Patients were asked to control an online two-class BCI by means of a voluntary CVSA-orienting task (Figure 1A). Each trial started with a fixation cross (size 3.12°) in the middle of the screen and a random image of real object (4.8°) at the bottom left (neglected side). After 3000 ms, patients were instructed by a cue (300 ms) to covertly focus their attention to the left target (attention condition) or to keep fixating the center of the screen (rest condition). After a random time (3000–4000 ms), the target was outlined in red. In case of left target selection, the image started moving toward the center. Patients were required to press a button with their right hand as soon as they perceived the target selection.

Each patient performed six recording sessions within two consecutive weeks. In average, each session consisted of 2.7 ± 0.8 runs and each run of 30 trials randomly shuffled between attention (20) and rest (10) condition (Supplementary Table 3). The first two sessions were devoted to the calibration of the BCI with a positive feedback always delivered to patients; in the following runs (online modality) the image selected at the end of the trial was based on the output of the CVSA BCI, as real-time feedback for the patients.

Electroencephalography signals were acquired with a 64-channel system at 2048 Hz (BioSemi, Amsterdam, Netherlands). Electrodes were placed according to the standard international 10–20 system. Eye movements were recorded by means of three electrodes placed at the outer canthi of the eyes and at the gabella. The CVSA BCI was similar to our previous work (Tonin et al., 2012, 2013). The envelope of the EEG was extracted in seven α sub-bands (8–14 Hz, with 3 Hz of bandwidth) by means of Hilbert transform and a Laplacian filter was applied. Channels were pre-selected in the parieto-occipital regions (17 electrodes: P7–8, PO7–8, O1–2). Trial classification was based on data from the first 3000 ms after the cue. This period was split into windows of 150 ms. For each window, a quadratic discriminant analysis classifier was trained with the most discriminant features (frequency-channel pairs) selected during the calibration. In the online modality, classifiers were evaluated and the resulting posterior probabilities were integrated over time to deliver the final decision at the end of the trial.

We computed BCI performance as the percentage of trials that the BCI classified correctly during each online session. The average performance was 55 ± 3.9%, 60 ± 2.8% and 58.3 ± 3.7% (median and standard error) for patients P1, P2, and P3. Although patients’ BCI performance was low, it was above random for most sessions with an individual maximum accuracy of 76.6, 70, and 70% in their best runs. Furthermore, such a level of BCI accuracy is similar to that achieved by stroke patients during BCI-based motor rehabilitation (Prasad et al., 2010; Ramos-Murguialday et al., 2013; Pichiorri et al., 2015).

A posteriori analysis of ocular artifacts showed that horizontal eye movements were moderate for all three patients (9.4, 3.6, and 25% of the trials for P1, P2, and P3). Furthermore, the mean square contingency coefficient between the direction of horizontal eye movements and the CVSA tasks was not significant (P1: φ² = 0.53, p = 0.46; P2: φ²= 0.83, p = 0.36; P3: φ²= 0.29, p = 0.58).

Figures 1B,C report the analysis of the behavioral response to the button press across modalities (calibration or online) and over time for left CVSA task. Results showed a decrement of the reaction time (RT) for patients P2 and P3 (1.15 ± 0.15 s vs. 1.06 ± 0.17 s, p < 0.01 and 0.57 ± 0.20 s vs. 0.45 ± 0.10 s, p < 0.001). Patient P1 did not exhibit any change in RTs (0.50 ± 0.13 s vs. 0.50 ± 0.13 s; p = 0.72), however, he had the fastest initial RTs. A significant negative correlation between RTs and trial index was found for patients P2 and P3 (r = -0.42, p < 0.001 and r = -0.24, p < 0.01). The speedup in RTs cannot be explained as a consequence of behavioral training, since the duration of the trials was random.

We analyzed the spectral and spatial locations of the IAF in the parieto-occipital regions during the pre-cue interval for each patient, modality and hemisphere. During calibration the percentage of trials with missing α-peak was significantly higher in the affected hemisphere (1.73 ± 2.71% vs. 10.02 ± 7.68%, p < 0.01; mean and standard deviation for healthy vs. affected hemisphere). However, during the online modality, a reduction of this percentage seems to occur (1.73 ± 3.1% vs. 7.52 ± 7.90%, p < 0.01). This spatial asymmetry substantially differs from healthy populations, where a symmetric α-response is expected along the two hemispheres (Klimesch et al., 1998; de Munck et al., 2007).

