Dissociable effects of β-tACS on spectral power and network organization for proximal versus distal upper-limb recovery in subacute stroke: double-blind, randomized, sham-controlled trial with high-density EEG Authors: Xiao-Yu Liao, Yuxin Zhang, Yan Gong, Aisong Guo, Shangxiaoyue Li, Ailipinai Yasen, Surong Qian, Wenjun Qian, Renjie Xu, Guangxu Xu, Yu-Er Jiang, Yaxin Gao Published: 2026 ABSTRACT AbstractBackgroundUpper-limb recovery after stroke is not unitary, and proximal and distal functions may differ in their neural constraints. We tested whether 20-Hz β-tACS paired with rehabilitation differentially affects proximal versus distal recovery in subacute stroke and which EEG changes accompany improvement.MethodsIn this double-blind, randomized, sham-controlled trial, 61 first-ever stroke patients within 1–6 months after onset were enrolled; 56 (28 per group) completed the final analysis. Active stimulation consisted of 20-Hz tACS over ipsilesional M1 (2 mA peak-to-peak, 20 min/session, 10 sessions over 2 weeks) followed by standardized rehabilitation; sham used ramp-up/ramp-down only. FMA-UE total, proximal, and distal scores were assessed at baseline, post-intervention, and 1-month follow-up (F/U). High-density resting-state EEG quantified δ/β-band spectral power, connectivity, and graph metrics. Prespecified moderation analyses examined MEP status, lesion location, and stroke duration.ResultsCompared with healthy controls, stroke patients showed higher δ and θ power and lower β power at baseline. Relative to sham, β-tACS reduced whole-brain δ power (F = 4.60, q = 0.037), increased β power (F = 8.29, q = 0.009), and produced a δ-band network downshift, including lower network density (F = 6.73, q = 0.012), mean edge weight (F = 6.16, q = 0.010), clustering coefficient (F = 6.77, q = 0.012), and global efficiency (F = 6.87, q = 0.010). No FDR-significant β-band connectivity main effects were observed. Clinically, within the present 10-session protocol, β-tACS was associated with greater improvement in total FMA-UE at post (F = 7.24, p = 0.007) and F/U (F = 11.33, p = 0.001), with the clearest benefit in proximal FMA-UE at post (F = 11.17, p = 0.002) and F/U (F = 24.37, p < 0.001); distal improvement did not differ significantly between groups. MEP status moderated total and proximal recovery, with larger effects in MEP-negative participants. In the active group, larger total and proximal gains were associated with larger reductions in δ-band network organization; no other EEG–behavior correlations survived FDR correction.ConclusionsWithin the present 10-session protocol, 20-Hz β-tACS over ipsilesional M1 paired with standardized rehabilitation was associated with clearer proximal than distal upper-limb benefits in subacute stroke. Its physiological signature was characterized by β-band target engagement together with reduced δ power and δ-band network organization. Whether longer treatment courses or higher cumulative doses would produce additional distal gains requires further study.Trial registration The study was registered in the Chinese Clinical Trial Registry (ID: ChiCTR2400083526) on April 26, 2024. The registration submitted on March 14, 2024. The first participant enrolled on May 14, 2024.Graphical Abstract Supplementary InformationThe online version contains supplementary material available at 10.1186/s12984-026-02040-4. Introduction Stroke is one of the leading causes of long-term disability worldwide, with upper extremity motor impairments significantly compromising patients’ daily activities and quality of life [1]. Recovery of upper-limb function after stroke is therefore a central goal of neurorehabilitation [2]. However, upper-limb recovery is not a unitary process. Improvements in proximal control and distal dexterity often follow different trajectories, likely because they depend on partially distinct anatomical substrates and network organizations [3–5]. Neurophysiological studies indicate that stroke is accompanied by abnormal oscillatory activity within distributed motor networks [6, 7]. Resting-state EEG studies have consistently shown a shift toward pathological low-frequency activity, particularly in the δ range, together with disruption of higher-frequency sensorimotor rhythms, including β-band activity [7–9]. Reduced β power, altered β coherence, and disturbed interhemispheric β dynamics have all been linked to impaired sensorimotor processing and poorer motor performance after stroke [7, 10–13]. At the same time, increased δ activity has emerged as one of the most robust electrophysiological signatures of post-stroke dysfunction and has been associated with lesion burden, cortical deafferentation, altered large-scale communication, and recovery status [8, 9, 14]. Because β oscillations are closely related to motor planning, movement execution, and sensorimotor integration [6, 15–18], whereas δ-band abnormalities may reflect a maladaptive low-frequency dominant state [8, 9, 14], examining both β and δ rhythms may help clarify not only target engagement by stimulation but also whether treatment is associated with a broader normalization of post-stroke network dynamics [8, 13]. The neural control of the upper limb further suggests that proximal and distal recovery may not respond identically to neuromodulatory interventions. Proximal movements, such as those involving the shoulder and elbow, depend on more distributed and bilateral motor-network organization, engaging primary motor, premotor, and supplementary motor regions as well as multiple descending systems [3, 19, 20]. In contrast, distal hand and finger function depends more strongly on high-fidelity corticospinal output and the integrity of more selectively organized cortico-muscular pathways, which are critical for fractionated movement and dexterity [3, 4, 20]. Accordingly, proximal function may retain greater capacity for compensation through distributed network recruitment after stroke, whereas distal recovery may be more constrained by corticospinal tract integrity and may require more task-specific or precision-oriented interventions [4, 5]. This conceptual distinction provides a mechanistic basis for asking whether a network-level neuromodulatory intervention such as β-tACS may preferentially benefit proximal rather than distal motor recovery. Transcranial alternating current stimulation (tACS) is a non-invasive neuromodulation technique designed to interact with ongoing neural oscillations in a frequency-dependent manner [21–23]. Within the β range, 20 Hz has been widely used as a canonical sensorimotor stimulation frequency in human motor-cortex studies, providing a mechanistically interpretable and literature-grounded fixed-frequency protocol [24–26]. Experimental work suggests that β-range stimulation can influence oscillatory timing, corticospinal excitability, and plasticity-related processes, although its behavioral effects are clearly state and context-dependent [23–26]. In subacute stroke, where β activity is often attenuated and low-frequency dominance is enhanced [7–9], repeated 20-Hz stimulation delivered before behavioral training may act more as a priming intervention that biases the motor system toward a more functionally favorable oscillatory state. Such a priming account