Fifteen healthy young adults (seven females, eight males; mean age = 24.38 ± 2.98 years) without any neurological disorders or other impairments voluntarily participated in this study. Due to a technical issue with the EEG system during data acquisition, specifically the failure of a EEG or EMG electrode, data from two subjects were excluded from further analysis. This malfunction resulted in unusable data for those channels, rendering the entire dataset for those participants unusable. This exclusion was based on the objective criterion of complete electrode failure. Consequently, the final dataset comprised 13 participants (seven females, six males). All participants provided informed consent prior to their inclusion in the experiment. The study protocol was conducted according to the Declaration of Helsinki.

During all sessions, monopolar EEG signals were recorded using an active EEG electrode system (g.GAMMAcap; g.tec Medical Engineering, GmbH, Austria) connected to a g.USBamp amplifier (g.tec Medical Engineering, GmbH, Austria). Ten electrodes were placed over the motor cortex according to the standard international 10–20 system: FP1, Fz, FC1, FC2, C3, Cz, C4, CP1, CP2, and Pz. The ground electrode was placed centrally on the frontopolar region (FPz), and the contralateral earlobe was used for reference. Surface EMG electrodes were placed on the musculus tibialis anterior (TA) to determine movement onset (Figure 1). All data were sampled at a rate of 1,200 Hz.

During the actual experiment, subjects interacted with a custom-made speller interface consisting of a 6 × 6 matrix design (Farwell and Donchin, 1988) with characters to choose from and a continuously moving selector (Figure 2). The selector consists of a gray bar that moves horizontally (x-axis) until a selection is made, then changes direction to move vertically (y-axis).

The BCI speller operates as a simulation, where letter selections were predefined, and the system automatically pauses at the corresponding column or row, assuming that the subject would initiate movement during the designated time. Each column and row selection lasted for 2 s. After each selection, the bar paused for 2 s before resuming movement for the next selection. The speller includes a blank area at the top displaying the selected letters or a predefined phrase, with the last selected letter highlighted.

For each condition, the target phrase was presented at the beginning of the experiment in the Phrase Area of the speller interface (Figure 2). The selector moved from left to right, sliding over the columns (which were internally divided within the speller matrix). Once the selector reaches the end of the predefined column of the target letter, it automatically changed direction and moved from bottom to top to stop at the end of the corresponding row. The rows and columns were visually marked with dividers to aid user orientation (Figure 2). Once a letter was selected the bar reset its position at the left of the screen to restart the sliding. Participants were cued to execute a movement when the selector reached the target location (column/row) of the letter. This ensured that MRCPs were aligned with voluntary movement at the desired letter.

All measurements were conducted at the neurophysiological laboratory of the Department of Sports and Sports Science of the University Freiburg.

The subjects were instructed to perform a ballistic dorsiflexion of the dominant leg, executing the movement when the bar was around the midpoint of the column or row containing the desired character. Each subject completed three different phrases, each consisting of 13 characters, resulting in 26 trials or 26 possible MRCPs per phrase (Table 2).

To evaluate MRCP performance in conjunction with a simulated spelling task, three conditions were tested. These conditions were designed to assess how MRCPs are affected by the varying demands of a spelling task, ranging from minimal cognitive load to more complex scenarios involving memory and letter selection. Cognitive load, in this study, refers to the mental effort imposed on participants by the task demands, including factors such as memory, attention, and decision-making. The three conditions were: a control condition using repeated selections of the letter “O” to isolate movement-related brain activity with minimal cognitive load; a phrase spelling condition with structured text (“HELLO IM FINE”) to simulate a meaningful spelling task with moderate cognitive load, reflecting the memory and planning involved in spelling a familiar phrase; and a random condition using a randomized sequence of letters to introduce higher task complexity, as participants needed to process and select letters without the aid of linguistic or semantic context, thus increasing demands on attention and working memory.