A comparison of the related IAP is reported in Figure 2 (first row) for each modality and hemisphere. Statistical analysis showed an increment in the affected hemisphere for all patients during the online modality (statistically significant for P1 and P3: p < 0.00001; marginally significant for P2: p < 0.05 without correction). For patient P3 the increment was significant also in the healthy hemisphere.

Furthermore, the aforementioned restoration of the α-peak in the pre-cue interval seems to anticipate the decrement of the RTs in the correctly classified left trials (Figure 2, second row). In fact, patients P2 and P3 showed a significant negative correlation between the RTs and the IAP in the affected hemisphere (r = -016, p < 0.05 and r = -0.22, p < 0.01). This did not apply for patient P1 (r = -0.04, p = 0.61) but he/she was the only one who showed constantly low RTs over all runs. These results are in line with the literature, where it has been demonstrated that exists a positive correlation between the strength of α-power in the pre-stimulus interval and a faster RT (Klimesch et al., 1996; Nenert et al., 2012).

Feature discriminancy is a common BCI metric to assess the subject’s ability of modulating channel-frequency pairs during different mental tasks (Galán et al., 2007; Leeb et al., 2015). Herein, we investigated possible inter-hemispheric changes in the spatial distribution of discriminancy across the two experimental modalities (calibration and online).

Topographic maps in Figure 3 show the spatial distribution of discriminancy during calibration and online modalities (only parieto-occipital channels exploited in the online BCI are reported; normalized values are shown for comparison purposes; low discriminancy in blue, high in red). An initial imbalance of the discriminancy toward the healthy hemisphere (patients P1 and P3) and toward the affected one (patient P2) seems to be attenuated in the online modality. This hypothesis is supported by analysis on the difference of discriminancy between homotopic regions (Figure 3, second row). Negative values correspond to asymmetry in the modulation toward the left (healthy) hemisphere. A general asymmetry reduction (values toward zero) was reported for all patients in the parietal or the occipital regions with statistical significance for P1 and P3 (p < 0.05) and marginal significance for P2 (p < 0.05, without correction). Furthermore, for all patients, the asymmetry toward the healthy hemisphere significantly decreased across runs (P1, parietal regions: r = 0.59, p < 0.01; P2 and P3, occipital regions: r = 0.52, p < 0.05 and r = 0.51, p < 0.05) (Figure 3, third row).

Inter-patient differences might be explained by the fact that the most discriminant channels are strictly subject-dependent during CVSA tasks, as already reported in literature (Tonin et al., 2012, 2013). These results are in line with the rivalry hypothesis of SN (Kinsbourne, 1993) with the additional advantage of contingency with respect to the attention task performed by the patients.

Several functional Magnetic Resonance Imaging (fMRI) studies reported a relation between FC in large scale resting-state networks and SN (Corbetta et al., 2005; He et al., 2007). On the other hand, our within-patient connectivity analyses focused on investigating the FC during the attention task (driven by the CVSA BCI) and possible changes between calibration and online trials by exploiting the high spectro-temporal resolution of EEG signals. A general inter-patient trend seems to appear, highlighting an increment of inter-hemispheric FC for nodes in the affected (right) hemisphere and decrement for those in the healthy (left) one (Figure 4). Topographic plots show the difference between online and calibration FC maps computed with respect to each of the selected nodes (electrodes highlighted in magenta). Boxplots illustrate the difference in calibration and online modality between the FC of each node and the FC averaged across the whole opposite hemisphere. Only statistically significant differences are reported. For patient P2 the inter-hemispheric FC increment happened in the right-occipital node (p < 0.05) and for patient P3 in right-frontal (p < 0.001) and right-parietal nodes (p < 0.001). Contrarily, for patient P1, decrement of FC occurred in left-parietal (p < 0.01) and left-occipital nodes (p < 0.05) and, again, for patient P2 marginally in left-frontal node (p < 0.05, without correction). Changes in FC did not correlate with RT for any patient or node. Previous studies in resting-state networks demonstrated that the increasing inter-hemispheric connectivity (from right parietal regions to left hemisphere) is a signature of recovery from SN (Ramsey et al., 2016). The missing correlation between connectivity changes and RT might be due to the task-dependent nature of our analysis and, consequently, to the task-induced specific correlation patterns between different cortical regions (Zanto et al., 2011) as already suggested in (Baldassarre et al., 2014).