is consistent with the view that the effects of non-invasive stimulation depend critically on the physiological state of the targeted network and may be amplified when stimulation is paired with behaviorally relevant training [24, 25]. The present study was therefore conceived as a hybrid neuromodulation-plus-rehabilitation intervention. β-tACS was administered immediately before standardized physiotherapy and occupational therapy to test whether frequency-targeted modulation of the ipsilesional motor system could create a neural context more conducive to subsequent rehabilitation-driven plasticity. We selected ipsilesional M1 as the stimulation target because it is the most established target in motor-cortex neuromodulation, is closely linked to sensorimotor β oscillations and corticospinal output [13, 24]. Another important source of heterogeneity in post-stroke recovery is corticospinal functional reserve, commonly indexed by motor evoked potential (MEP) status. MEP status is a clinically relevant biomarker of motor recovery potential and may moderate responsiveness to rehabilitation and neuromodulation [27, 28]. Because distal dexterity may be more consistently linked to corticospinal tract integrity, whereas proximal actions may be supported to a greater extent by distributed compensatory network recruitment [4, 5, 25], MEP status offers a biologically meaningful framework for examining differential treatment response [27, 28]. We therefore prespecified MEP status as a moderator of interest to test whether baseline corticospinal reserve shapes the behavioral and neurophysiological effects of β-tACS. In this study, we aimed to evaluate the effects of β-tACS paired with standardized rehabilitation on upper limb motor recovery in individuals after stroke, using a randomized, double-blind, sham-controlled design. High-density resting-state EEG was used to quantify changes in spectral power and functional network organization, and motor outcomes were assessed using the Fugl-Meyer Assessment for the Upper Extremity (FMA-UE), including total, proximal, and distal scores. Our primary objective was to examine whether β-tACS improves overall upper-limb motor function. We further sought to examine whether treatment effects differ between proximal and distal domains within the dose and observation window tested, and to characterize the oscillatory and network-level EEG signatures associated with behavioral change. We hypothesized that β-tACS would modulate oscillations and related network dynamics and that these physiological changes would be associated with improvements in upper-limb motor outcomes. Methods Participants This double-blind, randomized, sham-controlled trial was conducted at the Rehabilitation Medicine Centre of Suzhou Municipal Hospital between May 14, 2024 and May 31, 2025. Eligible participants were those aged 18–80 years, experiencing their first-ever stroke, and within 1–6 months post-onset. All participants met the AHA/ASA tissue-based clinical definition of stroke, confirmed through CT or MRI imaging. Patients were excluded if limb dysfunction resulted from other neurological disorders, or if they had severe cognitive impairment, language or psychiatric disturbances, reduced consciousness precluding cooperation, or contraindications to MRI, TMS, or tACS (e.g., cardiac pacemaker, intracranial metal implants, skull defects). Individuals taking medications known to influence cortical excitability, such as tricyclic antidepressants, neuroleptics, or antiepileptic drugs, were also excluded. A priori, the sample size was based on an ANCOVA for the proximal FMA-UE subscore at 1 month follow-up (F/U), with baseline score as a covariate. Assuming a common SD of 6.5 points and a between-group difference of 4.0 points [29], a two-sided α = 0.05 and 80% power required 22 participants per group using the Borm-Fransen-Lemmens formula [30]. Written informed consent was obtained from each participant or a legal representative. The study was approved by the Ethics Committee of Suzhou Municipal Hospital and was registered with the Chinese Clinical Trial Registry (ChiCTR2400083526) on April 26, 2024. In addition, resting-state EEG data were collected from 21 age-matched healthy controls recruited contemporaneously for this study, all without a history of neurological or psychiatric disorders. Clinical baseline assessment and lesion characterization At baseline, demographic and clinical variables were recorded, including age, time since stroke onset (weeks), and stroke severity measured by the National Institutes of Health Stroke Scale (NIHSS). For exploratory moderator analyses, stroke duration was additionally categorized a priori as 1–3 months versus > 3–6 months post-onset. Stroke subtype (ischemic/hemorrhagic), lesion hemisphere (right/left), and lesion location were determined from routine clinical CT/MRI reports. Lesion location was categorized a priori as cortical, subcortical, or brainstem for exploratory moderator analyses, based on the primary anatomical involvement described by neuroradiology and confirmed by an experienced stroke neurologist. Because routine clinical imaging reports did not allow reliable lesion segmentation or quantification of corticospinal tract lesion load, this classification was intended to capture broad lesion category rather than fine-grained lesion topography. NIHSS was administered by trained clinicians using standardized procedures. MEP status was assessed at baseline as an index of corticospinal functional reserve and was prespecified as a moderator of treatment response [27]. Single-pulse TMS was delivered over the motor cortex contralateral to the paretic hand while surface electromyography was recorded from the paretic first dorsal interosseous. Resting motor threshold was determined whenever obtainable; otherwise, active motor threshold was used. Participants were classified as MEP-positive if suprathreshold responses were present and as MEP-negative if no resting MEP could be elicited despite testing up to maximum stimulator output. Study design This study was a double-blind, randomized, sham-controlled trial. An independent investigator performed group allocation using sealed, opaque envelopes to assign eligible participants to either the tACS or sham group. A second investigator, who was not involved in any assessment or treatment procedures, programmed the stimulation device accordingly. Participants, treating therapists, and outcome assessors remained blinded to group allocation throughout the study (Fig. 1a). Study timeline, EEG ROI definition, and β-tACS montage with electric-field modeling. a A double-blind randomized sham-controlled study timeline is shown. Resting-state EEG was recorded at Baseline and Post. Motor performance was assessed with FMA-UE proximal and distal subscores at Baseline, Post, and F/U. Participants were assigned to the tACS group (n = 28) or the sham group (n = 28) and completed 10 stimulation sessions during the treatment phase. b The stimulation montage and individualized electric-field simulations are. presented. Active stimulation delivered a 20 Hz sinusoidal current at 2 mA peak-to-peak for 20 min, with the anode placed over ipsilesional M1 and the cathode placed over the contralateral upper