Fourteen healthy male adults (mean age = 27.4 ± 3.5 years) participated in the experiment. Informed written consent was obtained from all participants. One participant was excluded for not completing the entire experiment, resulting in a final sample of 13 participants. All participants reported normal or corrected-to-normal vision and were right-handed. The study was approved by the Biological and Medical Ethics Committee of Beihang University and was conducted in accordance with the World Medical Association Declaration of Helsinki.

Considering that humans are more susceptible to attentional lapses during prolonged continuous tasks than discrete ones (Rahman et al., 2021; Reteig et al., 2019), we designed a novel continuous force control task to induce mind wandering, instead of using the discrete paradigms from prior studies (Peng et al., 2021; Zhang et al., 2023). Herein, participants held a pen-shaped handle to exert force while tracking a periodically varying target force. As shown in Figure 1a, participants were seated in front of a computer screen and held the handle with their dominant hand. They maintained a naturalistic pen-holding posture, mimicking handwriting. The handle was an end effector of a haptic device (Touch, 3D Systems Inc., United States). The haptic device allowed six degrees of freedom in movement and three degrees of freedom in force feedback within a 26.5 × 24.1 × 8.9 cm workspace. A virtual 3D scene was constructed, displaying a gray circular object, an annular object, and a pen-shaped object. The circular and annular objects remained fixed, while the pen-shaped virtual object was attached to the real-world handle. As participants moved the handle in the real space, the virtual pen synchronously performed the same movement on the screen.

During the task, a black dot moved clockwise around the gray annular region at a constant speed of 20°/s. The diameter of the black dot (D_o, in millimeters, mm) scaled proportionally to the contact force (F_o, in newtons, N) exerted by the user’s virtual pen on the circular palette, as shown in Equation 1:

where k denotes a scaling constant (k = 3 mm/N). This value was empirically determined to ensure that the dot’s size varied within a perceptually noticeable but not overly intrusive range (approximately 2.0–10.0 mm) across the allowable force range of 0.5–3.17 N. A linear mapping was adopted due to its intuitiveness, allowing participants to quickly learn the relationship between applied force and visual feedback. This helped reduce the learning burden and enabled participants to focus on force output regulation. The width of the gray annular region varied nonlinearly over each 60-degree cycle. Participants were required to adjust their exerted force based on the dot’s position to align its trajectory with the target zone. The circular palette was established using sphere tree models with Solidworks (Dassault Systems Inc., United States) and 3D Studio Max (Autodesk Inc., United States). Real-time force feedback was implemented using a validated haptic rendering algorithm (Wang et al., 2013). Exerted forces were recorded at a 1,000 Hz sampling rate using the haptic device and custom scripts developed with Microsoft Foundation Classes (MFC).

The experimental procedure is illustrated in Figure 1b. Each participant completed three sessions of the force control task following a practice session. Each session lasted approximately 10 min, with a short break (1–2 min) to eliminate muscle fatigue. Participants were instructed to focus on the moving dot and modulate their exerted force to match the target zone’s width. At random intervals (40–50 s), a probe question appeared on the screen asking ‘What are you thinking about? Task or Something else?’ Participants rated their attentional states on a 0–100 scale, where 0 indicated being completely focused on the task and 100 indicated complete distraction (Kucyi et al., 2016). EEG and force data from the 3-s time window preceding each thought probe were extracted as individual trials for subsequent analyses.

During the continuous force control task, the exerted force data were recorded at a sampling rate of 1,000 Hz. Because the target force (F_target), corresponding to the width of the target zone, varied nonlinearly over a 3-s cycle (i.e., 60 degrees), the target force pattern in the 3 s preceding each probe was consistent across trials. As shown in Figure 2a, we calculated the force error within the 3-s period to assess the performance of force control for each trial. The force error (BE_NL) represented the relative difference between the output force (F_output) and the target force (F_target). It was computed at each sampling point and analyzed using sliding windows (1-s length, 90% overlap) to quantify the variation of force errors within each trial. BE_NL was computed as Equation 2:

where the value of BE_NL ranges from 0 to 1, with smaller values indicating better task performance. We also calculated the sum of BE_NL within each trial, termed BE_NL-trial, to assess the overall behavioral performance. Figure 2b shows examples of BE_NL-trial values for two trials, where a smaller value indicates better force control performance.

