The tACS and cathodal tDCS stimulation was delivered with conductive-rubber electrodes placed in saline-soaked sponges and connected with a NeuroConn MC8 stimulator. The tACS stimulation electrode (5 × 7 cm) and a reference electrode (10 × 10 cm) were placed at Fpz and at the right upper arm, respectively, according to 10–20 system EEG recordings. Stimulation started with a 5-Hz burst, and then frequency increased in steps of 1–30 Hz using a 48-s-long “rtACS block.” The tACS stimulation was given for 20 min per day with a maximum current of 1.5 mA (peak-to-peak), which was well above the phosphene threshold (47). The block was repeated for 20 min. In the tDCS condition, the cathode was positioned above the intact hemisphere, and stimulation was done for 10 min immediately before rtACS and set at 1 mA using one electrode placed at either O1 or O2 position (3 × 3 cm) with anode at Fpz. The impedance was kept below 10 kΩ.

Sham patients had the identical electrode montage and stimulation duration. The tACS sham condition was designed to induce (short-lasting) phosphenes that patients could subjectively report (47); that is, it was a minimal stimulation. In addition, occasional current bursts were given to create short but presumably therapeutically ineffective phosphenes (45) involving one 5-Hz burst/min with individual amplitude for phosphene perception as used in a previous study where none of the subjects could tell to which group they belonged (45). In contrast, the tDCS sham-condition was designed to elicit only cutaneous sensations that gradually disappear because of habituation (48). Here, the current was ramped up for 30 s, then stopped, and at the end of the session ramped down for another 30 s as shown in Figure 1A (45). Through this design, we ensured that all patients felt their skin comparable in degree and duration with the active tDCS. The combined tDCS/tACS stimulation was designed to indicate whether prior cathodal tDCS on the intact hemisphere (a kind of conditioning) could enhance rtACS effects compared to sham stimulation and rtACS without preceding tDCS (45). Here, cathodal tDCS was applied to reduce the interhemispheric imbalance by inhibiting the visual cortex of the intact hemisphere (Figure 1B). All sham patients had been offered to receive stimulation treatment after the final follow-up evaluation. The stimulation parameters were kept unchanged for 20–30 min per day during the 2 weeks' treatment (Figure 1C). Of note: for all stimulation conditions, the default setting of the neuroConn stimulator gives short pulse of 50 Hz at 0.5 μAmp every 2 s to monitor the skin resistance.

The relatively large surface area of electrodes during stimulation limited the maximum threshold of current densities compared with other studies. The maximum current density was 42 μA/cm² below AC stimulating electrodes and 15 μA/cm² below the reference electrode. In the case of tDCS, it was 111 μA/cm² below the stimulating electrodes, which corresponded to a total charge density lower than 0.1 C/cm², which was below the safety limits as described in the previous study (22). Safety guidelines for direct current applied to the human brain were reported (16, 22, 49).

The following undesirable events had been observed immediately after each stimulation session and the following day before the next stimulation session: rare cases of headache, dizziness, fatigue/drowsiness, skin sensation, blurred vision immediately after stimulation, and others. Patients were not asked to perform a visual task during stimulation sessions but just kept their eyes closed while sitting down.

In the graph theory, the shortest path length is the short distance from one node to another node, which related to network efficiency and information transfer rate (61); the characteristic path length (CPL) is the average shortest path length of all nodes in the network with a definition:

where l_i is the average path length of between node i and all other nodes.

The clustering coefficient is the fraction of triangles around a node and is equivalent to the fraction of the node's neighbors that are neighbors of each other (63).

where C_i is the clustering coefficient of node i (C_i = 0 for K_i <2).

We performed a two-way 3 (stimulation group: ACDC, AC, and sham) × 3 (time: Pre, Post, and FU) mixed-design ANOVA with repeated measures on the time variable of local node strength and long coherence (Figure 2A). A compound symmetry assumption was checked before statistical analysis was performed. The regular p-value calculations in the repeated measures were reported if the theoretical distribution of the response variables was of the same variance. p-value calculations were corrected with Greenhouse-Geisser approximation. The post-hoc test was estimated with a significant sign (p < 0.05) after a mixed-design ANOVA test, and the family-wise error rate was controlled by the Tukey–Kramer test after estimating homogeneity of variances.

To explore the role of brain functional network reorganization as potential mechanisms of recovery, we calculated the two global parameters CPL and “global clustering coefficient” using a 30% threshold of the connectivity matrix (67).

There was no significant interaction effect on the alpha band in the occipital lobe.

