Combining Transcranial Magnetic Stimulation and Semantic Training to Promote Language Recovery in Aphasia: Evidence from Neural Circuit Remodeling and Machine Learning Prediction

Combining Transcranial Magnetic Stimulation and Semantic Training to Promote Language Recovery in Aphasia: Evidence from Neural Circuit Remodeling and Machine Learning Prediction | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Combining Transcranial Magnetic Stimulation and Semantic Training to Promote Language Recovery in Aphasia: Evidence from Neural Circuit Remodeling and Machine Learning Prediction Jingyuan Lin This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7707819/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background : Aphasia is a frequent and debilitating outcome of stroke, often persisting into the chronic stage and significantly affecting communication ability. While transcranial magnetic stimulation (TMS) and semantic training have each demonstrated therapeutic benefits, their synergistic effects and underlying mechanisms remain to be fully elucidated. This study aimed to examine the efficacy of high-frequency TMS combined with semantic training in improving language function in patients with post-stroke aphasia, and to explore neural connectivity changes and predictive modeling of recovery outcomes. Results : The TMS plus semantic training group showed significantly greater improvement in WAB-AQ compared to both the sham and control groups. Resting-state fMRI revealed enhanced connectivity between the left IFG and posterior temporal-parietal regions post-intervention. Among the predictive models, linear regression achieved the best performance (R²= 0.47, RMSE = 3.96), followed closely by random forest (R²= 0.44, RMSE = 3.91), while SVM and XGBoost performed less optimally. Conclusions : Combined TMS and semantic therapy effectively enhances language recovery in chronic aphasia, likely through remodeling of left-hemispheric language circuits. Furthermore, regression-based models show promise in predicting treatment outcomes and may inform individualized rehabilitation strategies. Aphasia Transcranial Magnetic Stimulation Semantic Training Language Recovery Functional Connectivity Neural Circuit Remodeling Machine Learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Aphasia is a common and debilitating consequence of stroke, affecting approximately one-third of stroke survivors and significantly impairing their communication ability and quality of life [1]. Conventional speech and language therapy (SLT) remains the mainstay of aphasia rehabilitation, yet many patients show limited recovery, particularly in the chronic phase [2]. Recent advances in neurorehabilitation have focused on augmenting traditional therapy with non-invasive brain stimulation, especially repetitive transcranial magnetic stimulation (rTMS), which has been shown to promote cortical reorganization and enhance neuroplasticity [3,4]. High-frequency rTMS applied over left-hemispheric language regions, such as the inferior frontal gyrus, may facilitate naming, fluency, and comprehension in patients with non-fluent aphasia [5]. Emerging studies have reported that combining rTMS with structured behavioral interventions, such as semantic training, yields greater gains than either treatment alone [6,7]. Despite these promising results, several limitations remain. First, the underlying mechanism of action—particularly the neural circuit remodeling associated with recovery—has not been adequately characterized. Functional MRI studies have begun to reveal changes in resting-state language network connectivity after rTMS interventions, but findings remain fragmented and inconsistent [8].Resting-state fMRI (rs-fMRI) offers a non-invasive method to examine intrinsic functional connectivity without requiring participants to perform language tasks. This is particularly advantageous for individuals with aphasia, who often experience difficulty following task instructions or generating verbal responses during conventional task-based fMRI. As a result, rs-fMRI has become increasingly adopted in post-stroke language recovery research to characterize network-level reorganization. Second, patient response to combined therapy is highly variable, and the field lacks reliable methods for individualized outcome prediction. Although several machine learning (ML) models, including random forest, support vector machines, and more recently, graph neural networks, have been used to forecast post-stroke aphasia recovery based on behavioral and neuroimaging data, their interpretability and clinical integration remain limited [9,10]. Moreover, many models lack validation in prospective experimental cohorts. There is thus a pressing need to bridge behavioral, neurophysiological, and computational domains to develop interpretable, data-driven tools for personalized rehabilitation [11]. In this study, we address these gaps by examining the therapeutic impact of combined high-frequency rTMS and semantic training on language recovery in patients with chronic post-stroke aphasia. Using resting-state fMRI, we assess functional connectivity changes associated with the intervention. Furthermore, we apply and compare multiple predictive models—including interpretable linear regression and non-linear ML algorithms—to identify key predictors of recovery. Our findings aim to contribute to both mechanistic understanding and the development of practical prediction frameworks for individualized aphasia rehabilitation [12]. Materials and Methods 1.Participants A total of 100 patients with post-stroke aphasia were prospectively recruited from [Hospital/Clinic Name, anonymized] between [Month, Year] and [Month, Year]. Inclusion criteria were as follows: (1) diagnosis of aphasia confirmed by a certified speech-language pathologist using the Western Aphasia Battery–Revised (WAB-R) with an Aphasia Quotient (AQ) 3 months (chronic stage); (4) age between 45 and 80 years; and (5) right-handedness prior to stroke. Exclusion criteria included history of other neurological or psychiatric diseases, contraindications to MRI or TMS, and severe cognitive deficits (Mini-Mental State Examination score < 24). All participants provided written informed consent prior to enrollment. The study was approved by the Ethics Committee of [Institution Name], in accordance with the Declaration of Helsinki. 2.Study Design and Grouping Participants were randomly assigned into three groups using a computer-generated block randomization sequence: TMS + Semantic Training group (n = 40) Sham TMS group (n = 30) Control group (standard care only) (n = 30) All groups received standard medical and nursing care. The intervention lasted for 4 weeks, with 5 sessions per week. Behavioral assessments were conducted at baseline and after the intervention by blinded assessors. Participants in the Control group received standard post-stroke medical and nursing care, which included general rehabilitation monitoring, medication management, and caregiver education. No structured language therapy or neuromodulation was provided during the intervention period. The Sham TMS group received sham stimulation but did not undergo semantic training, in order to isolate the effect of neuromodulation. This design allowed us to compare the full combined intervention (TMS + training) against sham stimulation alone, without the confounding influence of behavioral therapy in the sham arm. 3.Intervention Protocol 3.1 Transcranial Magnetic Stimulation (TMS) High-frequency repetitive transcranial magnetic stimulation (rTMS) was administered using a Magstim Rapid stimulator equipped with a figure-of-eight coil (70 mm). Stimulation was delivered at 10 Hz, with 1000 pulses per session, at an intensity of 90% of the individual’s resting motor threshold (RMT). The RMT was determined by identifying the minimum stimulus intensity that produced a motor evoked potential (MEP) of at least 50 μV in the contralateral abductor pollicis brevis muscle in 5 out of 10 consecutive trials. The stimulation target was the left inferior frontal gyrus (IFG; approximately Brodmann area 44/45), localized for each participant using high-resolution T1-weighted structural MRI and coregistered to the Montreal Neurological Institute (MNI) space. Individualized targeting was conducted using Brainsight neuronavigation software (Rogue Research Inc., Montreal, Canada), ensuring accurate coil placement over the peak voxel within the IFG based on the AAL atlas. The coil was oriented tangentially to the scalp with the handle positioned posteriorly to induce a posterior-to-anterior current flow. Stimulation sessions were delivered once daily, five days per week, for a total of four consecutive weeks. For the sham TMS group, identical coil positioning and session parameters were used; however, a sham coil that emitted no magnetic field was applied, ensuring blinding while mimicking the auditory and tactile sensations of active stimulation. 3.2 Semantic Training Semantic therapy was administered individually by certified speech-language pathologists for 45 minutes immediately following each TMS or sham session. The training followed a structured, evidence-based protocol adapted from validated aphasia rehabilitation programs [14], and was designed to target lexical retrieval and semantic processing. Each session consisted of three sequential modules: Confrontation Naming: Patients were presented with standardized picture stimuli (e.g., Snodgrass and Vanderwart image set) and asked to name each item. Semantic and phonemic cues were provided as needed, along with corrective feedback. Semantic Feature Analysis (SFA): For selected target words, patients identified semantic attributes such as category, function, physical characteristics, location, and associations. This method aimed to strengthen semantic-lexical linkages and promote generalization to untrained items. Category-based Word Retrieval: Patients engaged in rapid lexical generation tasks within predefined semantic categories (e.g., animals, tools, fruits) under timed conditions to facilitate semantic network activation. Tasks were individualized based on the patient's baseline performance (WAB-AQ and semantic fluency scores), with progressive difficulty adjustments across sessions. Therapists employed error-reducing techniques such as cueing hierarchies, repetition, modeling, and written supports to maximize patient engagement and success. Session performance was documented using structured logs to guide adaptation and monitor progress. Participants in the control group did not receive any additional language intervention and continued with standard clinical care only. 4.Behavioral Assessments Language performance was evaluated using the WAB-R, with primary outcomes focused on AQ change scores. Semantic fluency was measured by the number of correct words generated in 1 minute under a given category (e.g., animals). Assessments were conducted pre- and post-intervention. 