Additionally, based on the participants’ self-reported ratings, trials were sorted by quartiles for each subject. Specifically, trials were divided into four groups based on the quartiles of the rating scores. Trials with ratings in the lowest quartile (≤25%) were classified as task-focused (onT), and those in the highest quartile (≥75%) as mind-wandering (MW).

All the behavioral and EEG features were first compared between conditions. To identify significant effects, we employed appropriate statistical tests (t-tests or cluster-based permutation testing), and only these significant features were retained for subsequent classification. Specifically, for the behavioral force errors, data were first averaged across trials for each condition within subjects, and then paired-sample t-tests were performed to evaluate condition-related differences. For EEG features (i.e., spectral power, alpha-theta ratios, PLV, and MI), cluster-based permutation testing was adopted to evaluate condition-related differences. This nonparametric statistical approach controls the family-wise type I error rate that arises from multiple comparisons across electrodes and frequency bins by employing Monte Carlo randomization. Briefly, the data were shuffled (1,000 permutations) to estimate a null distribution of effect sizes based on cluster-level statistics—specifically, the sum of t-values with the same sign across adjacent electrodes, frequencies, or ratios. The cluster-corrected p-value was defined as the proportion of permuted datasets in which the cluster-level statistic exceeded that of the original data (cluster-defining threshold: p < 0.05). Statistically significant features and electrodes were then selected as the final feature subset and used as inputs for the classifier. We chose cluster-based permutation testing because this nonparametric method can effectively control false positives under multiple comparison corrections while accounting for the spatial and spectral contiguity of EEG data (Maris and Oostenveld, 2007). This approach yields interpretable clusters of features rather than isolated points, which is particularly appropriate for EEG and connectivity analyses (Pernet et al., 2015).

Using the identified features as inputs and the attentional state labels (onT or MW) as outputs, we trained the classifiers using the support vector machine (SVM) algorithm to decode attentional states for each trial. Although previous studies have employed various machine learning algorithms, such as decision trees (Tasika et al., 2020), random forests (Chen et al., 2022), and artificial neural networks (Hosseini and Guo, 2019), this study applied SVM due to its suitability for small-sized datasets with low-dimensional data. Moreover, most prior studies on attention decoding have reported superior classification performance using SVM (Jin et al., 2019, 2020; Dong et al., 2021; Jothiral et al., 2025). A radial basis function (RBF) was selected as the kernel function, and the default parameter settings in LIBSVM were applied (i.e., penalty parameter C = 1 and kernel parameter γ = 1/feature dimension). To ensure comparability across features, all input features were z-score standardized within each training fold before model fitting, and the same transformation was applied to the corresponding test fold.

Two cross-validation strategies were employed to evaluate the trained models:

Finally, we computed five commonly used metrics—accuracy, recall, precision, F1-score, and the area under the receiver operating characteristic curve (AUC)—to comprehensively evaluate the effectiveness of the selected features in classifying attentional states. The classification performance for various feature combinations was also assessed using these metrics.

To assess whether the off-task ratings reported by participants increased over time during the force control task, we performed a linear regression analysis on the z-scored ratings across trials for each participant. Figure 3a illustrates the rating scores of a representative participant across all trials (blue dots), along with the corresponding regression-fitted curve (red line). Figure 3b presents the group-averaged slope values of the regression-fitted curves. A one-tailed one-sample t-test confirmed that the slope was significantly greater than zero [t(8) = 2.53, p = 0.018], indicating a significant upward trend in self-reported off-task ratings over time. For the trial-level force error (i.e., BE_NL-trial), we performed a similar linear regression on the z-scored values across trials for each participant. Figure 3c shows the BE_NL-trial values of a representative participant with the corresponding regression-fitted curve, and Figure 3d summarizes the slope values for all participants. A one-tailed one-sample t-test on these slope values revealed a significant increasing trend in force error over trials (t(8) = 3.16, p = 0.007). The z-scored MW ratings and BE_NL-trial data for all participants are provided in Supplementary Figures 1, 2, respectively.