All signal analysis was done with a custom MATLAB script and GUI (version 2022a, MathWorks, Natick, USA). The CZ electrode was chosen for the subsequent analysis, as it is the electrode that is highly associated with lower limb movement (Do Nascimento et al., 2006; Shibasaki et al., 1986; Shibasaki and Hallett, 2006). The EMG data was used as supplementary data for the detection of the movement and as part of the criteria for selecting the MRCPs.

A simplified signal processing pipeline was implemented, avoiding computationally expensive techniques such as Independent Component Analysis (ICA) for artifact detection. Instead, we employed filtering and epoch-based processing techniques that are compatible with real-time BCI applications. Both the MRCP traces and their features were extracted and analyzed (Figure 3), including peak negativity, readiness potential slopes, negative slope, motor potential, and overall MRCP amplitude (PN, NS1, NS2, rebound, and PP). The preprocessing steps are described as follows (Figures 4, 5):

Line noise was removed by applying a 50 Hz notch filter using a 2nd order Butterworth filter. EEG channels were then filtered using a 2nd order Butterworth band-pass filter ranging from 0.05 to 3 Hz. Next, EEG and EMG data were segmented into epochs 4 s before to 5 s after column/row onsets (Figure 6). A column/row onset was defined as the moment the selector bar intersected the column or row containing the target letter. Offset correction was applied per epoch by demeaning – that is, by calculating the mean for each of them and subtracting it. To improve spatial resolution, a large Laplacian transformation was applied to the Cz electrode. EMG onset detection was performed using an adapted version of Yang’s algorithm, which identifies movement initiation by detecting deviations from baseline activity and image processing techniques (Yang et al., 2017). This adapted version included modifications to the threshold levels for detecting EMG activity and adjustments to the window length for morphological image processing, such as opening and closing techniques. The window duration was set to 1.5 s, based on the participants’ intent to sustain movement for approximately 2 s. This method was used solely for confirming movement execution and not for training or classifying MRCPs in real-time.

Data analysis focused on the Cz electrode and the Laplacian result of this electrode for each trial. Column/row onsets were used to segment the data and define individual trials. The EMG signal was then analyzed to confirm that the subject performed the intended movement within the appropriate time window. Both segmentation and movement verification were performed using a custom MATLAB graphical user interface (GUI).

After confirming the movement, the segmented data (trials) were inspected with the GUI to verify the presence of MRCPs. The GUI was also used to identify the Peak Negativity (PN) within each trial. A trial was considered successful, indicating the presence of an MRCP, if it met all the following criteria: (1) PN and EMG onset occurred within the time window of −0.5–0.5 s relative to the column/row onset; (2) the MRCP exhibited two distinct negative slopes followed by a rebound phase; (3) PN corresponded to a local minimum in the MRCP; or (4) PN occurred more than 1 s before the column/row onset (Figure 3).

This stepwise approach ensured that data segmentation, movement verification, MRCP validation, and PN identification were carried out systematically, resulting in the reliable selection of valid MRCPs across all trials.

The number of successful trials per phrase, condition, and subject was calculated as a percentage of the total trials (n = 26). In this study, this metric is referred to as the success rate, which indicates the number of times a subject successfully elicited an MRCP, as determined based on the inclusion criteria outlined in the previous paragraph. Success rates were calculated separately for both Cz and Laplacian-processed data. To standardize onset, three datasets per phrase were created, with MRCPs aligned based on three different time points: the column/row onset, the EMG onset (defined as the start of EMG activity), and the Peak Negativity (PN) timing. Each dataset was centered around the corresponding time point (column/row, EMG, or PN) and consisted of epochs extending 2 s before and 1.5 s after the specified time point. The PN-shifted dataset was selected for further analysis due to phase cancelation observed in other shifts.