A significant main effect of strength was observed in the ACDC group with repeated time measures on occipital_sup of the lesioned hemisphere [F_{(2, 42)} = 5.31, p = 0.009]. We ascertained that the assumption of sphericity was not violated (W = 0.97, p = 0.74). Thus, in the ACDC group, the strength of three treatment time points on occipital_ sup_LH differed significantly. Post-hoc analysis showed that FU strength on occipital_sup _LH was significantly higher than Pre (median ± SEM = 0.84 ± 0.32, p = 0.044).

The significant main effect of strength in Sham group was observed with repeated time measures on occipital_mid of the lesioned hemisphere [F_{(2, 42)} = 4.486, p = 0.017] and occipital_sup of lesioned hemisphere [F_{(2, 42)} = 5.31, p = 0.009]. The assumption of sphericity was not violated (W = 0.99, p = 0.98; and W = 0.97, p = 0.74, respectively). This shows that if we only consider the sham group's treatment, the strength of three time points on occipital_mid_LH and occipital_sup_LH significantly differed. Post-hoc analysis showed that the node strength of occipital_mid_LH after Sham treatment was significantly higher than before treatment (median ± SEM = 1.01 ± 0.42, p = 0.050) and follow-up (median ± SEM = 1.35 ± 0.39 p = 0.007). Moreover, the occipital_sup_LH node strength after Sham treatment was also observed to be significantly lower than follow-up (median ± SEM = 1.30 ± 0.41, p = 0.011) (Figure 2A, right part).

According to graph theory, a network structure can be characterized by two opposing poles: a high cluster coefficient with long path length (an “ordered” network) and a low clustering with short path length (a “random” network). If the network is in between those two poles, it has a proper balance between “stability” and “efficiency.” Then it is called a small world network. Patterns of anatomical connectivity in neuronal networks are sometimes characterized by high clustering and a small path length (63). We calculate the global CPL and CC using the 30% threshold of connectivity matrix as a criterion for three treatment groups to identify the small world network dynamic changes. Using these network parameters, we performed a two-way 3 (group: ACDC, AC, and sham) × 3 (time: Pre, Post, and FU) mixed-design ANOVA with repeated measurements on the time variable. The compound symmetry assumption was checked before statistics was performed. The regular p-value calculations in the repeated measures were reported if the theoretical distribution of the response variables had the same variance, provided the compound symmetry assumption was not violated; p-value calculations were corrected with the Greenhouse–Geisser approximation, and the post-hoc test was estimated with a significance level of p < 0.05 after a mixed-design ANOVA test. The family-wise error rate was controlled by the Tukey–Kramer test following estimation of the homogeneity of variances. No significance was observed for the global CPL and CC (Figure 2B). However, a trend was noted in that the global CC of ACDC was decreased in FU, whereas the global CPL remained at the same level as before, a clear sign of a more efficient small-worldness network after ACDC treatment.

The functional meaning of the network dynamics can be explored with correlation analyses. We found a negative correlation between the intact visual field and CPL (global CPL), which was significantly different at post (r = −0.80, p = 0.017) in the ACDC group (Figure 3A); this indicates that a larger visual field is associated with lower CPL after treatment. Furthermore, a positive correlation was observed between RT in areas of residual vision and CPL at Pre (r = 0.70, p = 0.049), suggesting that slower RT of the residual visual field was associated with higher CPL.

The correlation between brain network measures at subregions of occipital lobe and RT in intact VF shows a trend of a treatment effect as well. Node strength correlated negatively with both the intact and the lesioned hemisphere (Figure 3B), indicating that better visual function is supported by higher node strength in the brain. As for centrality, which is a measure to quantify how many shortest path length go through one node, the results show that after treatment and follow-up better visual function is supported by a state where brain has a higher capacity to compensate spontaneously.

The centrality of cuneus (r = −0.88, p < 0.01) and lingual (r = −0.73, p < 0.01) in the intact hemisphere was significantly correlated with the RT (Figure 3B), demonstrating that ACDC may have long-lasting modulation effect than the other two groups (see follow-up). ACDC enhances both hemispheres' brain connectivity, and especially, we could assume that “silent” neurons were reactivated, more functional connectivity could be rescheduled and transferred around the lesion part.

The connection coherence from lesion occipital and intact occipital to other brain regions was calculated, as shown in Figure 3C. A two-way mixed-design ANOVA test was conducted on brain network measures between group (ACDC, AC, sham) and time (before treatment: Pre, after treatment: Post, and: FU). The post-hoc analysis has been performed for the pairwise comparison with a significant level of p < 0.05; the family-wise error rate was adjusted.