During the experiment, seven participants completed 36 trials and the other two participants completed 48 trials. As shown in Figure 3e, trials were classified based on rating quartiles: the trials in the lowest quartile (≤25%) were labeled as onT condition, and those in the highest quartile (≥75%) as MW condition. Consequently, two participants contributed 12 trials per condition, while the remaining seven participants contributed 9 trials per condition. Group-averaged force errors (BE_NL-trial) for the two conditions are presented in Figure 3f. Paired-sample t-tests revealed significantly higher force errors during the MW condition (0.24 ± 0.07) compared to the onT condition [0.17 ± 0.03; t(8) = 3.44, p = 0.004].

For the trials classified as onT and MW, power spectra were estimated at each electrode for every frequency point between 2 and 45 Hz in a step of 0.1 Hz. Paired-sample t-tests were conducted to assess the power differences between conditions, with t-values (onT minus MW) visualized as spatial-frequency topographies (Figure 4a). Cluster-based permutation tests were then applied, and t-values that survived the significance threshold (p-cluster <0.025) are shown in the lower panel of Figure 4a. The analysis revealed two significant clusters in the low alpha band (8–10 Hz). Cluster 1 was primarily distributed over the frontal region, encompassing electrodes FPz, Fz, FCz, FP1, AF7, AF3, F1, F3, F5, F7, FC1, FC3, FC5, FT7, FP2, AF8, AF4, F2, F4, F6, F8, FC2, FC4, and FC6. Cluster 2 was mainly distributed over the posterior region, involving electrodes POz, Oz, Pz, TP7, P7, P5, P3, PO7, PO5, PO3, TP8, CP6, P4, P6, P8, PO4, PO6, and PO8. Additionally, cluster-averaged power within the 8–10 Hz band was compared between the two conditions. As shown in Figure 4b, the cluster-averaged power during the MW condition was significantly higher than during the onT condition in both the frontal region (MW = 0.142 ± 0.079, onT = 0.068 ± 0.029; p < 0.001) and the posterior region (MW = 0.170 ± 0.089, onT = 0.074 ± 0.039; p < 0.001).

To further localize the cortical sources associated with attentional fluctuations during the force control task, we conducted source localization analysis using a beamformer algorithm implemented in FieldTrip (Oostenveld et al., 2011). For each participant, source activity estimates were obtained for both the onT and MW conditions within the 8–10 Hz band. Paired-sample t-tests were used to assess differences in neural activity between the two conditions, followed by cluster-based permutation testing (p-cluster <0.025). As shown in Figure 5, the average power in the left middle occipital gyrus (MNI: [−30–70 12]) was significantly higher during the MW condition (3.51 ± 0.39 dB) than during the onT condition [3.36 ± 0.40 dB; t(8) = 4.99, p = 0.0084].