Grand averages were computed per condition by grouping selected trials by condition, while Statistical Parametric Mapping (SPM) was conducted by dividing selected trials by subject and condition. For each subject-condition combination, a mean was calculated, yielding 39 average MRCPs (13 subjects * 3 conditions). These were analyzed using one-way ANOVA in MATLAB’s SPM toolbox (Pataky, 2022), with post-hoc paired t-tests and Bonferroni correction if significant results (p < 0.05) were found.

Descriptive features (PN, NS1, NS2, rebound, PP) were extracted (Figure 3), per trial and condition. Finally, statistical analysis involved computing the mean of each feature per subject and condition. Organized arrays were created with columns representing conditions and rows representing subjects. Five arrays, each representing a feature from both preprocessing methods (non-Laplacian and Laplacian), were analyzed using Friedman’s test. If significance was detected, a Wilcoxon signed-rank post-hoc test with Bonferroni correction was applied for the feature.

To assess whether the interface and spelling task affect the morphology of the MRCPs, features were extracted from all selected trials for each subject and condition. The mean of each feature was computed for each subject and condition, and repeated for all features. These means were then statistically compared for both processing methods and the results are shown in Table 5 (non-Laplacian) and Table 6 (Laplacian).

From the Friedman test results, only one feature showed significance (p < 0.05). This feature was the PP in the Laplacian dataset (p = 0.0231). A Wilcoxon rank-sum post-hoc test was performed for pairwise comparisons with a Bonferroni correction (adjusted significance level: p < 0.0167), with results presented in Table 7. No significant differences were found in this test. The comparisons of the features between the control condition and the phrase (p = 0.1272) and against the random condition (p = 0.0574) were the smallest, but neither showed significant differences.

To compare the three conditions, we calculated the grand average of the MRCPs for each condition and conducted an SPM analysis. The grand averages were obtained by averaging the individual MRCP traces over three conditions (Figure 8). The resulting grand averages across conditions exhibit similar morphology, with overlapping confidence intervals.

To further investigate these similarities, an SPM analysis was conducted using a one-way ANOVA. For this analysis, MRCPs were first averaged per subject for each condition, resulting in 39 averaged MRCPs (13 subjects * 3 conditions). The SPM analysis confirmed that there were no statistically significant differences between the three conditions, consistent with the observed grand averages.

The Laplacian data showed significant differences across conditions except between −1.56 s and −1.3 s (Figure 9). Post-hoc paired t-tests with Bonferroni correction (Supplementary Figure 1) revealed the following: For the control vs. phrase condition, only the interval from −1.55 s to −1.0 s was not significant (p ≤ 0.2). For the letter vs. random condition, significant differences were found between −2.0 s and −1.88 s, −0.53 s and −0.36 s, and −0.17 s and 1.5 s, with the letter showing the greatest difference. The phrase vs. random condition had a significant difference only between −0.94 s and −0.4 s (p ≤ 0.2).

In the field of BCI, selecting an appropriate control signal that accurately reflects the user’s intent is critical for the system’s effectiveness (Abiri et al., 2019; Krusienski et al., 2012; Wolpaw et al., 2002). In this study, we chose MRCPs as a potential control signal for BCI spellers.

We implemented the success rate, defined as the occurrence of an MRCP, as a metric to evaluate the correlation between the user’s intent and the actual brain output measured by EEG. High success rates suggest that the MRCP has the potential to serve as a reliable control signal, capable of effectively capturing the user’s intent.

Compared to the control condition, the success rates of the phrase spelling condition met expectations, with 9 out of 12 subjects achieving at least 69% success. Notably, six participants were novices, suggesting that MRCPs can be elicited without extensive user training (Jochumsen et al., 2018; Niazi et al., 2013). This trend persisted in the phrase spelling condition, where 10 participants matched or exceeded their performance, supporting the idea that MRCP-based control can be trainable (Jochumsen et al., 2017). While this may appear lower than classification accuracies reported in P300 and SSVEP studies (Abiri et al., 2019; Rezeika et al., 2018), it is important to note that participants in this study had no prior experience or training in eliciting MRCPs. Despite this lack of user training, MRCPs were successfully elicited, suggesting that performance could improve with additional practice and adaptation.