trapezius. Head-model renderings summarize the modeled current distribution and electric-field magnitude under the montage. c The 128-channel EEG layout and the a priori scalp ROI parcellation used for. spectral and connectivity analyses are illustrated, including ipsilesional and contralesional frontal (iF, cF), temporal (iT, cT), central (iC, cC), and posterior (iP, cP) regions Intervention Stimulation was delivered immediately before each rehabilitation session for 10 sessions over 2 weeks. This 10-session schedule was selected as a pragmatic early-phase fixed-dose protocol to examine short-term physiological target engagement and motor effects of 20-Hz β-tACS when paired with standardized rehabilitation. Given the heterogeneity of tACS protocols in stroke rehabilitation and the absence of an established dose–response standard, this regimen should be interpreted as an experimental dose within the tested treatment window [31]. Two 5 × 5 cm saline-soaked sponge electrodes were used. The stimulation electrode was centered over the ipsilesional primary motor cortex using the international 10–20 system (C3 for left-hemisphere lesions, C4 for right-hemisphere lesions), and the return electrode was placed over the contralateral upper trapezius (Fig. 1b). Active stimulation consisted of a 20-Hz sinusoidal current at 2 mA peak-to-peak for 20 min. Sham stimulation used the same schedule, with ramp-up and ramp-down only. Immediately after stimulation, all participants received the same standardized physiotherapy and occupational therapy program for 30–40 min. The rehabilitation program comprised therapist-guided upper-limb training modules drawn from routine clinical practice, including proximal control and postural stabilization, shoulder/elbow reaching and movement coordination, task-oriented upper-limb practice, hand-use and grasp–release training when feasible, and occupational-therapy exercises targeting functional task practice and activities of daily living. During F/U, home-based rehabilitation frequency was recorded and adverse events were monitored at each visit (Table S1). Outcome measures EEG acquisition Resting-state EEG (eyes open, central fixation, 5 min) was recorded in an electrically shielded room using a 128-channel HydroCel Geodesic Sensor Net (Philips EGI, USA) at a sampling rate of 1000 Hz (Fig. 1c). Preprocessing was performed in Python (MNE), including removal of the initial 30 s, 0.5–40 Hz band-pass filtering, identification and interpolation of bad channels, ICA-based removal of ocular, cardiac, and muscle artifacts, common-average re-referencing, 2-s epoching, and rejection of residual noisy epochs by visual inspection. For participants with right-hemisphere lesions, channel labels were mirrored to align lesion laterality across the dataset. Spectral power Power spectral density(PSD) was estimated using Welch’s method. Relative band power was calculated by dividing band-limited power by total power across 1–40 Hz. Based on a priori hypotheses, analyses focused on the δ (1–4 Hz) and β (13–30 Hz) bands. In addition to whole-brain estimates, electrodes were grouped into eight a priori scalp regions of interest (ROIs), comprising ipsilesional and contralesional frontal (iF, cF), temporal (iT, cT), central (iC, cC), and posterior (iP, cP) regions. The ROI scheme was based on the geometry of the HydroCel 128-channel net and was designed as spatially contiguous, left–right homologous clusters for hypothesis-driven regional summarization. Functional connectivity Frequency-specific functional connectivity was computed using the debiased weighted phase-lag index (dwPLI) [32] for the δ and β bands. Whole-brain mean connectivity was defined as the average of all off-diagonal channel-to-channel connections. ROI-level connectivity matrices were obtained by averaging channel-level connections within and between the eight predefined ROIs. Global graph-theoretical metrics included network density, mean edge weight, clustering coefficient, and global efficiency. Motor performance Upper-limb motor impairment was quantified using the FMA-UE, a standardized measure ranging from 0 to 66, with higher scores indicating superior motor function. The assessment is composed of a proximal subscale (maximum 36 points; assessing shoulder, elbow, and coordination) and a distal subscale (maximum 30 points; assessing wrist and hand function). To ensure comprehension and optimal performance, each task was first demonstrated and practiced on the unaffected limb. Participants then performed three trials with the affected limb, and the best score was recorded. Assessments were administered at three time points: 24 h prior to the first stimulation session (baseline), within 24 h following the final session (post), and at a one-month F/U. Statistical analysis All analyses were performed in Python 3.11. Baseline characteristics were compared using independent-samples t tests or Mann–Whitney U tests for continuous variables and χ2 or Fisher’s exact tests for categorical variables, as appropriate. Primary treatment effects on motor and EEG outcomes were evaluated at post-intervention and F/U using baseline-adjusted ANCOVA when model assumptions were met, or baseline-adjusted permutation ANCOVA (Freedman–Lane, 5000 permutations) otherwise. Effects are reported as adjusted mean differences with 95% confidence intervals and effect sizes. Within-group longitudinal changes were assessed using paired t tests or paired sign-flip permutation tests, depending on data distribution. For EEG outcomes, multiple comparisons were controlled using the Benjamini–Hochberg false discovery rate (FDR). Corrections were applied jointly across δ and β bands for whole-brain analyses and within each band for ROI-level analyses. Prespecified moderation analyses tested Group × MEP Status, Group × Lesion Location and Group × Group × Stroke Duration interactions. For motor outcomes, responder analyses based on prespecified MCID thresholds were conducted using baseline-adjusted logistic regression. For responder analyses, prespecified minimal clinically important difference (MCID) thresholds were defined as a change from baseline of ≥ 10 points for the total FMA-UE [33], ≥ 5.6 points for the proximal subscore [34], and ≥ 4.9 points for the distal subscore [34]. Brain–behavior associations were examined using Spearman correlations between change scores in behavioral and EEG metrics. Full statistical specifications, assumption checks, ANOVA sensitivity analyses, and correction procedures are provided in the Supplementary Methods. Results A total of 104 first-ever stroke patients were screened, 61 were randomized, and 56 (28 per group) completed the final analysis. Reasons for exclusion and attrition are shown in Fig. 2. Baseline characteristics and adverse events are summarized in Table 1. No phosphene-like or flickering visual sensations were reported. All participants completed the blinding questionnaire, and treatment guesses did not differ between groups (χ2(2) = 0.32, p = 0.85), indicating successful blinding (Table S2). Participants in the randomized controlled trial Table 1Demographic and clinical characteristics of the patientsCharacteristicstACS (n = 28)Sham (n = 28)P-valueAge (years)α60.36 ± 12.0061.18 ± 11.690.796aGender (female/male)β9/199/191.000cNIHSS scoresα4.39 ± 2.174.07 ± 1.880.752bOnset since stroke (weeks)α8.50 ± 5.677.89 ± 5.450.841bLesion location (Cortical/subcortical/brainstem)β4/18/66/17/50.771cHemisphere (right/left)β13/1514/140.789cstroke subtype (ischemic/hemorrhagic)β23/519/90.355cMEP status ( ±)β13/1514/140.789cFMA-UE total scoresα29.25 ± 18.5628.00 ± 17.690.987bFMA-UE proximal scoresα19.71 ± 9.2218.36 ± 8.490.740bFMA-UE distal scoresα9.54 ± 10.039.64 ± 9.630.723bContinuous variables are presented as mean ± SD for consistency; between-group tests were selected according to distributional assumptionsStatistical analysis by using aWelch t-testbMann-Whitney U testcChi-square testαData are presented in mean ± SDβData are presented in number Demographic and clinical characteristics of the patients Continuous variables are presented as mean ± SD for consistency; between-group tests were selected according to distributional assumptions Statistical analysis by using aWelch t-test bMann-Whitney U test cChi-square test αData are presented in mean ± SD βData are presented in number Spectral power Baseline spectral differences between stroke patients and healthy controls To characterize baseline post-stroke spectral abnormalities, we first compared patients with healthy controls. Relative to controls, patients showed higher whole-brain relative δ power, a smaller but significant increase in θ power, and lower relative β power (δ: q = 0.004; θ: q = 0.042; β: q = 0.004; Table S3). δ-band Baseline EEG spectral measures did not differ between the β-tACS and sham groups (Table S4). Following intervention, at the whole-brain level, baseline-adjusted permutation ANCOVA showed lower δ power in the β-tACS group than in the sham group at post-intervention (F = 4.60, p = 0.037, q = 0.037). Within-group analysis further showed a significant whole-brain δ reduction in the β-tACS group (p = 0.002, q = 0.003), whereas no FDR-significant whole-brain δ change was observed in the sham group. At the ROI level, the β-tACS group showed robust within-group δ reductions across all eight ROIs (all q = 0.003), but ROI-level between-group δ effects did not survive FDR correction (Table S5, Fig. 3). β-tACS induces a bidirectional spectral shift with reduced δ power and increased β power at whole-brain and ROI levels. Relative EEG power in the δ band and β band at baseline and post intervention for the tACS and sham groups. Light bars indicate baseline and darker bars indicate post intervention. Error bars represent the standard error of the mean. Asterisks denote effects that survived multiple-comparison correction, with * indicating q < 0.05, ** indicating q < 0.01, and *** indicating q < 0.001. a Whole-brain relative δ power. Group differences at post were tested using baseline-adjusted permutation ANCOVA, and within-group baseline-to-post changes were tested on paired change scores. Scalp topographies illustrate the spatial distribution of whole-brain δ relative power at each time point for each group. b Whole-brain relative β power, shown with the same conventions as in panel a. Scalp topographies depict whole-brain β relative power at baseline and post for each group. c ROI-level relative δ power across eight predefined regions of interest, including ipsilesional frontal (iF), contralesional frontal (cF), ipsilesional temporal (iT), contralesional temporal (cT), ipsilesional central (iC), contralesional central (cC), ipsilesional posterior (iP), and contralesional posterior (cP). d ROI-level relative β power across the same ROIs and time points β-band At the whole-brain level, β power was higher in the β-tACS group than in the sham group at post-intervention (F = 8.29, p = 0.004, q = 0.009), and the β-tACS group showed a significant within-group increase from baseline to post (p = 0.004, q = 0.004), whereas the sham group did not. ROI-level between-group effects survived FDR correction in all eight ROIs. The most prominent β-band increases were observed in the ipsilesional central region (iC: F = 10.15, q = 0.014) and the ipsilesional temporal region (iT: F = 9.34, q = 0.014), followed closely by the contralesional central region (cC: q = 0.020). Within the β-tACS group, β increases were again observed across all eight ROIs, with the strongest within-group effects in iC, iT, and iP (all q = 0.011). Functional connectivity (dwPLI and graph-theoretical analyses) δ-band A coherent δ-band network downshift was observed after β-tACS. Whole-brain mean δ connectivity decreased significantly within the β-tACS group (Δ = − 0.008, q = 0.001), but not in the sham group. The strongest between-group effects were observed for iF–iT (F = 12.37, q = 0.002), cF–cT (F = 11.80, q = 0.002), and cF–cP (F = 10.63, q = 0.002), indicating a widespread weakening of low-frequency coupling. At the graph level, the β-tACS group showed lower δ-band network density (F = 6.73, q = 0.012), mean edge weight (F = 6.16, q = 0.010), clustering coefficient (F = 6.77, q = 0.012), and global efficiency (F = 6.87, q = 0.010) than the sham group (Table S6-8, Figs. 4, 5). δ-band dwPLI connectivity at baseline and post intervention. a Whole-brain δ-band dwPLI connectivity matrix in the tACS group at baseline. Warmer colors indicate stronger dwPLI values. b Whole-brain δ-band dwPLI connectivity matrix in the tACS group at post intervention. c ROI-level δ-band mean dwPLI connectivity matrix in the tACS group at baseline, obtained by averaging channel-level d values within and between ROIs. d ROI-level δ-band mean dwPLI connectivity matrix in the tACS group at post intervention. e Whole-brain δ-band dwPLI connectivity matrix in the sham group at baseline. f Whole-brain δ-band dwPLI connectivity matrix in the sham group at post intervention. g ROI-level δ-band mean dwPLI connectivity matrix in the sham group at baseline. h ROI-level δ-band mean dwPLI connectivity matrix in the sham group at post intervention ROI-to-ROI δ-band dwPLI change from baseline to post intervention. Distributions of ROI-to-ROI δ-band dwPLI change scores, computed as post minus baseline, are shown separately for the tACS and sham groups. The vertical line at zero denotes no change. Negative values indicate reduced phase-lagged coupling, whereas positive values indicate increased coupling. Asterisks mark ROI pairs that remain significant after multiple-comparison correction, with * indicating q < 0.05, ** indicating q < 0.01, and *** indicating q < 0.001. Stars shown on the left of each row correspond to within-group tests, whereas stars shown on the right correspond to between-group comparisons β-band In contrast, β-band connectivity and graph-theoretical measures did not show significant main effects after multiple-comparison correction. Neither between-group differences nor within-group changes in whole-brain β connectivity or β graph metrics survived FDR correction. Motor outcomes FMA-UE total score β-tACS was associated with greater improvement in overall upper-limb motor outcome than sham within the present 10-session protocol. At post-intervention, the adjusted between-group mean difference in total FMA-UE score was 4.01 points (95% CI 1.02 to 7.00; d = 0.72; F = 7.24, p = 0.007), increasing to 5.99 points at F/U (95% CI 2.42 to 9.56; d = 0.90; F = 11.33, p = 0.001). Responder rates