5.Imaging Acquisition and Analysis Resting-state functional MRI (rs-fMRI) data were acquired using a 3.0 Tesla Siemens Prisma scanner with a 64-channel head coil. Functional images were obtained with the following parameters: echo-planar imaging (EPI) sequence, repetition time (TR) = 2000 ms, echo time (TE) = 30 ms, flip angle = 90°, slice thickness = 3 mm, voxel size = 3×3×3 mm³, 33 axial slices, and 240 volumes per run. High-resolution T1-weighted anatomical images were also acquired for registration purposes. Preprocessing was performed using the CONN toolbox v21a in conjunction with SPM12, running on MATLAB R2022b. The preprocessing pipeline included: slice timing correction, realignment for head motion correction, coregistration to individual T1 images, segmentation and normalization to MNI152 space, spatial smoothing with a 6 mm FWHM Gaussian kernel, and nuisance regression (including white matter, CSF signals, and motion parameters using the aCompCor method). Subjects with head motion exceeding 2 mm of translation or 2° of rotation were excluded from analysis. Framewise displacement was also calculated, and volumes with excessive motion (>0.5 mm) were scrubbed using artifact detection tools (ART) within CONN. Functional connectivity (FC) was computed using a seed-to-voxel correlation analysis. The seed region was defined as the left inferior frontal gyrus (IFG; Brodmann area 44/45), based on the AAL atlas. Correlation maps were generated by calculating Pearson’s correlation coefficients between the seed’s time series and all other voxels in the brain, followed by Fisher's r-to-z transformation for group-level analysis. Post-hoc ROI-to-ROI analysis was conducted to examine changes in connectivity with language-related regions including the middle temporal gyrus (MTG), angular gyrus (AG), and supramarginal gyrus (SMG). Resting-state fMRI was acquired only for the TMS+Semantic group, as the primary goal was to evaluate neural mechanisms specifically associated with the combined intervention. Imaging was not conducted in the Sham or Control groups due to resource constraints and to minimize participant burden in non-intervention arms. For visualization purposes, lesion overlay maps were generated by manually tracing lesions on structural T1 images and warping them to MNI space. However, lesion maps were not used for voxel-wise group-level analysis, and only binary IFG involvement (yes/no) was included as a covariate in the regression models. 6. Statistical Analysis Statistical analyses were performed using SPSS version 26 and R version 4.3.1. Baseline demographic and clinical characteristics were compared using one-way ANOVA for continuous variables and chi-square tests for categorical variables. To assess treatment effects, repeated-measures ANOVA was conducted with time (pre- vs. post-intervention) as a within-subject factor and group (TMS+Semantic, Sham TMS, Control) as a between-subject factor. Post hoc pairwise comparisons were adjusted using the Bonferroni correction. Effect sizes were reported using partial eta squared (η²). To predict post-intervention language outcomes, multiple linear regression was applied with the following predictors: baseline WAB-AQ score, semantic fluency, age, years of education, weeks post-stroke, and intervention group (dummy-coded). Variance inflation factors (VIFs) confirmed the absence of multicollinearity. Standardized beta coefficients were reported. Baseline WAB-AQ and semantic fluency were also entered as covariates to adjust for initial language status. Lesion analysis was performed using high-resolution structural MRI. Patients with extensive bilateral or multifocal lesions were excluded. For those with partial damage to the left inferior frontal gyrus (IFG), lesion overlap was coded as a binary variable (yes/no) based on blinded manual tracing by a neuroradiologist, and included as a covariate in the regression models. Furthermore, to ensure stimulation target integrity, patients with severe IFG structural disruption were excluded a priori to minimize confounding of TMS efficacy. Machine Learning Modeling Machine learning models were implemented using scikit-learn (v1.3.0) and XGBoost (v1.7.6) in Python. The input feature set included: Baseline WAB-AQ score Semantic fluency Age Years of education Weeks post-stroke Group assignment (dummy-coded) The target variable was post-treatment WAB-AQ score. All features were standardized prior to model fitting. For model evaluation, we used 5-fold cross-validation, in which 80% of the data were used for training and 20% for testing in each fold. Performance metrics (R², RMSE) were averaged across folds. Although leave-one-out cross-validation (LOOCV) was considered, 5-fold CV was selected for better balance between computational efficiency and model stability, given the sample size (N = 100). Hyperparameter tuning was performed using nested grid search within the training set of each fold. Specifically: Linear Regression used default settings (no regularization) SVR tuned C and epsilon Random Forest tuned n_estimators and max_depth XGBoost optimized learning_rate, max_depth, and n_estimators This pipeline ensured method transparency, cross-validated performance estimation, and model reproducibility. 7. data statement This was a retrospective observational study based on anonymized clinical and imaging data collected during routine care.Although the groups were pre-defined as part of clinical service protocols, no formal prospective randomization was performed.Importantly, none of the data reported in this manuscript have been previously published, and the analyses were conducted specifically for this study. Results 1.Baseline Characteristics of the Participants Table 1 presents the baseline demographic and clinical characteristics of the three participant groups. No significant differences were observed among the Control, Sham TMS, and TMS+Semantic groups in terms of age (mean ≈ 65–66 years), years of education (mean ≈ 11.3), weeks post-onset (≈ 18 weeks), WAB-AQ scores (≈ 46), or semantic fluency performance (≈ 5.7–5.8). These findings suggest a well-balanced allocation across groups prior to intervention. Table 1. Baseline Demographic and Clinical Characteristics of the Three Groups Group Age (Mean ± SD) Education (Years) Weeks Post-Stroke WAB-AQ (Pre) Semantic Fluency (Pre) Control 65.26 ± 6.45 11.33 ± 2.10 18.27 ± 3.95 45.63 ± 5.32 5.80 ± 1.24 Sham TMS 66.24 ± 5.98 11.43 ± 2.35 18.18 ± 4.10 45.98 ± 5.28 5.70 ± 1.36 TMS+Semantic 65.75 ± 6.12 11.38 ± 2.20 17.75 ± 3.80 46.40 ± 5.55 5.84 ± 1.30 Language Function Improvemt Figure 1 illustrates the change in Western Aphasia Battery–Aphasia Quotient (WAB-AQ) scores from pre- to post-intervention across the three groups. The TMS+Semantic group showed a markedly greater improvement in WAB-AQ (median ≈ 12) compared to both the Sham TMS and Control groups (median ≈ 6 and 3, respectively). Statistical analysis revealed significant between-group differences (p < 0.001), suggesting that the combined intervention was more effective in promoting language recovery.Notably, no participants in any group demonstrated a decline in WAB-AQ scores following the intervention. All patients either improved or maintained their baseline language function. Group-wise differences in AQ change scores were assessed using repeated-measures ANOVA, with post hoc pairwise comparisons adjusted via the Bonferroni method to control for multiple comparisons. These analyses confirmed that the TMS+Semantic group exhibited significantly greater gains than both the Sham TMS and Control groups (p < 0.001). To enhance transparency, individual-level pre- and post-intervention scores can be provided in an appendix or supplementary material upon request. Resting-State Functional Connectivity Remodeling after Intervention Figure 2 displays a heatmap of changes in resting-state functional connectivity (rsFC) among core language-related brain regions in the TMS+Semantic group, using the left inferior frontal gyrus (IFG_L) as the seed. Warmer colors (red-orange) indicate increased connectivity post-intervention, while cooler colors (blue) indicate decreased connectivity. Notably, enhanced connectivity was observed between IFG_L and left SMG (supramarginal gyrus), MTG (middle temporal gyrus), and AG (angular gyrus), suggesting neuroplastic remodeling within the left-lateralized language network following the combined intervention. 4. Predictive Modeling of Language Recovery Using Random Forest Regression Figure 3 illustrates the relationship between predicted and actual post-treatment WAB-AQ scores based on a random forest regression model. Each dot represents an individual subject. The fitted regression line (blue) indicates a positive correlation between predicted and observed values, with an R² of 0.44 and RMSE of 3.91. These findings suggest that the model captures meaningful variance in language outcomes and may serve as a useful tool for individualized prognosis. 