Synchronization between neural oscillations at different frequencies has been proposed as a core mechanism for the coordination and integration of neural systems. Mathematically, when two oscillators with different frequencies form a harmonic relationship (e.g., f₁/f₂ = 2), as opposed to a nonharmonic relationship (e.g., f₁/f₂ = 1.6), the harmonic arrangement allows for more frequent excitatory phase meetings, thereby facilitating cross-frequency synchronization. In line with this principle, recent theoretical frameworks suggest that shifts in oscillatory peak frequencies constitute a principal mechanism for implementing cross-frequency coupling and decoupling in the brain (Rodriguez-Larios and Alaerts, 2019, 2021; Rodriguez-Larios et al., 2020). Following this framework, we quantified cross-frequency coupling by analyzing peak frequency ratios between different bands and compared them between onT and MW conditions during the force control task. Given the established role of alpha-theta coupling in tasks involving attention and executive control, we specifically examined peak frequency ratios between the alpha and theta bands. Figure 6a illustrates trial-wise variability in alpha and theta peak frequencies, as well as their corresponding numerical ratios over a 3-s period for a representative participant and electrode. Figure 6b compares the spectral power in two representative trials: one showing a harmonic alpha-theta ratio (2.04) and the other a non-harmonic ratio (1.62). Figure 6c shows the distribution of alpha and theta peak frequencies across all trials for the same participant and electrode.

Differences between onT and MW conditions were further assessed using paired-sample t-tests for each electrode and cross-frequency ratio (ranging from 1.1 to 3.3 in a step of 0.1). Cluster-based permutation testing was applied to identify significant clusters based on adjacency in electrode space and cross-frequency ratio, while controlling for multiple comparisons. As described in Section 2.4.2.2, the distribution of alpha-theta ratios was represented by its probability density, where higher density at a given ratio reflects a greater likelihood of oscillatory alignment around that harmonic relationship. Figure 7a shows the probability density of each ratio averaged across all electrodes for the onT and MW conditions. Figure 7b visualizes the condition differences in probability density by plotting t-values for each ratio and electrode. Positive t-values (shown in cold colors) indicate higher probability density during the onT condition compared to the MW condition. A negative cluster (i.e., MW > onT) was observed at posterior electrodes within the lower ratio range (1.2–1.7), although it did not reach significance in the permutation test. However, a significant positive cluster was identified within the 2.6–3.0 ratio range, indicating a significantly higher probability density during the onT condition compared to the MW condition (p-cluster < 0.05). Figure 7c presents the topographical heat map of t-values averaged over the 2.6–3.0 ratio range, with PO3 and PO4 marked as significant electrodes from the permutation test. A repeated one-way ANOVA on the identified cluster confirmed that the cluster-averaged probability density within the 2.6–3.0 ratio range was significantly higher during the onT condition (0.299 ± 0.066) than during the MW condition [0.226 ± 0.065; F_(1,92) = 8.21, p = 0.0005; Figure 7d].

Given the significant condition-related differences observed in the 8–10 Hz band (Section 3.2), we analyzed PLVs between all electrode pairs to investigate functional connectivity differences within this frequency range. Paired-sample t-tests with permutation-based correction revealed no significant differences for the onT > MW comparison, but significantly higher connectivity during MW compared to onT at specific connections. The statistically significant PLV values and their corresponding electrode pairs are shown in Figure 8a.

To investigate condition-related connectivity differences from a network-level perspective, electrodes were grouped into four brain regions (Figure 8b): frontal lobe (F), parieto-occipital lobe (PO), left central motor area (LC), and right central motor area (RC). Community connectivity (CC) was computed by averaging PLVs within or between these defined regions. A repeated-measures one-way ANOVA revealed significantly higher intra-region CC within the LC during the MW condition compared to the onT condition [F_(1,100) = 19.76, p < 0.001]. Inter-region CC between PO-RC [F_(1,100) = 9.36, p = 0.003], F-LC [F_(1,100) = 6.84, p = 0.010], and LC-RC [F_(1,100) = 6.85, p = 0.010] was also significantly higher during the MW condition than during the onT condition. Descriptive statistics and corresponding significance values are summarized in Table 1.