based on the total-score MCID threshold were also higher in the β-tACS group at post (13/28 vs. 2/28, p = 0.017) and F/U(21/28 vs. 7/28, p < 0.001) (Table S9, S10; Fig. 6a). Motor recovery trajectories and responder analyses for FMA-UE proximal and distal subscores. Between-group effects for continuous outcomes were evaluated using baseline-adjusted Freedman–Lane permutation ANCOVA, and responder outcomes were tested using baseline-adjusted logistic models. Asterisks denote between-group significance, with *p < 0.05, **p < 0.01, and ***p < 0.001. In the trajectory plots, thin lines represent individual participants and thick lines indicate group-level mean change with the standard error of the mean. a FMA-UE total outcome. The left panel shows longitudinal trajectories across baseline, post-intervention, and follow-up (F/U) for the tACS and sham groups. The middle and right panels show responder rates at post and F/U, respectively, using the prespecified total-score MCID threshold of Δ from baseline ≥ 10 points. b FMA-UE proximal outcome. The left panel shows longitudinal trajectories across baseline, post-intervention, and F/U for the tACS and sham groups. The middle and right panels show responder rates at post and F/U, respectively, using the prespecified proximal MCID threshold of Δ from baseline ≥ 5.6 points. c FMA-UE distal outcome. The left panel shows longitudinal trajectories across baseline, post-intervention, and F/U for the tACS and sham groups. The middle and right panels show responder rates at post and F/U, respectively, using the prespecified distal MCID threshold of Δ from baseline ≥ 4.9 points FMA-UE proximal subscore At post-intervention, the adjusted mean difference was 2.64 points (95% CI 1.06 to 4.22; d = 0.90; F = 11.17, p = 0.002), increasing to 4.70 points at F/U (95% CI 2.80 to 6.61; d = 1.32; F = 24.37, p < 0.001). Proximal responder rates were also significantly higher in the β-tACS group at both post (13/28 vs. 2/28, p < 0.001) and F/U (23/28 vs. 8/28, p < 0.001) (Fig. 6b). FMA-UE distal subscore By contrast, distal improvement did not differ significantly between groups within the present 10-session protocol. The adjusted mean difference was 1.54 points at post (95% CI − 0.23 to 3.32; d = 0.47; F = 3.05, p = 0.089) and 1.51 points at F/U (95% CI − 0.65 to 3.67; d = 0.38; F = 1.97, p = 0.170). Responder rates based on the distal MCID threshold were numerically higher in the β-tACS group but did not differ significantly from sham at either time point (Fig. 6c). Moderation of treatment effects Treatment effects were moderated at both the network and behavioral levels. For spectral power, no Group × Lesion location, Group × MEP status or Group × Stroke duration interactions survived FDR correction in either δ or β bands (Table S11). By contrast, in network analyses, lesion location significantly moderated δ-band changes in network density and mean clustering coefficient (both q = 0.025), with cortical and subcortical lesions showing greater reductions than brainstem lesions. MEP status significantly moderated selected β-band network outcomes despite the absence of β-band main effects, including network density (F = 8.09, q = 0.011), mean clustering (F = 9.12, q = 0.011), whole-brain mean connectivity (F = 6.57, q = 0.027), and the iC–cC and iC–cP edges (both q ≤ 0.036), indicating that these β-band network shifts were more evident in MEP-positive participants (Table S12). For motor outcomes, lesion location and stroke duration did not significantly moderate treatment effects, and no significant moderation was observed for distal outcomes. MEP status, however, significantly moderated total and proximal recovery. For total FMA-UE, the Group × MEP interaction was significant at F/U (F = 8.35, q = 0.014), with a larger adjusted treatment effect in MEP-negative than in MEP-positive patients (10.66 vs. 1.76 points). For proximal FMA-UE, the interaction was significant at both post (F = 12.31, q = 0.005) and F/U (F = 12.59, q = 0.005), again favoring a larger treatment effect in MEP-negative participants (5.04 vs. 0.20 points at post; 7.59 vs. 1.89 points at F/U) (Table S13). Brain-behavior correlations In the β-tACS group, greater behavioral improvement was associated with larger reductions in δ-band network organization (Table S14, Fig. 7)). At post-intervention, total FMA-UE gain was significantly correlated with reductions in δ-band network density (ρ = − 0.59, q = 0.004), mean clustering coefficient (ρ = − 0.55, q = 0.005), mean connectivity (ρ = − 0.52, q = 0.005), and global efficiency (ρ = − 0.40, q = 0.047). Proximal gain showed a similar pattern, with significant correlations for δ-band network density (ρ = − 0.58, q = 0.005), mean clustering coefficient (ρ = − 0.55, q = 0.005), and mean connectivity (ρ = − 0.47, q = 0.011). At F/U, these associations remained significant for both total and proximal outcomes, particularly for δ-band network density and mean clustering coefficient. Specifically, total FMA-UE gain remained correlated with δ-band network density (ρ = − 0.53, q = 0.012) and mean clustering coefficient (ρ = − 0.51, q = 0.012), whereas proximal gain remained correlated with δ-band network density (ρ = − 0.45, q = 0.031) and mean clustering coefficient (ρ = − 0.46, q = 0.031). No other EEG measures remained significantly associated with behavioral improvement after FDR correction in these analyses. δ-band network downshift is associated with upper-limb motor recovery. This figure summarizes δ-band brain–behavior associations between changes in whole-brain network measures and changes in FMA-UE outcomes. Brain metrics are expressed as Δ(post − baseline). Behavioral outcomes are expressed as Δ(post − baseline) and Δ(F/U − baseline); for follow-up behavioral endpoints, correlations were computed against the same brain Δ(post − baseline). Correlation coefficients are Spearman’s ρ. Asterisks denote correlations that survived multiple-comparison correction, with indicating q < 0.05, indicating q < 0.01, and indicating q < 0.001. a Heatmap of Spearman correlations between changes in whole-brain δ-band graph metrics and changes in FMA-UE total, proximal, and distal scores at post and follow-up (F/U) in the β-tACS group. Cell values denote ρ. b Representative scatterplots showing the association between Δ whole-brain network density and Δ FMA-UE total score (left) and Δ FMA-UE proximal score (right) at post. Red, blue, and black lines represent the β-tACS group, sham group, and pooled sample, respectively; shaded bands indicate fitted uncertainty, and marginal density plots show the distributions of each variable Discussion We evaluated 20-Hz β-tACS paired with routine rehabilitation in subacute stroke and observed greater upper-limb recovery within the present 10-session protocol, supported by between-group differences in FMA-UE change and MCID responder rates. At the clinical level, the between-group separation was evident for overall FMA-UE improvement and was most pronounced for the proximal subscore at both post-intervention and the 1-month F/U, accompanied by higher proximal MCID responder rates in the active group. Distal performance improved over time in both groups, but the between-group contrasts were smaller and did not reach statistical