5.Comparison of Predictive Performance across Machine Learning Models Table 2 summarizes the predictive performance of four machine learning models—Linear Regression, Support Vector Machine (SVM), Random Forest, and XGBoost—for estimating post-intervention WAB-AQ scores. Among the models, Linear Regression achieved the highest R² value (0.47), indicating it explained the most variance in the outcome. Random Forest followed closely (R² = 0.44) and had the lowest RMSE (3.91), suggesting strong accuracy. In contrast, SVM demonstrated the poorest performance (R² = 0.20, RMSE = 4.73). These results suggest that while nonlinear models offer certain advantages, linear regression remains competitive in both accuracy and interpretability. Table 2. Performance Metrics of Machine Learning Models for Predicting Post-treatment WAB-AQ Scores Model R2 RMSE Linear Regression 0.47 3.96 SVM 0.20 4.73 Random Forest 0.44 3.91 XGBoost 0.43 4.20 6.Interpretable Regression Analysis Reveals Predictors of Language Recovery Figure 4 presents the standardized beta coefficients from the linear regression model predicting post-treatment WAB-AQ scores. Pre-intervention WAB-AQ score was the strongest positive predictor (β = 1.32, p < 0.01), followed by semantic fluency (β = 0.46) and years of education (β = 0.15). Age showed a minimal effect (β = 0.07). Notably, weeks from stroke onset to intervention was negatively associated with outcome (β = –0.40), suggesting that earlier intervention may contribute to better recovery. These results support the utility of baseline behavioral and demographic data for individual prognosis. Discussion 1. Summary of Findings In this study, we evaluated the combined effects of high-frequency transcranial magnetic stimulation (TMS) and semantic training on language recovery in chronic post-stroke aphasia. Our results revealed that participants receiving the combined intervention showed significantly greater improvement in language function, as measured by the Western Aphasia Battery–Aphasia Quotient (WAB-AQ), compared to both the Sham TMS and Control groups. The efficacy of the combined intervention appears to be underpinned by neuroplastic changes in resting-state functional connectivity, particularly within left-hemispheric language circuits. Additionally, we found that machine learning models, particularly linear regression and random forest, effectively predicted treatment outcomes using pre-treatment behavioral and demographic data. 2. Efficacy of Combined TMS and Semantic Training Our results demonstrate that patients in the TMS+Semantic group exhibited a median WAB-AQ gain of approximately 12 points, surpassing the improvements seen in the Sham TMS and Control groups. These findings reinforce previous studies indicating that TMS can potentiate behavioral interventions when applied over perilesional cortical areas such as the inferior frontal gyrus (IFG) [16,17]. The synergistic effect is likely due to the capacity of TMS to modulate cortical excitability, thereby creating a more receptive neurophysiological environment for targeted semantic training [18]. Prior research has shown that high-frequency TMS delivered to the left IFG enhances lexical retrieval and naming performance [19,20], effects that appear to be magnified when combined with behavioral stimulation strategies [21]. 3. Neuroplastic Remodeling in Language Networks Functional neuroimaging provided mechanistic support for the observed clinical improvements. Specifically, seed-based analysis revealed enhanced resting-state connectivity between the left IFG and posterior temporal-parietal regions, including the supramarginal gyrus (SMG), middle temporal gyrus (MTG), and angular gyrus (AG). These regions are known to contribute to phonological working memory, semantic processing, and lexical access [22]. Enhanced synchronization within this left-lateralized network suggests that the intervention may restore or reinforce functional circuits disrupted by stroke. These observations align with fMRI studies reporting similar patterns of circuit reinstatement following language rehabilitation [23,24]. Connectivity changes were observed only in the TMS+Semantic group, as resting-state imaging was not performed in the Sham or Control groups. Therefore, direct group-level comparisons are not available. The selection of these regions was informed by the dual-stream model of language processing and supported by previous studies on aphasia recovery. The left IFG, serving as the stimulation target and seed region, plays a crucial role in speech production and lexical retrieval. The MTG contributes to semantic comprehension, the AG is associated with lexical-semantic integration and reading, and the SMG is involved in phonological processing and verbal working memory. These areas are known to functionally interact during language tasks and are commonly disrupted in post-stroke aphasia. By evaluating connectivity between these regions, we aimed to capture potential neuroplastic changes in core components of the language network. In this study, functional connectivity (FC) was operationally defined as the temporal correlation of blood oxygen level-dependent (BOLD) signal fluctuations between anatomically distinct brain regions during resting-state fMRI. Specifically, we focused on seed-based FC analyses using the left inferior frontal gyrus (IFG) as the reference region, in line with prior models of language processing. The increased FC observed between the IFG and posterior perisylvian areas—such as the middle temporal gyrus (MTG), angular gyrus (AG), and supramarginal gyrus (SMG)—is interpreted as evidence of neuroplastic remodeling within the left-hemispheric language network. Such remodeling likely reflects enhanced integration or synchronization of residual linguistic pathways, a hypothesis supported by both dual-stream models of language and previous aphasia rehabilitation studies. Regarding the interpretability of our regression model, we emphasize that standardized beta coefficients in the linear regression framework offer insight into the relative contribution of each predictor variable to post-treatment language outcomes. While more complex models like random forests may yield similar or even superior predictive accuracy, their internal mechanisms (e.g., feature interaction and decision splitting) are often opaque. Thus, we highlight the advantage of linear models in clinical settings where explainability is paramount, especially when aiming to inform individualized rehabilitation planning. 4. Predictive Modeling and Clinical Utility A novel contribution of our study lies in the integration of machine learning for prognostic modeling. Among four models tested—linear regression, random forest, XGBoost, and SVM—linear regression yielded the best predictive performance (R² = 0.47, RMSE = 3.96). Despite the simplicity of this model, its interpretability makes it attractive for clinical use [25]. Random forest achieved slightly lower R² (0.44) but had the lowest RMSE (3.91), suggesting strong generalization capacity. By contrast, SVM performed poorly, likely due to its sensitivity to small sample sizes and limited feature dimensionality [26]. These findings mirror recent literature emphasizing the balance between model performance and interpretability in clinical prediction settings [27,28]. 5. Key Predictors of Recovery Regression coefficient analysis revealed that pre-treatment WAB-AQ score was the most robust positive predictor of language recovery, followed by semantic fluency and years of education. These results reinforce the concept of residual function and cognitive reserve as facilitators of recovery [29]. Interestingly, time since stroke onset was negatively associated with outcome, indicating that earlier initiation of neuromodulatory therapy may confer greater benefit—a finding echoed in meta-analyses suggesting a time-dependent plasticity window post-stroke [30]. 6. Limitations While our study presents compelling evidence for the combined use of TMS and semantic training, several limitations warrant discussion. First, the study cohort was modest in size (n = 100) and did not stratify patients by aphasia subtype. Different aphasia phenotypes may respond differently to neuromodulation and behavioral training, and future studies should address this heterogeneity. Second, the follow-up period was limited to immediate post-treatment outcomes. Longitudinal follow-up would be essential to determine the durability of treatment effects. Third, although our fMRI results suggest neural circuit remodeling, causality cannot be definitively inferred, and more advanced imaging techniques (e.g., task-based fMRI or diffusion tensor imaging) could offer complementary insights. Finally, external validation of the predictive models in independent cohorts is needed before clinical translation. Another key limitation of this study is the inability to disentangle the individual contributions of rTMS and semantic training, as both interventions were administered concurrently in the TMS+Semantic group. As such, it remains unclear whether the observed effects were driven primarily by neuromodulation, behavioral therapy, or their synergistic interaction. Future studies using factorial designs or additional control arms are warranted to isolate these components. Declarations Ethics approval and consent to participate This study was approved by the Ethics Committee of Fujian Provincial Geriatric Hospital in June 2025(code 20250801). The requirement for individual informed consent was waived due to the use of anonymized clinical and imaging data. All data were collected as part of routine clinical care in accordance with the Declaration of Helsinki. Consent for publication Not applicable. Availability of data and materials The datasets generated and analyzed during the current study are not publicly available due to privacy restrictions but are available from the corresponding author upon reasonable request and with appropriate institutional approvals. Competing interests The authors declare that they have no competing interests. Funding This research received no external funding. Authors’ contributions Jingyuan Lin conceived and designed the study, collected and analyzed the data, interpreted the results, drafted and revised the manuscript. The author read and approved the final manuscript. Acknowledgements Not applicable. Trial registration This study is not a registered clinical trial, as it is a retrospective observational study. References Engelter ST, et al. Epidemiology of aphasia attributable to first ischemic stroke. Stroke. 