Mutual information (MI) between the within-trial force error (BE_NL) and EEG power amplitude was analyzed to assess neural-behavioral synchronization. Figure 9a presents paired-sample t-test values for each electrode and frequency point (2–45 Hz in 0.1 Hz steps), where positive t-values (shown in cold colors) indicate higher MI during the onT condition compared to the MW condition. Cluster-based permutation testing identified a significant negative cluster within the 7.2–8.8 Hz range (p-cluster < 0.05), indicating significantly stronger neural-behavioral synchronization during the onT condition compared to the MW condition. This negative cluster was located in the anterior region, including electrodes F1, F3, F5, F2, F4, and F6. Figure 9b shows the topographical t-value map averaged across these six electrodes within the 7.2–8.8 Hz range. A repeated-measures one-way ANOVA on the identified cluster further confirmed significantly higher MI during the onT condition (0.650 ± 0.085) compared to the MW condition [0.530 ± 0.054; F_(1,92) = 10.69, p = 0.0015; Figure 9c].

An SVM-RBF model was trained to classify binary attentional states and evaluated using two cross-validation strategies (i.e., LOSO and 5-fold cross-validation). Each feature that passed the aforementioned statistical tests was evaluated individually in separate classification models. Table 2 summarizes the classification performance metrics for each individual feature under both validation strategies. The chance-level accuracy is 50% because the dataset was balanced using the quartile-based labeling approach. Here, MW was defined as the positive class. The single-feature models showed generally comparable classification performance across both validation strategies. Among the four features, spectral power yielded the highest accuracy, precision, F1-score, and AUC. Considering comparability with existing literature, this study primarily focuses on the accuracy and AUC metrics. For the cross-participant classification (i.e., LOSO strategy), the power feature achieved a mean accuracy of 66.99 ± 12.34% and a mean AUC of 76.37 ± 19.26%. Similarly, in the within-participant classification (i.e., 5-fold strategy), the power feature yielded a mean accuracy of 68.28 ± 7.14% and a mean AUC of 71.86 ± 7.05%. For the MI-only model, the recall for the MW class was high, whereas overall accuracy, precision, and AUC remained low. This indicates a bias toward predicting MW trials, leading to numerous false positives for onT trials, as illustrated by the confusion matrices shown in Supplementary Figure 3.

To evaluate whether synchronization features (i.e., ratio, PLV, and MI) could enhance the classification performance, we combined the spectral power feature with these synchronization features as inputs to the SVM-RBF models. Table 3 presents the classification results from all possible combinations of spectral power and synchronization-related features. Overall, combining spectral power with any of the synchronization features improved classification accuracy and AUC. Under LOSO cross-validation, the combination of spectral power, ratio, and PLV achieved the highest accuracy (71.57 ± 10.79%) and AUC (79.87 ± 15.59%). In 5-fold cross-validation, the combination of all four features yielded the highest accuracy (75.53 ± 8.40%), while the combination of spectral power, ratio, and PLV achieved the highest AUC (76.58 ± 7.07%).

The proposed continuous force control task effectively induced MW episodes, as evidenced by a significant upward trend in off-task ratings over trials and a marked degradation in behavioral performance (i.e., increased BE_NL-trial errors) during the MW condition compared to the onT condition. In terms of neural activity, we observed increased alpha power (8–10 Hz) over frontal and parieto-occipital regions during MW, with source localization revealing the increased activity in the left middle occipital gyrus. This widespread alpha power increase during MW aligns with the “perceptual decoupling” hypothesis (Schooler et al., 2011; Smallwood, 2013) and numerous EEG studies using visual vigilance tasks such as the SART (Compton et al., 2019; Jin et al., 2019). Notably, our findings are consistent with recent observations by Luna et al. (2023), who reported increased alpha power in left occipital regions prior to missed targets compared to correct detections in a visual vigilance task, suggesting that alpha increases may reflect attentional disengagement across both visual and visuo-motor tasks. Regarding the haptic domain, while Peng et al. (2021) reported increased frontal-central alpha associated with off-task states in a discrete force task, we observed a more distributed pattern involving both frontal and posterior areas. The localized increase in the left middle occipital gyrus indicates that even in a force-focused task, MW involves disengagement of visual processing regions, likely reflecting the visuo-haptic integration demands of our paradigm, since force adjustments in our task were guided by visual feedback. These findings highlight the modality-independent nature of alpha increases as a potential neural marker of attentional disengagement.