significance at either time point. Physiologically, the intervention produced a bidirectional spectral shift toward higher relative β power and lower relative δ power, alongside a δ-band dominant network downshift characterized by reduced functional coupling. Importantly, larger improvements in FMA-UE total and proximal function were associated with larger reductions in δ-band network organization. A significant treatment-by-phenotype interaction suggested moderation by MEP status, whereas lesion-location and stroke-duration effects were less consistent. In the healthy brain, β activity has been linked to maintaining the current sensorimotor set through inhibitory control mechanisms, consistent with the “status quo” account [18, 35–37]. Surround inhibition may help suppress competing motor representations during voluntary movement and support more selective muscle activation [38, 39]. Stroke can disrupt this balance and is often accompanied by altered β-band dynamics and increased low-frequency activity, including elevated δ power, which has been linked to injury burden, recovery trajectories, altered cortical state, and disrupted large-scale communication [8, 40, 41]. Pathological δ dominance has been interpreted as reflecting disrupted large-scale afferent drive and thalamo-cortical network dysfunction after focal injury, so a δ decrease is compatible with a shift away from a deafferented or low-responsiveness network regime toward a more responsive state that can better support information processing and recovery [14]. Accordingly, 20 Hz β-tACS can be considered a frequency-targeted intervention intended to bias oscillatory dynamics toward a more functional operating point. By externally applying β-range rhythmic drive, tACS may bias cortical excitability and timing toward β-consistent dynamics rather than directly generating movement [26]. This priming effect does not generate movement itself, but it may provide a permissive physiological context in which task practice may more effectively engage Hebbian plasticity mechanisms [42]. This framework is consistent with the view that stimulation effects are amplified when paired with concurrent training that provides behaviorally relevant synaptic drive [43, 44]. Our spectral power findings are consistent with frequency-matched engagement and partial normalization of post-stroke resting physiology. This interpretation is supported by our baseline comparison, which confirmed that patients exhibited significantly elevated δ and suppressed β activity relative to healthy controls. The increase in relative β power was close to the stimulation frequency, which is compatible with resonance or entrainment-like effects. This aligns with evidence linking sensorimotor β rhythms to the maintenance of motor set and inhibitory control, and with reports that β-tACS can modulate corticospinal excitability and plasticity-related processes [24–26, 45]. Concurrently, the reduction in relative δ power is compatible with attenuation of post-stroke spectral slowing that has been associated with injury and diaschisis-related network disruption [8]. This departure from δ dominance may reflect a shift away from a low-responsiveness state toward a more reactive cortical mode that is better suited for information processing [46, 47]. Although ROI-level δ effects did not consistently survive correction, the significant whole-brain reduction supports a more global attenuation of slow-wave dominance rather than a purely focal effect. More specifically, the δ-band pattern appeared to be spatially distributed rather than driven by a single dominant subregion. The active group showed robust within-group δ reductions across all ROIs, whereas the sham group showed smaller same-direction changes in several regions, which likely reduced the magnitude of ROI-wise between-group contrasts after correction while preserving the whole-brain effect. By contrast, the β-band effect was most robust at the spectral level, with strongest ROI effects in the ipsilesional central and temporal regions, consistent with target engagement in an M1-centered but spatially distributed sensorimotor network. Beyond immediate entrainment, repeated 20 Hz stimulation may interact with synaptic plasticity mechanisms, including spike-timing-dependent plasticity (STDP)-like timing effects. Recent models propose that tACS may bias spike timing relationships rather than merely adding a background rhythm [48, 49]. By promoting β-range rhythmicity across the motor network, tACS may, in principle, increase phase consistency across distributed regions, including premotor and parietal cortices. This temporal alignment could create windows in which rehabilitation-related inputs are more likely to coincide with higher-excitability phases. Such phase-consistent modulation could, in theory, facilitate Hebbian strengthening of task-relevant connections during training [36, 48, 50]. This state-dependent perspective offers a plausible account of frequency-specific benefits and supports pairing stimulation with active rehabilitation to translate physiological modulation into behavioral gains. The treatment-by-MEP interaction observed in our moderator analysis supports a mechanistically informed interpretation of why proximal and distal functions may respond differently. In this trial, participants without recordable MEPs showed larger treatment-associated gains than those with preserved MEPs, suggesting that response may not be limited to individuals with higher corticospinal reserve. This pattern is consistent with the possibility that recovery in individuals with reduced corticospinal reserve may rely more heavily on compensatory motor-system recruitment [51]. However, the present data do not indicate that MEP-negative and MEP-positive patients were engaging fundamentally different descending pathways in any direct physiological sense; rather, they suggest that baseline corticospinal reserve may shape the extent to which a distributed neuromodulatory intervention translates into measurable clinical benefit. Although relative β power increased in a frequency-matched manner, β-band connectivity and graph metrics did not remain significant after multiple-comparison correction, suggesting that spectral power modulation and network-level reconfiguration may be partially dissociable. In contrast, the δ-band showed a coherent downshift across both coupling and topology, and the magnitude of this δ-band reduction tracked clinical improvement, which is more consistent with attenuation of maladaptive low-frequency hypersynchrony than with global network degradation [8, 14, 47]. Importantly, decreases in density, clustering, or efficiency are not inherently deleterious [52, 53]. When the pre-intervention state is dominated by pathological hypersynchrony, lower values can plausibly reflect desynchronization and greater dynamical flexibility rather than loss of healthy integration [8, 14, 41, 52, 53]. This also helps explain the dissociation between β power and β connectivity: relative power indexes local or regional oscillatory magnitude, whereas dwPLI indexes phase-lagged interregional coupling; therefore, β-power increases need not be accompanied by parallel β-band connectivity changes [54]. This interpretation aligns