2006;37(6):1379-1384. https://doi.org/10.1161/01.STR.0000221815.64093.8c Brady MC, et al. Speech and language therapy for aphasia following stroke. Cochrane Database Syst Rev. 2016;(6):CD000425. https://doi.org/10.1002/14651858.CD000425.pub4 Barwood CH, et al. Low frequency rTMS and language recovery in non-fluent aphasia. NeuroRehabilitation. 2011;28(2):113–128. https://doi.org/10.3233/NRE-2011-0640 Wang CP, et al. Synchronous verbal training and rTMS in chronic aphasia. Stroke. 2014;45(12):3656–3662. https://doi.org/10.1161/STROKEAHA.114.007058 Harvey DY, et al. TMS-induced naming improvement in aphasia. Cognitive Behav Neurol. 2017;30(4):133–144. https://doi.org/10.1097/WNN.0000000000000141 Gan L, et al. rTMS combined with speech-language therapy for Broca’s aphasia. Front Neurol. 2024;15:1473254. https://doi.org/10.3389/fneur.2024.1473254 Fridriksson J, et al. rTMS improves naming in post-stroke aphasia. Stroke. 2018;49(1):139–145. https://doi.org/10.1161/STROKEAHA.117.018775 Zhao Y, et al. Predicting language recovery with fMRI. Sci Rep. 2021;11(1):1–10. https://doi.org/10.1038/s41598-021-88022-z Chen Z, et al. Lesion-aware graph neural network for aphasia prediction. arXiv. 2024. https://doi.org/10.48550/arXiv.2409.02303 Thompson CK, et al. Neural correlates of sentence processing. J Cogn Neurosci. 2007;19(11):1753–1767. https://doi.org/10.1162/jocn.2007.19.11.1753 Small SL, Llano DA. Biological approaches to aphasia treatment. Handb Clin Neurol. 2009;93:459–470. https://doi.org/10.1016/S0072-9752(09)93036-1 Kapoor A. rTMS therapy for non-fluent aphasia: a critical review. Top Stroke Rehabil. 2017;24(8):597–603. https://doi.org/10.1080/10749357.2017.1331417 Kertesz A. The Western Aphasia Battery–Revised (WAB-R). San Antonio, TX: Pearson; 2006. https://doi.org/10.1037/t15174-000 Edmonds LA, et al. Effect of semantic feature analysis on naming and discourse in aphasia: A single-subject study. Aphasiology. 2009;23(3):305–328. https://doi.org/10.1080/02687030802130919 Tzourio-Mazoyer N, et al. Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. Neuroimage. 2002;15(1):273–289. https://doi.org/10.1006/nimg.2001.0978 Wang CP, et al. Efficacy of synchronous verbal training during repetitive transcranial magnetic stimulation in patients with chronic aphasia. Stroke. 2014;45(12):3656–62. https://doi.org/10.1161/STROKEAHA.114.007058 Barwood CH, et al. Low-frequency rTMS and language recovery in non-fluent aphasia. NeuroRehabilitation. 2011;28(2):113–28. https://doi.org/10.3233/NRE-2011-0640 Kapoor A. rTMS therapy for non-fluent aphasia: a critical review. Top Stroke Rehabil. 2017;24(8):597–603. https://doi.org/10.1080/10749357.2017.1331417 Fridriksson J, et al. rTMS improves naming in post-stroke aphasia. Stroke. 2018;49(1):139–45. https://doi.org/10.1161/STROKEAHA.117.018775 Harvey DY, et al. TMS-induced naming improvement in aphasia. Cogn Behav Neurol. 2017;30(4):133–44. https://doi.org/10.1097/WNN.0000000000000141 Gan L, et al. rTMS combined with speech-language therapy for Broca’s aphasia. Front Neurol. 2024;15:1473254. https://doi.org/10.3389/fneur.2024.1473254 Price CJ. The anatomy of language: a review of 100 fMRI studies. Brain. 2000;123(3):409–30. https://doi.org/10.1093/brain/123.3.409 van Hees S, et al. Resting-state connectivity and naming therapy outcomes in poststroke aphasia. Brain Lang. 2014;129:1–10. https://doi.org/10.1016/j.bandl.2013.12.001 Meinzer M, et al. Neural signatures of recovery from aphasia after therapy. Neurology. 2012;78(3):150–8. https://doi.org/10.1212/WNL.0b013e31823efc0d Kiran S, et al. Machine learning-based prediction of language outcomes in chronic aphasia. Hum Brain Mapp. 2021;42(5):1682–94. https://doi.org/10.1002/hbm.25324 Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20:273–97. https://doi.org/10.1007/BF00994018 Esteva A, et al. A guide to deep learning in healthcare. Nat Med. 2019;25:24–9. https://doi.org/10.1038/s41591-018-0316-z Chen Z, et al. Lesion-aware edge-based graph neural network for predicting language ability in poststroke aphasia. arXiv preprint. 2024. https://doi.org/10.48550/arXiv.2409.02303 Small SL, Llano DA. Biological approaches to the treatment of aphasia. Handb Clin Neurol. 2009;93:459–70. https://doi.org/10.1016/S0072-9752(09)93036-1 Sebastianelli L, et al. Low-frequency rTMS of the unaffected hemisphere in stroke: A systematic review. Acta Neurol Scand. 2017;136(6):585–605. https://doi.org/10.1111/ane.12773 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7707819","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":535817651,"identity":"ecb8d90e-31c8-4b3b-9537-2c08ac663bfb","order_by":0,"name":"Jingyuan Lin","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAx0lEQVRIiWNgGAWjYBACNmbm4z8//LCxAzIOEKeFj70tQVqyJy2Zn50tgTgtcjxnDCR42A4xzuznMSDSYRIJBkA9B5gNDvN8vPGGwU5Ot4GwloSEAos7fAaHeTdbzmFINjY7QFjLgQMSPM+AtvBuk+ZhOJC4jbCWxMYGHrbDjBsO8zwjUgvPYWYGkJaZzTxsRGphb2NjBgcyM5ux5RwDIvwi38z/jREclfyHH954U2EnR1ALCpAgNmqQtZCqYxSMglEwCkYEAAA/UDwZQFXX3wAAAABJRU5ErkJggg==","orcid":"","institution":"Fujian Provincial Geriatric Hospital","correspondingAuthor":true,"prefix":"","firstName":"Jingyuan","middleName":"","lastName":"Lin","suffix":""}],"badges":[],"createdAt":"2025-09-25 02:23:19","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7707819/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7707819/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":94728360,"identity":"91c05e20-1fb5-4aa8-a836-f6a38f882341","added_by":"auto","created_at":"2025-10-30 07:03:37","extension":"png","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":56706,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1WABChange.png","url":"https://assets-eu.researchsquare.com/files/rs-7707819/v1/8081d69c30383021461bcc14.png"},{"id":94681545,"identity":"22ae8673-8291-4a77-a0ee-8e7b1c18c91e","added_by":"auto","created_at":"2025-10-29 14:52:14","extension":"doc","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":339966,"visible":true,"origin":"","legend":"","description":"","filename":"COMBIN1.doc","url":"https://assets-eu.researchsquare.com/files/rs-7707819/v1/52a9e59f9cdc826ff28c7bd4.doc"},{"id":94681541,"identity":"8484ee9b-03a5-4442-8d04-fe143756398c","added_by":"auto","created_at":"2025-10-29 14:52:14","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":29447,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2FCHeatmap.png","url":"https://assets-eu.researchsquare.com/files/rs-7707819/v1/e276b58d2711f216db4342cd.png"},{"id":94681547,"identity":"66fd4594-82a0-47f8-9a8f-98cae4beae1b","added_by":"auto","created_at":"2025-10-29 14:52:14","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":124785,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3PredictedvsActual.png","url":"https://assets-eu.researchsquare.com/files/rs-7707819/v1/5027eef41573f59141298fbd.png"},{"id":94728668,"identity":"3e3a4a17-338f-4ef6-852c-8230add73976","added_by":"auto","created_at":"2025-10-30 07:04:09","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":101325,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4LMCoefficientsImproved.png","url":"https://assets-eu.researchsquare.com/files/rs-7707819/v1/24fba4059bd1e6fe76f6bda5.png"},{"id":94681546,"identity":"22629ae8-42c6-4645-bd0a-b9031ba14eb9","added_by":"auto","created_at":"2025-10-29 14:52:14","extension":"json","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3387,"visible":true,"origin":"","legend":"","description":"","filename":"9fd3552fcbe04f039502de65ff27c58c.json","url":"https://assets-eu.researchsquare.com/files/rs-7707819/v1/d9429d3f532615ab4775b37a.json"},{"id":94681540,"identity":"ad6ab223-dc66-498f-bb11-96dcf17c9eef","added_by":"auto","created_at":"2025-10-29 14:52:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":77511,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGroup-wise Changes in WAB-AQ Scores after Intervention\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Figure1WABChange.png","url":"https://assets-eu.researchsquare.com/files/rs-7707819/v1/3937012a65516c947feeaba5.png"},{"id":94681539,"identity":"d2fe2028-b7b8-41dd-80ed-1ac61a8ff271","added_by":"auto","created_at":"2025-10-29 14:52:14","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":29447,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePost-intervention Changes in Resting-State Functional Connectivity among Language-Related Regions (Seed: Left IFG)\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Figure2FCHeatmap.png","url":"https://assets-eu.researchsquare.com/files/rs-7707819/v1/cc88a054d11a6beaa0d9c2db.png"},{"id":94728091,"identity":"53309e08-bdc5-4641-bc05-5411e7c3a89f","added_by":"auto","created_at":"2025-10-30 07:03:04","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":150013,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePredicted vs Actual WAB-AQ Scores in the Random Forest Model\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Figure3PredictedvsActual.png","url":"https://assets-eu.researchsquare.com/files/rs-7707819/v1/dacf6656621c85ed5d433a8e.png"},{"id":94681543,"identity":"9020b8c7-88d2-49e6-a425-85d53c2a0940","added_by":"auto","created_at":"2025-10-29 14:52:14","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":130436,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStandardized Coefficients from the Linear Regression Model Predicting Post-treatment WAB-AQ Scores\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Figure4LMCoefficientsImproved.png","url":"https://assets-eu.researchsquare.com/files/rs-7707819/v1/c9a43ae2966b9bbc6b2a139e.png"},{"id":98062565,"identity":"0a7a93d2-5808-4316-ad12-02dfcc682ed8","added_by":"auto","created_at":"2025-12-12 11:09:47","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1326842,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7707819/v1/d9d51ad7-bf09-4152-a683-1cebf302f9ce.