a condition decoupling feedback from intent-to-select interactions to ascertain the feedback’s role in SPN elicitation. However, prior work has demonstrated the SPN’s absence without feedback [11] and the necessity of temporal predictability or instilled memory of the target providing feedback for SPN emergence [8, 21, 22]. This aligns with our findings in Figure 9’s right plot and as stated earlier, phase 5 ERP shows greater negativity than phase 1 ERP, due to a more instilled memory of the feedback by phase 5. Furthermore, the SPN’s prominence in the occipital region, known for processing visual stimuli, confirms that it is linked to the visual feedback [8]. Future research can examine the SPN’s behavior in the sudden cessation of feedback during intent-to-select interactions.