Beyond single-frequency analyses, we evaluated cross-frequency coupling and found significantly reduced probability density of high alpha-theta ratios (2.6–3.0) under the MW state, particularly over parietal-occipital electrodes PO3 and PO4. While prior work by Rodriguez-Larios and Alaerts (2019, 2021) and Rodriguez-Larios et al. (2020) linked increased alpha-theta phase synchrony to MW during meditation, our focus on harmonic frequency ratios reveals a different aspect of cross-frequency organization. Harmonic ratios near 3.0 may support on-task attention by enabling stable cross-frequency phase coupling. This precise phase alignment likely optimizes communication between neural assemblies supporting top-down control (theta) and those involved in sensory inhibition (alpha), thereby facilitating efficient information integration. The reduction of these stable harmonic ratios during MW suggests a breakdown in this coordinated cross-frequency mechanism. This view aligns with theories proposing that optimal cognitive control relies on harmonic cross-frequency arrangements enabling effective communication between neural assemblies (Fries, 2005; Palva and Palva, 2017). The parietal-occipital localization (PO3/PO4) further implicates visuospatial processing networks in maintaining this rhythmic coordination during force tracking. Nevertheless, whether the near 3.0 ratios observed here also support focused attention in other motor tasks or in other modalities (such as pure visual or auditory tasks) remains to be determined by future studies. Direct cross-task comparisons and source-level CFC analyses would be required to assess the generalizability.

Additionally, MW was associated with enhanced functional connectivity across sensorimotor networks. Within the 8–10 Hz band, MW states exhibited significantly stronger phase locking within the left central motor area and between the PO-RC, F-LC, and LC-RC communities. One plausible interpretation is the compensatory recruitment of task-relevant sensorimotor assemblies when top-down control wanes; such localized synchronization could transiently support baseline performance despite attentional lapses. This finding extends previous fMRI research highlighting the dominance of the DMN during mind wandering by suggesting a context-dependent compensatory mechanism: in the force control task requiring continuous sensorimotor engagement, attenuated top-down attention may trigger localized synchronization within task-relevant networks to sustain baseline performance (Christoff et al., 2009). This aligns with recent findings demonstrating a dynamically interdependent relationship between external (sensory and motor processing) and internal cognition (mind wandering) (Long et al., 2025). In accordance with the recently proposed “Baseline model of internal and external cognition” (Northoff et al., 2022), the observed hyperconnectivity likely reflects inefficient neural resource reallocation—where heightened “noisy” processing in sensorimotor circuits fails to fully compensate for attentional lapses—as evidenced by concurrent increases in behavioral errors. Our results underscore that MW dynamically redistributes neural resources with sensorimotor synchronization representing a signature of embodied attentional fluctuations. Nevertheless, alternative explanations for increased sensor-level connectivity—such as contamination by muscle activity, volume conduction, field spread, or other recording artifacts—cannot be ruled out (Haufe et al., 2012). Importantly, our connectivity estimates were derived at the sensor level without full source-level leakage correction; therefore, these results should be interpreted with caution as preliminary evidence. Future studies should complement sensor-level FC with source reconstruction, leakage-robust metrics, and muscle activity monitoring to better distinguish neural coupling from confounds.

In addition to neural synchronization metrics, we also examined the coupling between neural activity and behavioral performance by assessing the mutual information between the force error (i.e., BE_NL) and EEG power. During MW, the MI was significantly reduced within the 7.2–8.8 Hz band over several frontal electrodes, suggesting a breakdown in the real-time coupling between brain dynamics and motor output during attentional lapses. This observation is conceptually novel. Although previous studies have linked behavioral variability (e.g., RT variability) to MW (Esterman et al., 2013; Peng et al., 2021), none have quantified the dynamic synchronization between continuous neural signals and high-frequency motor performance. The 7.2–8.8 Hz band overlaps with the low-alpha/mu rhythm, known to reflect motor cortical excitability and somatosensory processing. Reduced NBS likely reflects a weakened predictive relationship between fluctuations in this rhythm and moment-to-moment force control accuracy weakens during MW. This decoupling might provide an objective neurobehavioral signature of attentional disengagement specific to active motor tasks and represents an advance beyond static behavioral error measures.