On the other hand, the random condition revealed that just seven subjects achieved an equal or higher success rate than 69%. This decrease could be attributed to the diversion of attention (Aliakbaryhosseinabadi et al., 2017a; Aliakbaryhosseinabadi et al., 2017b), or the complexity of selecting random characters due to the lack of goal-oriented spelling (Frömer et al., 2012; Olsen et al., 2021).

The Laplacian filter improved or maintained the success rate in six subjects across at least two of three conditions but decreased in the others. As a spatial filter, the Laplacian reduces spatial noise by emphasizing local signal sources and minimizing distant ones. While effective in subjects with well-localized MRCPs, the filter can introduce distortions in cases where signals are more diffuse. Additionally, the Laplacian has been shown to counteract attention diversion during tasks, potentially explaining its improved performance in some subjects (Aliakbaryhosseinabadi et al., 2017a; McFarland, 2015).

The grand average MRCPs were computed based on three different onsets (column/ row, PN and EMG), revealing a phase-canceling effect for the column/row and EMG onset. The PN serves as the crucial point marking the end of the negative shift and the beginning of the positive shift (Do Nascimento et al., 2006; Shibasaki and Hallett, 2006). Therefore, it proves to be an excellent reference point for extracting the MRCP features with consistency. Additionally, this selection aligns with the goal of enabling non-muscular control signals as it remains present during motor imagery (Cunnington et al., 1996), which are essential for BCI applications, especially in patients with motor impairments.

Once an onset was selected the MRCPs were analyzed with and without Laplacian filtering. In the non-Laplacian data, MRCPs behaved as expected, showing no statistically significant differences across conditions, either in feature extraction or SPM analysis. Prior research indicates that MRCP morphology, particularly in features like NS1 and NS2, can change with attention diversion (Aliakbaryhosseinabadi et al., 2017a; Aliakbaryhosseinabadi et al., 2017b; Wright et al., 2012). The absence of significant changes here suggests that the spelling task did not impose a cognitive load that interfered with the movement timing or focus.

The use of the Laplacian filter in processing MRCP signals presents both benefits and challenges, particularly when comparing tasks with different cognitive demands. During spelling tasks, which are likely to engage broader cortical regions due to the increased complexity, the Laplacian filter may show amplitude differences in the MRCP. This occurs because the filter enhances localized neural activity but may smooth out signals spread across multiple channels, which is more common in cognitively demanding tasks. In contrast, monotonic tasks such as the control condition, are less affected by the Laplacian because they involve more focused neural activity that tends to produce clearer MRCP signals (Aliakbaryhosseinabadi et al., 2017a; Aliakbaryhosseinabadi et al., 2017b; McFarland, 2015).

Interestingly, in our analysis, while no significant differences were found in the extracted MRCP features, the SPM analysis of the Laplacian-processed data did show differences across conditions. This suggests that the Laplacian filter, while effective at reducing noise, might amplify specific aspects of the MRCPs and reduce their overall amplitude. This suggests that while Laplacian filtering can refine the signal by reducing noise, it may also alter the MRCP morphology, warranting careful consideration in future studies.

P300 and SSVEP spellers have established themselves as reliable BCI paradigms, consistently achieving accuracy rates of 80–95% with information transfer rates of 20–25 bits/min (Gannouni et al., 2022; Kundu and Ari, 2022; Rezeika et al., 2018). However, they require multiple stimulus presentations for each selection and can cause visual fatigue during extended use (Abiri et al., 2019; Gannouni et al., 2022; Rezeika et al., 2018).

