Participants were seated calmly in front of a computer screen 80 cm from their eyes and asked to focus on a central fixation point displayed on a white background on the screen. The task consisted of a 2-min eye-open scan followed by a 2-min eye-closed wakefulness scan, each prompted by spoken and written instructions.

Participants were instructed to name an object shown as a line drawing on the computer screen as soon as they heard a sound signal, which occurred 4 s after the image appeared. Before the sound signal, participants were asked to simply view the image without responding. Each trial allowed a 20-s window for participants to respond. The next image was displayed only after the participant’s response, with the examiner manually controlling the presentation. The images used were taken from the Chinese version of the Boston Naming Test (Cheung et al., 2004). The task included 5 practice trials and 30 actual trials.

Participants were asked to name as many single words as possible within a given semantic category. The categories included both high- (e.g., land animals) and low-frequency (e.g., birds) conditions and were prompted by spoken and written instructions. Participants were given 1 min to respond for each trial. The task consisted of 1 practice trial and 4 actual trials.

Participants were instructed to spontaneously produce monologues based on 7 discourse prompts selected from the Cantonese AphasiaBank (Kong and Law, 2019). These included two well-known narratives, two single-picture descriptions, two sequential-picture descriptions, and one procedural discourse. Participants were allowed unlimited time during the preparation stage to review the stimulus before beginning their monologue. Once the monologue began, a 3-min time limit was applied for each discourse task. All picture stimuli were displayed on the screen, with the examiner managing the timing of their presentation.

All 15 participants from phase 1 underwent an EEG session utilizing the tasks and stimuli detailed in the “Section 2.1 Materials.” EEG data was preprocessed and segmented into time series for connectivity estimation. Network-based statistic (NBS) was applied to identify aphasia-related dysfunctomes. These results informed restoration-based and enhancement-based individualized stimulation targets for each participant, to be used in the phase 2 clinical trial.

All pre-processing procedures were performed using in-house routines and the EEGLAB toolbox (Delorme and Makeig, 2004; RRID:SCR_007292) on MATLAB version 9.14 (The MathWorks Inc, 2022; RRID:SCR_001622). The raw EEG signals were subjected to a bandpass filtering process with cut-off frequencies of 0.1 and 40 Hz. Bad channels were identified through visual inspection and corrected using spherical interpolation (Perrin et al., 1987). The data was then down-sampled to 250 Hz. To correct artifacts, an independent component analysis was conducted. Component classification was performed using the ICLabel plugin from EEGLAB, rejecting components labeled as “eye,” “muscle,” or “channel noise” with a probability greater than 50%. On top of that, we also rejected components showing higher-than-majority correlation with the EMG signals recorded from the lip muscles (Porcaro et al., 2015). The cleaned data were then reconstituted into channel space and underwent a final visual inspection to ensure quality. Subsequently, the signals were converted to an average reference. EEG signals were filtered into different frequency bands, including delta (1–4 Hz), theta (4–8 Hz), alpha (8–12 Hz), beta (15–30 Hz), and low-gamma (31–40 Hz) ranges. The EEG time series of each frequency band was then obtained by segmenting the signals into 4-s intervals. Time series with amplitude exceeding ±50 μV were rejected.

Connectivity matrices were constructed in sensor space, in which each electrode channel was defined as a node, and the connectivity between each pair of electrodes was defined as an edge. Phase Lag Index (PLI) was used as the connectivity estimator, which ranges from 0 to 1 and is robust against the effects of volume conduction (Peraza et al., 2012; Hardmeier et al., 2014). The calculation of the PLI was performed using in-house MATLAB code. A 61 × 61 connectivity matrix was constructed within each frequency band in each EEG task for each individual, resulting in a total of 20 sets connectivity matrices (5 frequency bands × 4 EEG tasks) for exploration in subsequent network-based statistical analyses.

All 20 connectivity matrices were included in the exploratory analysis using the Network-Based Statistic toolbox in MATLAB (Zalesky et al., 2010; RRID:SCR_002454). The goal of this analysis was to identify sub-networks comprising interconnected edges significantly predict aphasia severity (i.e., aphasia dysfunctome), as measured by aphasia quotient (AQ). As illustrated in Figure 2, the process began by performing mass univariate regression models on all edges to generate a single “statistic matrix,” where each element of the matrix represented the t-value of the corresponding edge. A primary threshold (a t-value threshold) was then applied to identify edges that showed a significant association with the AQ. This resulted in a binary statistic matrix, where edges exceeding the primary threshold were assigned a value of “1,” and all other edges were assigned a value of “0.”

Using this binary statistic matrix, network components were formed by linking all interconnected edges. The size of the largest network component was quantified by calculating the number of edges involved in the component. According to the assumptions of the NBS, cognitive processes rely on interconnected sub-networks serving as neural substrates, rather than on isolated connections distributed across the brain. Consequently, individual edges that were detached from the largest network component were considered less critical in supporting the cognitive process of interest.