The SVM classification achieved optimal performance when traditional power features were combined with synchronization metrics. Within-participant models (5-fold CV) using all features reached 75.53% accuracy (AUC = 76.12%), while cross-participant models (LOSO CV) using power + alpha-theta ratio + FC features achieved 71.57% accuracy (AUC = 79.87%). These results are comparable to or slightly better than performance reported in similar binary classification studies (Jin et al., 2019, 2020; Dong et al., 2021; Chen et al., 2022). Several factors likely contributed to the classification performance. First, the combination of commonly used features (i.e., power) and novel synchronization features (cross-frequency ratio, FC, NBS) offered complementary information. Notably, adding synchronization features consistently boosted performance over power features alone (e.g., LOSO AUC increased from 76.37 to 79.87% with power + ratio + FC), highlighting the value of capturing distributed network dynamics and brain-behavior interactions. Second, the proposed continuous force task provided a rich stream of behavioral data (1,000 Hz) tightly synchronized with EEG data. This enabled the calculation of NBS, a feature unavailable in discrete response tasks and proved to be a useful feature for classification. Third, the continuous, dynamic nature of the force control task likely elicited more pronounced and ecologically valid MW states compared to simpler vigilance tasks, leading to clearer neural dissociations.

However, direct comparisons are challenging due to differences in tasks, probing methods, and classification approaches (e.g., LOPOCV vs. LOSO vs. within-participant). Our cross-participant accuracy (71.57%) highlights the challenge of generalizing models across individuals, a common limitation in the field. Future work should explore more advanced normalization or domain adaptation techniques.

While this study offers novel insights, the findings should be considered with several limitations. First, the final sample size of nine participants (after artifact rejection) is relatively small and all participants were male. Future studies should recruit larger, more diverse cohorts, including females and individuals from broader age ranges, to enhance generalizability and to explore potential sex differences in neural correlates of MW during motor tasks. Second, while synchronization features improved classification, their neurobiological interpretation remains complex. For instance, increased FC within motor areas during MW could reflect different processes (e.g., maladaptive noise, inefficient compensation, or muscle artifacts). Given the sensor-level nature of our connectivity analyses, we emphasize the need for source-space reconstructions and leakage-robust measures in future work to distinguish true inter-region interactions from sensor-level mixing. Future research should also integrate computational modeling or causal interventions (e.g., transcranial magnetic stimulation) to elucidate the functional significance of these connectivity patterns. Third, EEG epochs were extracted from the 3 s preceding thought-probes, which occurred at 40–50 s intervals. This quasi-periodicity could have induced anticipatory effects or strategic attention re-engagement just before probes, potentially affecting the MW vs. onT contrast. However, the overall monotonic increase in MW reports across trials, coupled with the corresponding behavioral degradation, suggests that anticipatory effects alone are unlikely to fully account for our main findings. Nevertheless, future studies are necessary to employ randomized probe timing to eliminate this potential confound.

Finally, our findings pertain to a specific visuo-haptic force control task. The generalizability of the identified synchronization signatures to other haptic tasks or modalities remains to be investigated. Future studies should systematically compare MW signatures across different task types, such as pure haptic vs. visuo-haptic or force vs. texture discrimination. Future studies should also explore more sophisticated synchronization measures, such as directed connectivity (Tafreshi et al., 2019), graph theory metrics (Moon et al., 2020), and advanced machine learning algorithms that may better capture complex spatiotemporal dynamics in EEG data.