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Combining Transcranial Magnetic Stimulation and Semantic Training to Promote Language Recovery in Aphasia: Evidence from Neural Circuit Remodeling and Machine Learning Prediction","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAphasia is a common and debilitating consequence of stroke, affecting approximately one-third of stroke survivors and significantly impairing their communication ability and quality of life [1]. Conventional speech and language therapy (SLT) remains the mainstay of aphasia rehabilitation, yet many patients show limited recovery, particularly in the chronic phase [2]. Recent advances in neurorehabilitation have focused on augmenting traditional therapy with non-invasive brain stimulation, especially repetitive transcranial magnetic stimulation (rTMS), which has been shown to promote cortical reorganization and enhance neuroplasticity [3,4].\u003c/p\u003e\n\u003cp\u003eHigh-frequency rTMS applied over left-hemispheric language regions, such as the inferior frontal gyrus, may facilitate naming, fluency, and comprehension in patients with non-fluent aphasia [5]. Emerging studies have reported that combining rTMS with structured behavioral interventions, such as semantic training, yields greater gains than either treatment alone [6,7]. Despite these promising results, several limitations remain. First, the underlying mechanism of action\u0026mdash;particularly the neural circuit remodeling associated with recovery\u0026mdash;has not been adequately characterized. Functional MRI studies have begun to reveal changes in resting-state language network connectivity after rTMS interventions, but findings remain fragmented and inconsistent [8].Resting-state fMRI (rs-fMRI) offers a non-invasive method to examine intrinsic functional connectivity without requiring participants to perform language tasks. This is particularly advantageous for individuals with aphasia, who often experience difficulty following task instructions or generating verbal responses during conventional task-based fMRI. As a result, rs-fMRI has become increasingly adopted in post-stroke language recovery research to characterize network-level reorganization.\u003c/p\u003e\n\u003cp\u003eSecond, patient response to combined therapy is highly variable, and the field lacks reliable methods for individualized outcome prediction. Although several machine learning (ML) models, including random forest, support vector machines, and more recently, graph neural networks, have been used to forecast post-stroke aphasia recovery based on behavioral and neuroimaging data, their interpretability and clinical integration remain limited [9,10]. Moreover, many models lack validation in prospective experimental cohorts. There is thus a pressing need to bridge behavioral, neurophysiological, and computational domains to develop interpretable, data-driven tools for personalized rehabilitation [11].\u003c/p\u003e\n\u003cp\u003eIn this study, we address these gaps by examining the therapeutic impact of combined high-frequency rTMS and semantic training on language recovery in patients with chronic post-stroke aphasia. Using resting-state fMRI, we assess functional connectivity changes associated with the intervention. Furthermore, we apply and compare multiple predictive models\u0026mdash;including interpretable linear regression and non-linear ML algorithms\u0026mdash;to identify key predictors of recovery. Our findings aim to contribute to both mechanistic understanding and the development of practical prediction frameworks for individualized aphasia rehabilitation [12].\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003e\u003cstrong\u003e1.Participants\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 100 patients with post-stroke aphasia were prospectively recruited from [Hospital/Clinic Name, anonymized] between [Month, Year] and [Month, Year]. Inclusion criteria were as follows: (1) diagnosis of aphasia confirmed by a certified speech-language pathologist using the Western Aphasia Battery\u0026ndash;Revised (WAB-R) with an Aphasia Quotient (AQ) \u0026lt; 93.8 [13]; (2) single ischemic stroke confirmed by MRI/CT located in the left cerebral hemisphere; (3) stroke onset time \u0026gt; 3 months (chronic stage); (4) age between 45 and 80 years; and (5) right-handedness prior to stroke. Exclusion criteria included history of other neurological or psychiatric diseases, contraindications to MRI or TMS, and severe cognitive deficits (Mini-Mental State Examination score \u0026lt; 24).\u003c/p\u003e\n\u003cp\u003eAll participants provided written informed consent prior to enrollment. The study was approved by the Ethics Committee of [Institution Name], in accordance with the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.Study Design and Grouping\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eParticipants were randomly assigned into three groups using a computer-generated block randomization sequence:\u003c/p\u003e\n\u003cp\u003eTMS + Semantic Training group (n = 40)\u003c/p\u003e\n\u003cp\u003eSham TMS group (n = 30)\u003c/p\u003e\n\u003cp\u003eControl group (standard care only) (n = 30)\u003c/p\u003e\n\u003cp\u003eAll groups received standard medical and nursing care. The intervention lasted for 4 weeks, with 5 sessions per week. Behavioral assessments were conducted at baseline and after the intervention by blinded assessors.\u003c/p\u003e\n\u003cp\u003eParticipants in the Control group received standard post-stroke medical and nursing care, which included general rehabilitation monitoring, medication management, and caregiver education. No structured language therapy or neuromodulation was provided during the intervention period.\u003c/p\u003e\n\u003cp\u003eThe Sham TMS group received sham stimulation but did not undergo semantic training, in order to isolate the effect of neuromodulation. This design allowed us to compare the full combined intervention (TMS + training) against sham stimulation alone, without the confounding influence of behavioral therapy in the sham arm.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.Intervention Protocol\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.1 Transcranial Magnetic Stimulation (TMS)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHigh-frequency repetitive transcranial magnetic stimulation (rTMS) was administered using a Magstim Rapid stimulator equipped with a figure-of-eight coil (70 mm). Stimulation was delivered at 10 Hz, with 1000 pulses per session, at an intensity of 90% of the individual\u0026rsquo;s resting motor threshold (RMT). The RMT was determined by identifying the minimum stimulus intensity that produced a motor evoked potential (MEP) of at least 50\u0026nbsp;\u0026mu;V in the contralateral abductor pollicis brevis muscle in 5 out of 10 consecutive trials.\u003c/p\u003e\n\u003cp\u003eThe stimulation target was the left inferior frontal gyrus (IFG; approximately Brodmann area 44/45), localized for each participant using high-resolution T1-weighted structural MRI and coregistered to the Montreal Neurological Institute (MNI) space. Individualized targeting was conducted using Brainsight neuronavigation software (Rogue Research Inc., Montreal, Canada), ensuring accurate coil placement over the peak voxel within the IFG based on the AAL atlas. The coil was oriented tangentially to the scalp with the handle positioned posteriorly to induce a posterior-to-anterior current flow.\u003c/p\u003e\n\u003cp\u003eStimulation sessions were delivered once daily, five days per week, for a total of four consecutive weeks.\u003c/p\u003e\n\u003cp\u003eFor the sham TMS group, identical coil positioning and session parameters were used; however, a sham coil that emitted no magnetic field was applied, ensuring blinding while mimicking the auditory and tactile sensations of active stimulation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2 Semantic Training\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSemantic therapy was administered individually by certified speech-language pathologists for 45 minutes immediately following each TMS or sham session. The training followed a structured, evidence-based protocol adapted from validated aphasia rehabilitation programs [14], and was designed to target lexical retrieval and semantic processing.\u003c/p\u003e\n\u003cp\u003eEach session consisted of three sequential modules:\u003c/p\u003e\n\u003cp\u003eConfrontation Naming: Patients were presented with standardized picture stimuli (e.g., Snodgrass and Vanderwart image set) and asked to name each item. Semantic and phonemic cues were provided as needed, along with corrective feedback.\u003c/p\u003e\n\u003cp\u003eSemantic Feature Analysis (SFA): For selected target words, patients identified semantic attributes such as category, function, physical characteristics, location, and associations. This method aimed to strengthen semantic-lexical linkages and promote generalization to untrained items.\u003c/p\u003e\n\u003cp\u003eCategory-based Word Retrieval: Patients engaged in rapid lexical generation tasks within predefined semantic categories (e.g., animals, tools, fruits) under timed conditions to facilitate semantic network activation.\u003c/p\u003e\n\u003cp\u003eTasks were individualized based on the patient\u0026apos;s baseline performance (WAB-AQ and semantic fluency scores), with progressive difficulty adjustments across sessions. Therapists employed error-reducing techniques such as cueing hierarchies, repetition, modeling, and written supports to maximize patient engagement and success. Session performance was documented using structured logs to guide adaptation and monitor progress.\u003c/p\u003e\n\u003cp\u003eParticipants in the control group did not receive any additional language intervention and continued with standard clinical care only.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.Behavioral Assessments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLanguage performance was evaluated using the WAB-R, with primary outcomes focused on AQ change scores. Semantic fluency was measured by the number of correct words generated in 1 minute under a given category (e.g., animals). Assessments were conducted pre- and post-intervention.