Next, to evaluate the statistical significance of the observed network component, a permutation test was performed. In this test, the data labels were randomly shuffled across participants, and the size of the largest network component was recalculated for each permutation. This process was repeated 10,000 times to generate a null distribution of component sizes that could arise by chance. The observed network component was deemed statistically significant if its size exceeded the 95th percentile of the null distribution, corresponding to a p-value below 0.05.

The choice of primary threshold is considered arbitrary and open to exploration. In particular, applying a higher primary threshold is expected to extract smaller network components, and vice versa. There is no definitive rule for selecting primary threshold, as different cognitive processes may involve varying degree of focality in their neural substrates. A stringent threshold is appropriate when the effect is expected to be strong but localized, whereas a less stringent threshold is better suited for detecting weaker effects that are distributed across broader regions. Since predicting the nature of these effects is inherently challenging, analyzing the data with a range of thresholds allowed for greater flexibility and insight into the network patterns (Fornito et al., 2016; Mehraram et al., 2023). Therefore, the above-described procedure was repeated with multiple primary thresholds, starting from 2 to 4 with increments of 0.2 per step.

Finally, we evaluated which significant network component identified via NBS had the highest correlation with AQ. For this purpose, the individual averaged connectivity within each significant network component was calculated and entered into a Pearson’s correlation analysis with AQ. The network component with the highest correlation coefficient was selected as the target dysfunctome.

The dysfunctome-based individualized targeting consists of three steps: (1) dysfunctome identification, (2) individual dysfunctome examination, and (3) edge prioritization (Figure 3). This method requires two input information, one is the target dysfunctome defined as a binary matrix derived from the network-based linear regression, another is the individual’s dysfunctome profile defined as a weighted matrix. This method outputs a single edge which serves as the stimulation target for dual-site in-phase tACS with the stimulating electrodes positioned at the two nodes connected by the target edge.

After the exploratory analysis in phase 1, we identified an aphasia-predictive theta dysfunctome during divergent-naming task (see Section “3 Results” for details). The primary goal of the stimulation was to enhance the connectivity within this target dysfunctome. Given that the mechanism of dual-site in-phase tACS is to increase the connectivity between two stimulation sites, the individualized stimulation target was defined as a single edge within the target dysfunctome that was both “topologically important” and “individually relevant” based on two key factors: edge betweenness centrality and connectivity.

Edge betweenness centrality defines how an edge is considered to be topologically important based on the structure of the network. Particularly, it quantifies the importance of an edge based on the proportion of “shortest paths” that pass through it (Girvan and Newman, 2002). Edges with higher betweenness centrality are considered “bottlenecks” in the network which play a critical role in the flow of information of the network (Freeman, 1977). Edges with higher edge betweenness centrality in the target dysfunctome were prioritized for stimulation.

To determine which edges are individually relevant, the connectivity of the edges within an individual dysfunctome profile was considered. As discussed, it is indefinitive about whether stimulation should aim for restoring weak connections that were damaged from the stroke or further enhancing strong connections that showed functional importance in the reorganized network, therefore, two distinct targets were chosen based on connectivity. For restoration-based targeting, edges with lower connectivity were prioritized, while for enhancement-based targeting, edges with higher connectivity were prioritized.

To equally weigh the influence of both edge betweenness centrality and connectivity in the edge prioritization process, all edges within the target dysfunctome were ranked and normalized based on these two variables. This process resulted in a 0-to-1 value for both centrality ranking and connectivity ranking for each edge within the target dysfunctome. By averaging these two rankings, a priority index ranging from 0 to 1 was obtained for each edge, representing its priority for stimulation. The edge with the highest priority index was selected as the target for the dual-site in-phase tACS.

In phase 2, a single-session double-blinded sham-controlled trial was conducted with procedure illustrated in Figure 4. All participants underwent four 30-minute tACS stimulation sessions, with each session corresponding to one of the following stimulation conditions: (1) Restoration-based In-phase (RI), (2) Enhancement-based In-phase (EI), (3) Enhancement-based Anti-phase (EA), and (4) Sham (SH). During each session, a concurrent 30-min SLT was conducted by trained student clinicians. All student clinicians and participants were blinded to the stimulation condition. A 1-week washout period was implemented between sessions to minimize carryover effects, and the order of stimulation conditions was counterbalanced across participants. Five trained speech therapy students blinded to the experimental condition were responsible for carrying out all assessment, scoring, and the SLT. The first author was responsible for operating the stimulation machine and was not involved in the SLT and any outcome scoring.

Within each session, EEG scans were implemented immediately before and after stimulation to assess immediate effects of the intervention. The procedure and analytic method of EEG were consistent with those used in phase 1 but included only the divergent naming task (which derived the target dysfunctome). The same set of items was used for both pre- and post-intervention assessments within a session. To reduce practice effects, two distinct sets of items were alternated across sessions. After each session, a questionnaire was given to obtain participant’s subjective sensation, perception about being stimulated and adverse events associated with the stimulation. This questionnaire was adapted from the recommended template provided in the tES safety guidelines (Antal et al., 2017).