A significant coherence was observed between the lesioned occipital (LH) and the intact occipital (IH) region in three measurements (F_{(2, 42)} = 6.509, p = 0.003), and the neural correlation in the delta band was significantly declined between the lesion occipital and intact occipital lobe after ACDC. Post-hoc analysis indicates that coherence after Post was lower than the Pre (MD ± SEM = −0.014 ± 0.005, p = 0.036), with a trend at FU (MD ± SEM = −0.016 ± 0.007, p = 0.069).

A significant coherence was also observed between intact occipital (IO) and intact temporal (IT) in three measurements (F_{(2, 42)} = 6.16, p = 0.004). The coherence between IO and IT was enhanced after Post and significantly declined at follow-up in the ACDC group in the low beta band. Post-hoc analysis indicates that the coherence of FU was lower than at Post (MD ± SEM = −0.017 ± 0.006, p = 0.018). And there was a trend of coherence enhancement after Post when compared with Pre-treatment (MD ± SEM = −0.018 ± 0.007, p = 0.054).

To compare the local node strength and centrality changes in responders with non-responders, we performed a two-way repeated ANOVA (groups: responder and non-responder, time: baseline and FU). In the lesioned hemisphere, the low alpha-band node strength in calcarine (F_{(1, 22)} = 6.42, p = 0.018) and lingual lobes (F_{(1, 22)} = 7.38, p = 0.012) was significantly different between responders and non-responders during follow-up. The post-hoc test showed in responders higher node strength in calcarine (MD ± SEM = −1.56 ± 0.57, p = 0.013) and lingual area (MD ± SEM = −1.68 ± 0.49, p = 0.002) (Figure 4C). In the intact hemisphere, both low alpha band node strength in calcarine (F_{(1, 22)} = 9.60, p = 0.005) and lingual lobes (F_{(1, 22)} = 5.76, p = 0.025) was significantly different between both groups during follow-up. Post-hoc, they were higher in responders for node strength in calcarine (MD ± SEM = −1.56 ± 0.65, p = 0.026) and lingual (MD ± SEM = −1.503 ± 0.427, p = 0.002). No significance was observed for brain network measure centrality. However, centrality of intact middle occipital at FU was lower than Pre in responders (p = 0.32, MD = −44).

Global network features are those that describe the state of the whole brain, irrespective of region. We first calculated the global clustering coefficient and global characteristic path length, followed by two-way repeated ANOVA (two groups/three time points). The p-value was corrected by the Tukey test in post-hoc analysis. The only significant finding was that global CC in FU was significantly lower than the Post (MD ± SEM = −0.0068 ± 0.0027, p = 0.05) in the high alpha band (Figure 5A). This suggests that responders need fewer connections to handle the neuronal synchronization in the resting state.

To investigate the long coherence fluctuation irrespective of the stimulation protocols, a two-way repeated-measure ANOVA was performed (three groups: control, responder, and non-responder, and two time points: Pre and FU). The p-value was adjusted for multiple comparisons by Tukey–Kramer test for post-hoc analysis.

A main effect on coherence was observed between intact occipital and lesion frontal in the alpha band (F_{(1, 45)} = 4.032, p = 0.05) (Figure 5B). Post-hoc, the FU coherence between the occipital_IH and frontal_LH was significantly lower than baseline in responders (median ± SEM = 0.0069 ± 0.003, p = 0.025), and an interaction effect was observed when investigating the difference between the control and responders during FU (F_{(2, 45)} = 4.04, p = 0.024) in the alpha band. Furthermore, at FU, the coherence between the occipital_IH and temporal_IH was significantly higher in non-responders when compared to controls (MD ± SEM = 0.0140 ± 0.0053, p = 0.030).

Pearson correlations were calculated to investigate the relationship between visual functionality measurements (FOV: Figure 6 and HRP: Figure 7) and node strength. We observed significant correlations both in the lingual_LH (r_resFU = −0.783, p = 0.007) and middle occipital_IH (r_resFU = −0.725, p = 0.018), where responders with higher FOV showed lower node strength in the delta band and higher node strength in non-responders in lingual_LH (r_resFU = 0.609, p = 0.021) and middle occipital_IH (r_resFU = 0.573, p = 0.032). This suggests that in both hemispheres, responders with higher FOV had lower delta band strength, whereas non-responders had higher delta band strength.

The same delta band pattern was also noted in both hemisphere of calcarine (lesion hemisphere: r_nonresFU = 0.608, p_nonresFU = 0.023, intact hemisphere: r_nonresFU = 0.539, p_nonresFU = 0.047) in non-responders. Delta band node strength was positively correlated with FOV in non-responders. This may be one possible reason why non-responders failed to improve their vision because of the delta band oscillation in visual cortex.