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.Imaging Acquisition and Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eResting-state functional MRI (rs-fMRI) data were acquired using a 3.0 Tesla Siemens Prisma scanner with a 64-channel head coil. Functional images were obtained with the following parameters: echo-planar imaging (EPI) sequence, repetition time (TR) = 2000 ms, echo time (TE) = 30 ms, flip angle = 90\u0026deg;, slice thickness = 3 mm, voxel size = 3\u0026times;3\u0026times;3 mm\u0026sup3;, 33 axial slices, and 240 volumes per run. High-resolution T1-weighted anatomical images were also acquired for registration purposes.\u003c/p\u003e\n\u003cp\u003ePreprocessing was performed using the CONN toolbox v21a in conjunction with SPM12, running on MATLAB R2022b. The preprocessing pipeline included: slice timing correction, realignment for head motion correction, coregistration to individual T1 images, segmentation and normalization to MNI152 space, spatial smoothing with a 6 mm FWHM Gaussian kernel, and nuisance regression (including white matter, CSF signals, and motion parameters using the aCompCor method).\u003c/p\u003e\n\u003cp\u003eSubjects with head motion exceeding 2 mm of translation or 2\u0026deg;\u0026nbsp;of rotation were excluded from analysis. Framewise displacement was also calculated, and volumes with excessive motion (\u0026gt;0.5 mm) were scrubbed using artifact detection tools (ART) within CONN.\u003c/p\u003e\n\u003cp\u003eFunctional connectivity (FC) was computed using a seed-to-voxel correlation analysis. The seed region was defined as the left inferior frontal gyrus (IFG; Brodmann area 44/45), based on the AAL atlas. Correlation maps were generated by calculating Pearson\u0026rsquo;s correlation coefficients between the seed\u0026rsquo;s time series and all other voxels in the brain, followed by Fisher\u0026apos;s r-to-z transformation for group-level analysis. Post-hoc ROI-to-ROI analysis was conducted to examine changes in connectivity with language-related regions including the middle temporal gyrus (MTG), angular gyrus (AG), and supramarginal gyrus (SMG).\u003c/p\u003e\n\u003cp\u003eResting-state fMRI was acquired only for the TMS+Semantic group, as the primary goal was to evaluate neural mechanisms specifically associated with the combined intervention. Imaging was not conducted in the Sham or Control groups due to resource constraints and to minimize participant burden in non-intervention arms.\u003c/p\u003e\n\u003cp\u003eFor visualization purposes, lesion overlay maps were generated by manually tracing lesions on structural T1 images and warping them to MNI space. However, lesion maps were not used for voxel-wise group-level analysis, and only binary IFG involvement (yes/no) was included as a covariate in the regression models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e6. Statistical Analysis\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eStatistical analyses were performed using SPSS version 26 and R version 4.3.1. Baseline demographic and clinical characteristics were compared using one-way ANOVA for continuous variables and chi-square tests for categorical variables.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo assess treatment effects, repeated-measures ANOVA was conducted with time (pre- vs. post-intervention) as a within-subject factor and group (TMS+Semantic, Sham TMS, Control) as a between-subject factor. Post hoc pairwise comparisons were adjusted using the Bonferroni correction. Effect sizes were reported using partial eta squared (\u0026eta;\u0026sup2;).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo predict post-intervention language outcomes, multiple linear regression was applied with the following predictors: baseline WAB-AQ score, semantic fluency, age, years of education, weeks post-stroke, and intervention group (dummy-coded). Variance inflation factors (VIFs) confirmed the absence of multicollinearity. Standardized beta coefficients were reported. Baseline WAB-AQ and semantic fluency were also entered as covariates to adjust for initial language status.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Lesion analysis was performed using high-resolution structural MRI. Patients with extensive bilateral or multifocal lesions were excluded. For those with partial damage to the left inferior frontal gyrus (IFG), lesion overlap was coded as a binary variable (yes/no) based on blinded manual tracing by a neuroradiologist, and included as a covariate in the regression models.\u003c/p\u003e\n\u003cp\u003eFurthermore, to ensure stimulation target integrity, patients with severe IFG structural disruption were excluded a priori to minimize confounding of TMS efficacy. \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMachine Learning Modeling Machine learning models were implemented using scikit-learn (v1.3.0) and XGBoost (v1.7.6) in Python. The input feature set included:\u003c/p\u003e\n\u003cp\u003eBaseline WAB-AQ score\u003c/p\u003e\n\u003cp\u003eSemantic fluency\u003c/p\u003e\n\u003cp\u003eAge\u003c/p\u003e\n\u003cp\u003eYears of education\u003c/p\u003e\n\u003cp\u003eWeeks post-stroke\u003c/p\u003e\n\u003cp\u003eGroup assignment (dummy-coded) \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe target variable was post-treatment WAB-AQ score. All features were standardized prior to model fitting. For model evaluation, we used 5-fold cross-validation, in which 80% of the data were used for training and 20% for testing in each fold. Performance metrics (R\u0026sup2;, RMSE) were averaged across folds. Although leave-one-out cross-validation (LOOCV) was considered, 5-fold CV was selected for better balance between computational efficiency and model stability, given the sample size (N = 100).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eHyperparameter tuning was performed using nested grid search within the training set of each fold. Specifically:\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eLinear Regression used default settings (no regularization)\u003c/p\u003e\n\u003cp\u003eSVR tuned C and epsilon\u003c/p\u003e\n\u003cp\u003eRandom Forest tuned n_estimators and max_depth\u003c/p\u003e\n\u003cp\u003eXGBoost optimized learning_rate, max_depth, and n_estimators \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis pipeline ensured method transparency, cross-validated performance estimation, and model reproducibility.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7. data statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis was a retrospective observational study based on anonymized clinical and imaging data collected during routine care.Although the groups were pre-defined as part of clinical service protocols, no formal prospective randomization was performed.Importantly, none of the data reported in this manuscript have been previously published, and the analyses were conducted specifically for this study.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003e1.Baseline Characteristics of the Participants\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTable 1 presents the baseline demographic and clinical characteristics of the three participant groups. No significant differences were observed among the Control, Sham TMS, and TMS+Semantic groups in terms of age (mean \u0026asymp; 65\u0026ndash;66 years), years of education (mean \u0026asymp; 11.3), weeks post-onset (\u0026asymp; 18 weeks), WAB-AQ scores (\u0026asymp; 46), or semantic fluency performance (\u0026asymp; 5.7\u0026ndash;5.8). These findings suggest a well-balanced allocation across groups prior to intervention.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1. Baseline Demographic and Clinical Characteristics of the Three Groups\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003eGroup\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003eAge (Mean \u0026plusmn; SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003eEducation (Years)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003eWeeks Post-Stroke\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003eWAB-AQ (Pre)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003eSemantic Fluency (Pre)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003eControl\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e65.26 \u0026plusmn; 6.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e11.33 \u0026plusmn; 2.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e18.27 \u0026plusmn; 3.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e45.63 \u0026plusmn; 5.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e5.80 \u0026plusmn; 1.24\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003eSham TMS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e66.24 \u0026plusmn; 5.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e11.43 \u0026plusmn; 2.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e18.18 \u0026plusmn; 4.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e45.98 \u0026plusmn; 5.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e5.70 \u0026plusmn; 1.36\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003eTMS+Semantic\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e65.75 \u0026plusmn; 6.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e11.38 \u0026plusmn; 2.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e17.75 \u0026plusmn; 3.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e46.40 \u0026plusmn; 5.