In all conditions, a dual-site 3 × 1 HD electrode montage was employed. This setup involved two stimulating electrodes positioned at the target nodes required for synchronization (i.e., the target edge), while three return electrodes were placed at the nearest channels surrounded each stimulating electrode (based on the configuration of the 64-channel EEG cap) to create a focal electrical field (Helfrich et al., 2014; Keator, 2022). The current intensity was set to 0.5 mA for both stimulating electrodes, and 0.17 mA for all returning electrodes, which had opposite polarity to the stimulating electrode they surrounded. In the in-phase conditions, both stimulating electrodes were set to the same polarity, whereas in the anti-phase condition, the two stimulating electrodes were assigned opposite polarities.

As we identified a theta (4–8 Hz) target dysfunctome in phase 1, the stimulation frequency for all conditions was set to 7 Hz. This frequency was chosen because previous studies suggested that a higher frontal-centric theta peak is associated with greater performance in cognitive tasks among healthy individuals (Moran et al., 2010).

The stimulation target of the sham condition followed enhancement-based target. During sham stimulation, the stimulator ramped up to the target intensity within the first 30 s of the session to mimic the stimulation sensation and then immediately ramped down to a negligible intensity for the remainder of the stimulation period. This procedure was repeated during the final minute of the session. This “fade-in, short-stimulation, fade-out” (FSF) protocol is the most commonly-used sham procedure across clinical trials and is considered effective for blinding when the stimulation intensity does not exceed 1 mA (Ambrus et al., 2012).

The 30-min SLT focused on speech production training. A two-phase repetitive training routine was conducted. The first phase consisted of a cueing-hierarchy naming training (McNeil et al., 1995), which utilized a series of increasingly less explicit cues to help the participant retrieve the correct name of an object depicted in a picture. Once the participant demonstrated retention of the trained target words, the training transitioned to the second part of the routine, which utilized the Response Elaboration Training (RET) (Kearns, 1985) approach. In this phase, the therapist asked the participant to describe a realistic, action-depicting picture that included a few target words trained in the first phase. Following the participant’s initial verbal response, the therapist provided a verbal model and asked follow-up wh-questions to prompt an expanded version of the sentence. This process was repeated multiple times until the participant was unable to further elaborate. Through this iterative method, the participant’s speech was shaped to become more detailed and complex.

All SLT stimuli were obtained from open-source picture banks available online. The target words were selected based on the individual participant’s ability level. One week prior to the stimulation sessions, a pre-treatment probe of 140 key words was administered to each participant to identify the most appropriate training items. Priority was given to words that the participant initially had difficulty producing but could successfully retrieve with cueing.

To investigate the effects of the stimulation on the target dysfunctome, the primary neurological outcome measures included the (1) connectivity of the target edge (estimated by phase lag index), (2) target node strength (calculated as the average node strength of the two nodes connected by the target edge), and (3) average connectivity within the target dysfunctome. Additionally, to determine whether the stimulation effects specifically modulated the target connections without influencing other connections, changes in the normalized connectivity of the target edge, normalized target node strength, and the average normalized connectivity within the target dysfunctome were also analyzed.

In addition to the connectivity changes within the target dysfunctome, we also explored the changes in global network properties during the divergent-naming EEG (the task-dependent EEG that derived the target dysfunctome). This included clustering coefficients, global efficiency, small-worldness, and modularity. These global network metrics represent important characteristics of the brain’s overall network organization and efficiency (Bullmore and Sporns, 2009; Ismail and Karwowski, 2020). The clustering coefficient reflects the tendency of nodes to form tightly connected groups, which is indicative of local specialization (Kaiser, 2008). Global efficiency measures the brain’s ability to integrate information across distant regions, serving as a marker of functional integration (Bullmore and Sporns, 2012). Small-worldness captures the balance between local specialization and global integration, a hallmark of efficient brain networks (Bullmore and Sporns, 2012). While modularity quantifies the degree to which the network is organized into distinct communities or modules that support specialized processing (Newman and Girvan, 2004). By examining these network properties, we aimed to assess whether the treatment not only influenced the target dysfunctome but also induced broader changes in the brain’s functional organization during a language task. All global network metrics were evaluated in unweighted connectivity matrix with a density threshold of 35% as binarization criteria. This threshold was determined based on prior exploratory analysis of multiple density thresholds ranging from 5% to 50% with 5% increments. The density threshold of 35% captured the most significant post-treatment differences in global network metrics across conditions. These measures were calculated by Brain Connectivity Toolbox (Rubinov and Sporns, 2010; RRID:SCR_004841) in MATLAB.

Apart from neurological changes, it was also important to investigate whether changes in the neural network corresponded to language facilitation. To this end, we measured the average number of hyponyms generated per minute performed during the divergent-naming EEG.