High alpha-band node strength correlated significantly with FOV in lingual_IH (r_nonresFU = −0.686, p = 0.004) and middle frontal_LH (r_nonresFU = −0.686, p = 0.007) in non-responders. This indicates that alpha-band node strength decreases with higher FOV measure in non-responders.

Cortical network reorganization after an injury is a widely recognized phenomenon (79). One study reported hemianopia patients with damage on the left primary visual cortex who showed greater activation on other lobes in the lesion hemisphere and intact hemisphere associated with the visual cortex (80). Another study showed that bilateral SMA activation was increased after intensive rehabilitation of postural balance (81). Yet, others suggested that plastic reorganization of cognitive resources serves to compensate for impairment in stroke patients during motor rehabilitation tasks (82). However, such studies of cortical balance and recovery using EEG were not carried out with hemianopic stroke patients. Most recently, another study reported that vision restoration training can improve visual field defects in chronic hemianopia and that this correlated with functional brain network reorganization as measured by MRI in precuneus, which may help quantify patient's ability to direct spatial attention (83). In the present study, we first applied the cathodal tDCS over the intact visual hemisphere with the goal to inhibit the increased excitability caused by brain network reorganization after the stroke, which was then immediately followed by tACS applied at Fpz to entrain oscillations for both hemisphere. In the ACDC group, the middle and superior occipital lobe of the lesion hemisphere had significantly higher strength in Post and FU compared to baseline and showed greater network strength in the intact middle occipital lobe. An enhancement of the lesioned hemisphere in the ACDC group was also observed in the superior occipital lobe. The strength of both AC and sham group fell back to slightly below the original level during follow-up. Both occipital lobes' enhanced strength could demonstrate that interhemispheric connections were more balanced in the ACDC group. Because this correlated with visual performance in the ACDC group, this observation is compatible with the hypothesis that postlesion FCN plasticity reduces the imbalance between the lesioned and the intact hemisphere. As we showed, the unique protocol design of ACDC is able to modulate brain plasticity between the lesion and intact hemisphere, where the continuous stimulation can produce sustained and long-lasting neuronal modulation of the brain's neural networks. However, in stroke patients, ACS alone is largely ineffective, except for a benefit of foveal sensitivity (35).

The strength of calcarine_IH increased in the ACDC group, whereas the other two groups did not show similar patterns. The same change was also observed in the responder group. The centrality of both calcarine_IH and lingual_IH positively correlated with the RT in Pre and correlated negatively with RT both in Post and RT. This shows that the ACDC modulation enhanced the efficiency of information transfer on these two brain regions, which is associated with recovery of visual function (faster RT).

Regarding the issue of global brain network modulation, a lower CC was observed only in the ACDC group compared to baseline, where ACDC reduced alpha-band whole-brain connections. Our interpretation is that reduced connections are a sign of greater efficiency of visual processing, where less connectivity (greater processing efficiency) could comprise a possible mechanism of visual recovery.

CPL is the average shortest path length in the network; a lower CPL indicates that fewer intermediary nodes are needed to transfer information between two unlinked nodes. In this case, the efficacy within a network is considered to be high. We found a significant positive correlation between the gray dots' RT and alpha-band CPL in the ACDC group at baseline, which disappeared at follow-up. This suggests that vision processing after ACDC modulation needs fewer nodes, i.e., fewer steps to process neural information. In the ACDC group, a significant negative correlation was observed between the number of white dots in HRP and CPL after treatment. This also suggests that ACDC decreased the whole-brain alpha-band CPL, and this was associated with an enlarged visual field. In contrast, CPL shows a very low negative correlation at FU. In summary, we suggest that ACDC can enhance processing efficiency in the alpha band, which also contributes to vision recovery.

The changes of the brain network in global coherence at both Post and FU show that NIBS can trigger brain plasticity by altering functional interaction between multiple brain regions. Specifically, ACDC reduced the functional connectivity between the lesion and the intact occipital lobe in the delta band and enhanced the connectivity between the intact occipital and the intact temporal lobe in the low beta band. In contrast, in the sham and ACS groups, no significant network changes were observed. Therefore, the combination of tACS and tDCS is apparently able to modulate neural plasticity by increasing the efficiency of communication between remote regions of the brain, possibly by improving interhemispheric balance.