with prior work linking post-stroke low-frequency activity to injury burden and recovery trajectories [8, 41, 47], and with guidance that graph metrics should be interpreted in the context of the underlying network state rather than as monotonic markers of “better” or “worse” organization [52, 53]. In addition, the β-band interaction observed in MEP-positive participants may be compatible with preserved capacity for stimulation-related network adjustment, whereas the smaller clinical gains might reflect heterogeneity in rehabilitation responsiveness, baseline functional reserve, or the degree to which routine therapy sufficiently targets distal control [43, 55]. Collectively, these observations support considering a tiered rehabilitation approach in which open-loop β-tACS may be more suitable for proximal deficits and for individuals with reduced corticospinal output [4, 5, 55], while distal goals may require more intensive hand-specific practice and state-informed stimulation strategies designed to better engage timing-sensitive corticospinal dynamics [43, 48–50, 55–57]. Fine distal dexterity depends heavily on high-fidelity corticospinal projections that support fractionated finger movements [57]. In MEP-negative patients, reduced functional corticospinal output is likely to limit distal recovery potential, particularly for dexterous finger control [55]. In contrast, proximal actions such as reaching can recruit more distributed bilateral cortical networks and additional descending pathways. One prominent alternative pathway is the cortico-reticulospinal system. These systems can contribute to gross motor synergies and compensatory control when corticospinal output is compromised [4, 5, 58, 59]. That said, the contribution of the reticulospinal tract to post-stroke recovery remains debated, and our study did not directly assess reticulospinal function. We therefore interpret the present proximal–distal dissociation more cautiously as reflecting differences in network organization, compensatory potential, and corticospinal dependence, rather than as evidence that β-tACS directly modulated a specific non-corticospinal descending pathway [4, 51, 60]. Lesion topography may also have contributed to the observed proximal–distal pattern. Damage involving distal-related motor regions, particularly the posterior precentral gyrus/hand area, has been associated with distal upper-extremity motor deficits after stroke, whereas proximal deficits may relate to more distributed subcortical white- and gray-matter injury [61]. Damage to corticospinal tract-related white-matter pathways, including those traversing the corona radiata and internal capsule, may also disproportionately limit hand and distal motor recovery [55, 60]. Although our exploratory moderator analysis did not show significant moderation of motor outcomes by the broad cortical/subcortical/brainstem classification, this classification is not sufficient to rule out the influence of specific lesion locations or corticospinal tract lesion load. The observed proximal–distal dissociation should also be interpreted in relation to stimulation dose. The present 10-session protocol demonstrated significant effects on total and proximal FMA-UE outcomes, but it may not have provided sufficient cumulative stimulation or hand-specific practice to detect distal gains. A recent systematic review of tACS in stroke rehabilitation highlighted substantial heterogeneity in stimulation parameters and emphasized the need to optimize stimulation frequency, dose, and treatment schedules [31]. Therefore, a longer treatment course or higher cumulative dose could potentially produce clearer distal effects and attenuate the apparent proximal–distal dissociation observed under the present 10-session protocol. Our findings suggest that β-tACS may influence recruitment of compensatory systems by modulating global oscillatory state. Specifically, attenuation of δ-band hypersynchrony may reduce maladaptive coupling and thereby make distributed motor networks more amenable to compensatory recruitment. This modulation may facilitate the recruitment of broad motor synergies essential for proximal recovery [4]. This process may be particularly relevant for MEP-negative patients, in whom corticospinal output is reduced and compensatory strategies may play a larger role. Conversely, in MEP-positive patients with preserved corticospinal output, standard rehabilitation may already engage the dominant pathway more effectively. This could contribute to smaller incremental effects of neuromodulation in this subgroup [4, 28, 55]. Furthermore, because tACS may bias global coupling and timing rather than selectively strengthening specific microcircuits, it may be better suited to supporting distributed coordination relevant to proximal function than the precise temporal control required for distal dexterity [5, 62, 63]. We observed δ-band reductions in whole-brain density, mean connectivity, mean edge weight, clustering, and global efficiency after β-tACS, a pattern that is consistent with attenuation of maladaptive over-coupling and release from a rigid slow-frequency regime rather than a loss of healthy network integrity. After stroke, increased low-frequency coupling has been associated with altered information processing and network topology [64], and longitudinal work has reported a normalization of increased low-frequency coupling alongside restitution of more complex network organization during recovery [64]. From a graph-theoretical perspective, density, clustering, and efficiency summarize how strongly and how locally or globally a network is interconnected; therefore, decreases in these measures can reflect a shift in segregation and integration rather than simple deterioration, particularly when the pre-intervention state is dominated by hypersynchrony [52, 53]. This interpretation is also consistent with evidence that stimulation frequency can differentially bias network topology. In chronic stroke, a graph-theoretical fMRI study comparing 10 Hz and 20 Hz tACS reported frequency-dependent effects on network segregation and integration, with 20 Hz stimulation showing changes more consistent with enhanced inter-community exchange [45]. Specifically, 20 Hz stimulation was associated with greater inter-community exchange and more differentiated modular organization, consistent with the view that β-tACS may bias network-level configuration rather than simply reducing nonspecific activity. The persistence of the proximal advantage at 1-month F/U is more consistent with use-dependent plasticity mechanisms than with a purely transient excitability shift. This suggests that the durability of the observed behavioral gains may reflect plastic after-effects that extend beyond the stimulation window [65]. This interpretation aligns with state-dependent neuromodulation frameworks proposing that NIBS efficacy is shaped by the concurrent functional state of the targeted network. The exogenous oscillation may have provided a temporal scaffold that biased network timing during rehabilitation tasks [66]. This could support plasticity processes relevant to motor learning and retention [67]. Recent work suggests that 20 Hz β-tACS can engage NMDA receptor-dependent plasticity mechanisms