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e5.84 \u0026plusmn; 1.30\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003col start=\"2\"\u003e\n \u003cli\u003e\u003cstrong\u003eLanguage Function Improvemt\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eFigure 1 illustrates the change in Western Aphasia Battery\u0026ndash;Aphasia Quotient (WAB-AQ) scores from pre- to post-intervention across the three groups. The TMS+Semantic group showed a markedly greater improvement in WAB-AQ (median\u0026nbsp;\u0026asymp;\u0026nbsp;12) compared to both the Sham TMS and Control groups (median\u0026nbsp;\u0026asymp;\u0026nbsp;6 and 3, respectively). Statistical analysis revealed significant between-group differences (p \u0026lt; 0.001), suggesting that the combined intervention was more effective in promoting language recovery.Notably, no participants in any group demonstrated a decline in WAB-AQ scores following the intervention. All patients either improved or maintained their baseline language function. Group-wise differences in AQ change scores were assessed using repeated-measures ANOVA, with post hoc pairwise comparisons adjusted via the Bonferroni method to control for multiple comparisons. These analyses confirmed that the TMS+Semantic group exhibited significantly greater gains than both the Sham TMS and Control groups (p \u0026lt; 0.001).\u003c/p\u003e\n\u003cp\u003eTo enhance transparency, individual-level pre- and post-intervention scores can be provided in an appendix or supplementary material upon request.\u003c/p\u003e\n\u003col start=\"3\"\u003e\n \u003cli\u003e\u003cstrong\u003eResting-State Functional Connectivity Remodeling after Intervention\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eFigure 2 displays a heatmap of changes in resting-state functional connectivity (rsFC) among core language-related brain regions in the TMS+Semantic group, using the left inferior frontal gyrus (IFG_L) as the seed. Warmer colors (red-orange) indicate increased connectivity post-intervention, while cooler colors (blue) indicate decreased connectivity. Notably, enhanced connectivity was observed between IFG_L and left SMG (supramarginal gyrus), MTG (middle temporal gyrus), and AG (angular gyrus), suggesting neuroplastic remodeling within the left-lateralized language network following the combined intervention.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.\u003c/strong\u003e\u003cstrong\u003ePredictive Modeling of Language Recovery Using Random Forest Regression\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 3 illustrates the relationship between predicted and actual post-treatment WAB-AQ scores based on a random forest regression model. Each dot represents an individual subject. The fitted regression line (blue) indicates a positive correlation between predicted and observed values, with an R\u0026sup2; of 0.44 and RMSE of 3.91. These findings suggest that the model captures meaningful variance in language outcomes and may serve as a useful tool for individualized prognosis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.Comparison of Predictive Performance across Machine Learning Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTable 2 summarizes the predictive performance of four machine learning models\u0026mdash;Linear Regression, Support Vector Machine (SVM), Random Forest, and XGBoost\u0026mdash;for estimating post-intervention WAB-AQ scores. Among the models, Linear Regression achieved the highest R\u0026sup2; value (0.47), indicating it explained the most variance in the outcome. Random Forest followed closely (R\u0026sup2; = 0.44) and had the lowest RMSE (3.91), suggesting strong accuracy. In contrast, SVM demonstrated the poorest performance (R\u0026sup2; = 0.20, RMSE = 4.73). These results suggest that while nonlinear models offer certain advantages, linear regression remains competitive in both accuracy and interpretability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Performance Metrics of Machine Learning Models for Predicting Post-treatment WAB-AQ Scores\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003eR2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003eRMSE\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003eLinear Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003e0.47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003e3.96\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003e0.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003e4.73\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003e0.44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003e3.91\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003e0.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 182px;\"\u003e\n \u003cp\u003e4.20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cstrong\u003e6.Interpretable Regression Analysis Reveals Predictors of Language Recovery\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 4 presents the standardized beta coefficients from the linear regression model predicting post-treatment WAB-AQ scores. Pre-intervention WAB-AQ score was the strongest positive predictor (\u0026beta; = 1.32, p \u0026lt; 0.01), followed by semantic fluency (\u0026beta; = 0.46) and years of education (\u0026beta; = 0.15). Age showed a minimal effect (\u0026beta; = 0.07). Notably, weeks from stroke onset to intervention was negatively associated with outcome (\u0026beta; = \u0026ndash;0.40), suggesting that earlier intervention may contribute to better recovery. These results support the utility of baseline behavioral and demographic data for individual prognosis.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003e\u003cstrong\u003e1. Summary of Findings\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn this study, we evaluated the combined effects of high-frequency transcranial magnetic stimulation (TMS) and semantic training on language recovery in chronic post-stroke aphasia. Our results revealed that participants receiving the combined intervention showed significantly greater improvement in language function, as measured by the Western Aphasia Battery\u0026ndash;Aphasia Quotient (WAB-AQ), compared to both the Sham TMS and Control groups. The efficacy of the combined intervention appears to be underpinned by neuroplastic changes in resting-state functional connectivity, particularly within left-hemispheric language circuits. Additionally, we found that machine learning models, particularly linear regression and random forest, effectively predicted treatment outcomes using pre-treatment behavioral and demographic data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2. Efficacy of Combined TMS and Semantic Training\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOur results demonstrate that patients in the TMS+Semantic group exhibited a median WAB-AQ gain of approximately 12 points, surpassing the improvements seen in the Sham TMS and Control groups. These findings reinforce previous studies indicating that TMS can potentiate behavioral interventions when applied over perilesional cortical areas such as the inferior frontal gyrus (IFG) [16,17]. The synergistic effect is likely due to the capacity of TMS to modulate cortical excitability, thereby creating a more receptive neurophysiological environment for targeted semantic training [18]. Prior research has shown that high-frequency TMS delivered to the left IFG enhances lexical retrieval and naming performance [19,20], effects that appear to be magnified when combined with behavioral stimulation strategies [21].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3. Neuroplastic Remodeling in Language Networks\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFunctional neuroimaging provided mechanistic support for the observed clinical improvements. Specifically, seed-based analysis revealed enhanced resting-state connectivity between the left IFG and posterior temporal-parietal regions, including the supramarginal gyrus (SMG), middle temporal gyrus (MTG), and angular gyrus (AG). These regions are known to contribute to phonological working memory, semantic processing, and lexical access [22]. Enhanced synchronization within this left-lateralized network suggests that the intervention may restore or reinforce functional circuits disrupted by stroke. These observations align with fMRI studies reporting similar patterns of circuit reinstatement following language rehabilitation [23,24]. Connectivity changes were observed only in the TMS+Semantic group, as resting-state imaging was not performed in the Sham or Control groups. Therefore, direct group-level comparisons are not available.\u003c/p\u003e\n\u003cp\u003eThe selection of these regions was informed by the dual-stream model of language processing and supported by previous studies on aphasia recovery. The left IFG, serving as the stimulation target and seed region, plays a crucial role in speech production and lexical retrieval. The MTG contributes to semantic comprehension, the AG is associated with lexical-semantic integration and reading, and the SMG is involved in phonological processing and verbal working memory. These areas are known to functionally interact during language tasks and are commonly disrupted in post-stroke aphasia. By evaluating connectivity between these regions, we aimed to capture potential neuroplastic changes in core components of the language network.\u003c/p\u003e\n\u003cp\u003eIn this study, functional connectivity (FC) was operationally defined as the temporal correlation of blood oxygen level-dependent (BOLD) signal fluctuations between anatomically distinct brain regions during resting-state fMRI. Specifically, we focused on seed-based FC analyses using the left inferior frontal gyrus (IFG) as the reference region, in line with prior models of language processing.\u003c/p\u003e\n\u003cp\u003eThe increased FC observed between the IFG and posterior perisylvian areas\u0026mdash;such as the middle temporal gyrus (MTG), angular gyrus (AG), and supramarginal gyrus (SMG)\u0026mdash;is interpreted as evidence of neuroplastic remodeling within the left-hemispheric language network. Such remodeling likely reflects enhanced integration or synchronization of residual linguistic pathways, a hypothesis supported by both dual-stream models of language and previous aphasia rehabilitation studies.