The design of a proper sham condition is one of the biggest challenges in NIBS studies because NIBS can elicit cutaneous sensations that gradually disappear due to habituation, and tACS induces phosphene perception (47). We used 5-Hz burst/min current bursts rather than “no stimulation” in the Sham tACS group. The current level was ramped up for the 30 s, then stopped, and at the end of the session ramped down for another 30 s in the sham tDCS group (45). In this way, we ensured a comparable effect and duration of cutaneous sensations for all the patients during stimulation with this unique design. In the sham group, however, the strength of temporal_mid_IH increased significantly after 10 days of stimulation and at follow-up returned back toward baseline levels. This suggests that the sham condition was not neutral but altered the strength, which might mean that the temporal lobe is sensitive to slow burst current in sham tACS, although no long-lasting effect was observed. Yet, node strength in the occipital cortex did not change after ACS.

The total group of responders (i.e., irrespective to which group each patient belonged) showed significantly enhanced strength in the low alpha band in the lingual and calcarine lobe of both hemispheres, which non-responders did not. The lingual gyrus located between the middle of the temporal lobe and occipital lobe is relevant for complex visual processing such as object shape and contour information (86, 87). The calcarine sulcus is mainly involved in the primary visual cortex (V1) with a role in early-stage visual processing, creating a bottom-up saliency map from visual inputs to guide the shifts of attention (88, 89). The strength enhancement in both hemispheres could be a sign of compensation following the occipital damage. Similarly, the strength of the middle occipital region of the intact hemisphere was enhanced. We conclude that the reorganization occurs in two hemispheres symmetrically as a consequence of the occipital lobe. The correlation between network strength and behavior performance indicates that the delta band and alpha band play a vital role in vision recovery. Possibly regions with less alpha and higher delta are less responsive to the NIBS-induced oscillations, at least with regard to behavior output.

While in the delta band of the lingual and calcarine node non-responders had higher connection weights with higher FOV values, responders had fewer connection weights with higher FOV. Thus, delta band connectivity might play a critical role in enhancing visual functions and be a possible recovery biomarker of brain network reorganization. The same was noted in the intact middle occipital lobe; in responders, better vision was associated with lower delta band connectivity strength of the intact middle occipital, whereas the non-responders had higher node strength. The pattern is consistent with the neural correlation between the intact and lesion occipital lobe in the ACDC group: here, a lower coherence in the delta band between two occipital lobes was associated with visual field parameter improvement. Similarly, the global CC of the responder group in FU was significantly decreased in high alpha band compared to baseline. Thus, this finding also suggests that responders (with better vision recovery) needed fewer whole-brain connections in the high alpha band.

In non-responders, greater visual acuity correlated with lower strength and FOV in the alpha band, especially at the frontal and lingual region. We also observed that the correlation between intact occipital and lesioned frontal lobe was decreased in responders. Thus, a local and global pattern of decreased connections between the intact occipital cortex and the lesioned frontal cortex signal seems to be a physiological correlate of vision recovery, and it shows that the middle frontal lobe plays an important role as a visual information-processing bridge, which was weakened after ACDC treatment.

In responders, the FU coherence between intact occipital and intact temporal lobe was significantly higher than at baseline. It is known that the temporal lobe is responsible for handling perception and movement identification. Therefore, an enhanced connection between two lobes following NIBS suggests that the intact temporal lobe adjusts the internal information transmission state more rapidly. It may temporarily disengage connections with other regions that are less important, providing more support for the visual cortex to process visual movement information. Centrality of the intact temporal lobe demonstrates how much information is transferred through this area. In responders, we noticed a trend of local node enhancement in centrality and strength during FU compared with baseline and control. In responders, the centrality of Occipital_mid_IH remained unchanged compared to baseline, whereas centrality of Temporal_mid_IH increased. Considering the global coherence enhancement between the intact occipital and intact temporal regions, this suggests that the communication between the intact occipital and temporal lobe plays a critical role in visual function enhancement, and it is this enhancement that seems to be adaptive.

In summary, occipital damage following stroke creates partial vision loss with brain network alteration in the delta and alpha band, and ACDC, but not AC alone, improves visual functions. As we now showed, brain network plasticity patterns such as interhemispheric (im-) balance and long-lasting plasticity were consistent with behavior performance in the ACDC group. This demonstrates that modulating brain network plasticity is a promising tool to induce vision recovery. An analysis of responders vs. non-responders (irrespective of the treatment they received) also helped us understand NIBS effects, highlighting the role of a reduction of the coherence between the LH and IH occipital lobes in the delta band and a reduced high alpha-band coherence between the frontal and occipital lobe; these two FCN patterns might comprise biomarkers of vision recovery and shed light on the role of coherence between intact occipital and intact temporal regions. Future experiments are needed to confirm this proposal.