in human motor cortex [25]. This supports the view that stimulation may act as a catalyst for training-driven plasticity processes. More broadly, β rhythms have been proposed to stabilize the current sensorimotor state, so externally reinforcing β-range dynamics may help restore a more functional operating point that rehabilitation can exploit to drive lasting motor gains [18, 50, 66]. Several limitations should be considered. First, mechanistic inference is constrained by the measurement level. Scalp EEG can capture oscillatory state and large-scale coupling, but it cannot directly demonstrate strengthening of a specific descending pathway, and the absence of concurrent diffusion imaging or TMS measures limits pathway-level interpretation. Second, the physiological measures were obtained at rest rather than during movement. As a result, the observed normalization of resting-state activity cannot be assumed to directly reflect movement-related control processes or functional arm use. Third, the stimulation should be interpreted as M1-centered rather than focal. Given the conventional sponge-electrode montage, extracephalic return, and lesion-related variability in current flow, the induced field was likely relatively distributed, which limits strong claims about target specificity. Fourth, lesion location was categorized only broadly as cortical, subcortical, or brainstem. This classification cannot exclude confounding by specific lesion topography or corticospinal tract lesion load, particularly for distal recovery. Future studies should include lesion segmentation, diffusion imaging, corticospinal tract lesion-load measures, and voxel-wise lesion–symptom mapping. Fifth, the stimulation dose was fixed at 10 sessions over 2 weeks. Although this allowed us to test short-term target engagement and behavioral effects, it does not establish the optimal dose or dose–response relationship of β-tACS after stroke. Given the heterogeneity of current tACS protocols, distal effects may require a longer treatment course, higher cumulative dose, or more intensive hand-specific pairing; therefore, the non-significant distal effect should not be interpreted as evidence that distal recovery is unresponsive to β-tACS [56]. Finally, although responder analyses were based on prespecified MCID thresholds, MCID estimates for the FMA-UE vary across populations and methods; these results should therefore be regarded as supportive rather than definitive [34]. Future work should combine individualized electric-field modeling with pathway-resolved stratification (DTI/TMS), longer F/U, and task-based neurophysiology to test whether the observed resting-state normalization predicts movement-locked dynamics and functional use. For distal hand outcomes, more targeted strategies may be required, including higher-definition montages, pairing with intensive hand-specific training, and closed-loop or phase-specific protocols aligned to voluntary movement attempts to better engage corticospinal timing demands [68]. Future studies should also directly compare M1-centered stimulation with premotor- or supplementary-motor-area-centered approaches, particularly if the goal is to differentially target proximal control or broader network recruitment. In conclusion, this randomized, double-blind, sham-controlled trial suggests that within the present 10-session protocol, 20-Hz β-tACS applied over ipsilesional M1 and paired with standardized rehabilitation was associated with clearer benefits for proximal than distal upper-limb recovery in subacute stroke, with gains maintained at one month F/U. The intervention was associated with a resting-state EEG pattern characterized by higher relative β power, lower relative δ power, and a δ-band dominant reduction in functional coupling and global graph organization. Importantly, participants showing larger clinical improvements, especially in proximal function, also tended to show larger reductions in δ-band network organization, linking attenuation of slow frequency coupling to behavioral benefit. These findings support a model in which network-level state normalization may create conditions that rehabilitation can capitalize on, while the non-significant distal effect should be interpreted cautiously because distal dexterity may be more constrained by corticospinal pathway integrity, lesion topography, stimulation dose, and the need for targeted task-specific pairing. Supplementary Information Supplementary Material 1. Supplementary Material 2. Acknowledgements We are grateful to the patients who participated in this trial and to their families for their trust and commitment. We thank our clinical colleagues for their support—rehabilitation physicians and attending physicians, physical therapists, occupational therapists. We also acknowledge Dr. Bu and colleagues for their article, "Common and specific effects in brain oscillations and motor symptoms of tDCS and tACS in Parkinson’s disease”, which inspired the design of the graphical abstract for this manuscript. Abbreviations Fugl-Meyer assessment Primary motor cortex National Institutes of Health Stroke Scale Power spectral density Transcranial alternating current stimulation Transcranial magnetic stimulation Non-invasive brain stimulation Debiased weighted phase-lag index Author contributions X-YL: drafting/revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data. YZ: revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data. YG: drafting/revision of the manuscript for content. AG: drafting/revision of the manuscript for content, including medical writing for content; analysis or interpretation of data. SL: drafting/revision of the manuscript for content, including medical writing for content; analysis or interpretation of data. A. Yasen: major role in the acquisition of data. S. Qian: analysis or interpretation of data. W. Qian: major role in the acquisition of data. R. Xu: major role in the acquisition of data. G. Xu: study concept or design; analysis or interpretation of data. Y-E. Jiang: drafting/revision of the manuscript for content, including medical writing for content; analysis or interpretation of data. Y. Gao: study concept or design. Funding This work was supported by the Suzhou Applied Basic Research (Medical and Health) Technology Innovation Project (grant number SYW2024115), the Gusu District Health Talent Training Project (GSWS2022073). Data availability Study protocols and deidentified participant data can be made available on request. Declarations Ethics approval and consent to participate The study was approved by the Ethics Committee of Suzhou Municipal Hospital and was registered with the Chinese Clinical Trial Registry (ChiCTR2400083526) on April 26, 2024. All participants provided informed consent before participating in the study. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Xiao-Yu Liao and Yuxin Zhang authors have contributed equally to this work and share first authorship. Contributor Information Xiao-Yu Liao, Email: xiaoyuliao06@163.com. Yu-Er Jiang, Email: jiangye2023@ion.ac.cn. Yaxin Gao, Email: gaoyax809@126.com. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Data Availability Statement Study protocols and deidentified participant data can be made available on request.