In contrast, MRCP-based BCIs offer single-trial detection capability with accuracy rates of 75–85%, though their inherent low-frequency characteristics (~2 s per decision) may result in lower information transfer rates depending on the temporal analysis window (Jochumsen et al., 2020; Olsen et al., 2021; Savić et al., 2021). Recent advances in MRCP detection, including predictive algorithms for low-latency detection and hybrid approaches combining MRCP with ERD/ERS features to increase detection accuracy, show promise in reducing these speed and accuracy limitations (Savić et al., 2014; Savić et al., 2020).

MRCPs offer several advantages for BCI control, including reduced dependence on external stimulation, real-time feedback, and short latency, which could enhance usability and effectiveness in real-world applications (Savić et al., 2020, 2021). However, unlike SSVEP- and P300-based BCIs, which achieve high classification accuracy with minimal user training, MRCP-based systems may require further refinement and validation to reach comparable reliability. Although this study did not implement an online algorithm or automated classification, previous research has demonstrated the successful integration of MRCPs into BCI systems, highlighting their potential as a viable control signal (Savić et al., 2021).

While this study offers promising insights into the development of an MRCP-based BCI speller, several limitations should be considered for future research. Participants performed the spelling task without any prior training. Additional training could improve the reliability of MRCPs elicitation, as some participants exhibited increased success after just one session (Jochumsen et al., 2017). Furthermore, the experimental sessions lasted only 5–10 min, which does not reflect the real-world use of a BCI speller. Investigating MRCP viability for longer periods, accounting for fatigue, is essential for practical applications.

This study demonstrated that MRCPs can be elicited in BCI-naïve users during a simulated BCI-speller task. However, a crucial limitation is that the analyzed MRCPs were derived from *executed* movements, not imagined or attempted ones. This may constrain the interpretation of our results, as they primarily reflect the feasibility of MRCP-based BCIs in scenarios where voluntary motor execution is possible. While MRCPs can be elicited by both executed movements and motor imagery (Do Nascimento et al., 2006; Gu et al., 2013; Shakeel et al., 2015), MRCPs elicited during imagined movements are often weaker and exhibit different characteristics (Do Nascimento et al., 2006; Gu et al., 2013). However not so much is known with regards to attempted movement – thus where patients unable to move due to paralysis attempt to perform the task. Previous research has demonstrated the feasibility of using MRCPs elicited by motor intention (i.e., motor imagery) (Aliakbaryhosseinabadi et al., 2020; Jiang et al., 2015; Niazi et al., 2013; Savić et al., 2021), however, the inherent differences in MRCP characteristics between executed and imagined movements, particularly the potentially weaker signal strength during motor imagery, are critical considerations for future development of MRCP-based BCIs for communication and control. Future research must validate the applicability of this approach in individuals with limited motor abilities.

Even though MRCPs have the potential to be used in BCIs without external stimulation, the specific implementation in this study used a visual speller interface with moving selectors to guide letter selection. This visual interface, while different from flashing stimuli, could also potentially contribute to visual fatigue during extended use, although this was not directly assessed in the present study.

In this study, the presence of the MRCPs was determined through manual inspection of the EEG data using exclusion criteria predefined by the researchers. Each trial was evaluated based on these criteria, and only trials meeting the specified conditions were counted toward the success rate. This systematic approach ensured consistency in MRCP identification. To advance the practical application of MRCP-based BCIs, future work will focus on developing and implementing machine learning algorithms for automated MRCP detection and classification, which could improve accuracy and efficiency in both executed and imagined movement scenarios.

Additionally, while high-pass filters have been used to reduce eye and head movement artifacts, MRCPs are in the frequency range of these artifacts meaning the filters may not be fully effective. Future research should integrate eye tracking and head movement measurements to improve signal quality. The current study was conducted offline, and the next step is to perform online trials to assess real-time performance. In an online setting, MRCPs may be influenced by factors like fatigue, feedback and user errors, making another metrics such as false positives, precision, and accuracy crucial for evaluating system effectiveness.