\u003c/p\u003e\n\u003cp\u003eRegarding the interpretability of our regression model, we emphasize that standardized beta coefficients in the linear regression framework offer insight into the relative contribution of each predictor variable to post-treatment language outcomes. While more complex models like random forests may yield similar or even superior predictive accuracy, their internal mechanisms (e.g., feature interaction and decision splitting) are often opaque. Thus, we highlight the advantage of linear models in clinical settings where explainability is paramount, especially when aiming to inform individualized rehabilitation planning.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4. Predictive Modeling and Clinical Utility\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA novel contribution of our study lies in the integration of machine learning for prognostic modeling. Among four models tested\u0026mdash;linear regression, random forest, XGBoost, and SVM\u0026mdash;linear regression yielded the best predictive performance (R\u0026sup2; = 0.47, RMSE = 3.96). Despite the simplicity of this model, its interpretability makes it attractive for clinical use [25]. Random forest achieved slightly lower R\u0026sup2; (0.44) but had the lowest RMSE (3.91), suggesting strong generalization capacity. By contrast, SVM performed poorly, likely due to its sensitivity to small sample sizes and limited feature dimensionality [26]. These findings mirror recent literature emphasizing the balance between model performance and interpretability in clinical prediction settings [27,28].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5. Key Predictors of Recovery\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eRegression coefficient analysis revealed that pre-treatment WAB-AQ score was the most robust positive predictor of language recovery, followed by semantic fluency and years of education. These results reinforce the concept of residual function and cognitive reserve as facilitators of recovery [29]. Interestingly, time since stroke onset was negatively associated with outcome, indicating that earlier initiation of neuromodulatory therapy may confer greater benefit\u0026mdash;a finding echoed in meta-analyses suggesting a time-dependent plasticity window post-stroke [30].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e6. Limitations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWhile our study presents compelling evidence for the combined use of TMS and semantic training, several limitations warrant discussion. First, the study cohort was modest in size (n = 100) and did not stratify patients by aphasia subtype. Different aphasia phenotypes may respond differently to neuromodulation and behavioral training, and future studies should address this heterogeneity. Second, the follow-up period was limited to immediate post-treatment outcomes. Longitudinal follow-up would be essential to determine the durability of treatment effects. Third, although our fMRI results suggest neural circuit remodeling, causality cannot be definitively inferred, and more advanced imaging techniques (e.g., task-based fMRI or diffusion tensor imaging) could offer complementary insights. Finally, external validation of the predictive models in independent cohorts is needed before clinical translation.\u003c/p\u003e\n\u003cp\u003eAnother key limitation of this study is the inability to disentangle the individual contributions of rTMS and semantic training, as both interventions were administered concurrently in the TMS+Semantic group. As such, it remains unclear whether the observed effects were driven primarily by neuromodulation, behavioral therapy, or their synergistic interaction. Future studies using factorial designs or additional control arms are warranted to isolate these components.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Ethics Committee of Fujian Provincial Geriatric Hospital in June 2025(code 20250801). The requirement for individual informed consent was waived due to the use of anonymized clinical and imaging data. All data were collected as part of routine clinical care in accordance with the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u0026nbsp;\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe datasets generated and analyzed during the current study are not publicly available due to privacy restrictions but are available from the corresponding author upon reasonable request and with appropriate institutional approvals.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding \u0026nbsp;\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research received no external funding.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; contributions \u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJingyuan Lin conceived and designed the study, collected and analyzed the data, interpreted the results, drafted and revised the manuscript. The author read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements \u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTrial registration\u003c/strong\u003e\u003cbr\u003e\u0026nbsp;This study is not a registered clinical trial, as it is a retrospective observational study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eEngelter ST, et al. Epidemiology of aphasia attributable to first ischemic stroke. Stroke. 2006;37(6):1379-1384. https://doi.org/10.1161/01.STR.0000221815.64093.8c\u003c/li\u003e\n \u003cli\u003eBrady MC, et al. Speech and language therapy for aphasia following stroke. Cochrane Database Syst Rev. 2016;(6):CD000425. https://doi.org/10.1002/14651858.CD000425.pub4\u003c/li\u003e\n \u003cli\u003eBarwood CH, et al. Low frequency rTMS and language recovery in non-fluent aphasia. NeuroRehabilitation. 2011;28(2):113\u0026ndash;128. https://doi.org/10.3233/NRE-2011-0640\u003c/li\u003e\n \u003cli\u003eWang CP, et al. Synchronous verbal training and rTMS in chronic aphasia. Stroke. 2014;45(12):3656\u0026ndash;3662. https://doi.org/10.1161/STROKEAHA.114.007058\u003c/li\u003e\n \u003cli\u003eHarvey DY, et al. TMS-induced naming improvement in aphasia. Cognitive Behav Neurol. 2017;30(4):133\u0026ndash;144. https://doi.org/10.1097/WNN.0000000000000141\u003c/li\u003e\n \u003cli\u003eGan L, et al. rTMS combined with speech-language therapy for Broca\u0026rsquo;s aphasia. Front Neurol. 2024;15:1473254. https://doi.org/10.3389/fneur.2024.1473254\u003c/li\u003e\n \u003cli\u003eFridriksson J, et al. rTMS improves naming in post-stroke aphasia. Stroke. 2018;49(1):139\u0026ndash;145. https://doi.org/10.1161/STROKEAHA.117.018775\u003c/li\u003e\n \u003cli\u003eZhao Y, et al. Predicting language recovery with fMRI. Sci Rep. 2021;11(1):1\u0026ndash;10. https://doi.org/10.1038/s41598-021-88022-z\u003c/li\u003e\n \u003cli\u003eChen Z, et al. Lesion-aware graph neural network for aphasia prediction. arXiv. 2024. https://doi.org/10.48550/arXiv.2409.02303\u003c/li\u003e\n \u003cli\u003eThompson CK, et al. Neural correlates of sentence processing. J Cogn Neurosci. 2007;19(11):1753\u0026ndash;1767. https://doi.org/10.1162/jocn.2007.19.11.1753\u003c/li\u003e\n \u003cli\u003eSmall SL, Llano DA. Biological approaches to aphasia treatment. Handb Clin Neurol. 2009;93:459\u0026ndash;470. https://doi.org/10.1016/S0072-9752(09)93036-1\u003c/li\u003e\n \u003cli\u003eKapoor A. rTMS therapy for non-fluent aphasia: a critical review. Top Stroke Rehabil. 2017;24(8):597\u0026ndash;603. https://doi.org/10.1080/10749357.2017.1331417\u003c/li\u003e\n \u003cli\u003eKertesz A. The Western Aphasia Battery\u0026ndash;Revised (WAB-R). San Antonio, TX: Pearson; 2006. https://doi.org/10.1037/t15174-000\u003c/li\u003e\n \u003cli\u003eEdmonds LA, et al. Effect of semantic feature analysis on naming and discourse in aphasia: A single-subject study. Aphasiology. 2009;23(3):305\u0026ndash;328. https://doi.org/10.1080/02687030802130919\u003c/li\u003e\n \u003cli\u003eTzourio-Mazoyer N, et al. Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. Neuroimage. 2002;15(1):273\u0026ndash;289. https://doi.org/10.1006/nimg.2001.0978\u003c/li\u003e\n \u003cli\u003eWang CP, et al. Efficacy of synchronous verbal training during repetitive transcranial magnetic stimulation in patients with chronic aphasia. Stroke. 2014;45(12):3656\u0026ndash;62. https://doi.org/10.1161/STROKEAHA.114.007058\u003c/li\u003e\n \u003cli\u003eBarwood CH, et al. Low-frequency rTMS and language recovery in non-fluent aphasia. NeuroRehabilitation. 2011;28(2):113\u0026ndash;28. https://doi.org/10.3233/NRE-2011-0640\u003c/li\u003e\n \u003cli\u003eKapoor A. rTMS therapy for non-fluent aphasia: a critical review. Top Stroke Rehabil. 2017;24(8):597\u0026ndash;603. https://doi.org/10.1080/10749357.2017.1331417\u003c/li\u003e\n \u003cli\u003eFridriksson J, et al. rTMS improves naming in post-stroke aphasia. Stroke. 2018;49(1):139\u0026ndash;45. https://doi.org/10.1161/STROKEAHA.117.018775\u003c/li\u003e\n \u003cli\u003eHarvey DY, et al. TMS-induced naming improvement in aphasia. Cogn Behav Neurol. 2017;30(4):133\u0026ndash;44. https://doi.org/10.1097/WNN.0000000000000141\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eGan L, et al. rTMS combined with speech-language therapy for Broca\u0026rsquo;s aphasia. Front Neurol. 2024;15:1473254. https://doi.org/10.3389/fneur.2024.1473254\u003c/li\u003e\n \u003cli\u003ePrice CJ. The anatomy of language: a review of 100 fMRI studies. Brain. 2000;123(3):409\u0026ndash;30. https://doi.org/10.1093/brain/123.3.409\u003c/li\u003e\n \u003cli\u003evan Hees S, et al. Resting-state connectivity and naming therapy outcomes in poststroke aphasia. Brain Lang. 2014;129:1\u0026ndash;10. https://doi.org/10.1016/j.bandl.2013.12.001\u003c/li\u003e\n \u003cli\u003eMeinzer M, et al. Neural signatures of recovery from aphasia after therapy. Neurology. 2012;78(3):150\u0026ndash;8. https://doi.org/10.1212/WNL.0b013e31823efc0d\u003c/li\u003e\n \u003cli\u003eKiran S, et al. Machine learning-based prediction of language outcomes in chronic aphasia. Hum Brain Mapp. 2021;42(5):1682\u0026ndash;94. https://doi.org/10.1002/hbm.25324\u003c/li\u003e\n \u003cli\u003eCortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20:273\u0026ndash;97. https://doi.org/10.1007/BF00994018\u003c/li\u003e\n \u003cli\u003eEsteva A, et al. A guide to deep learning in healthcare. Nat Med. 2019;25:24\u0026ndash;9. https://doi.org/10.1038/s41591-018-0316-z\u003c/li\u003e\n \u003cli\u003eChen Z, et al. Lesion-aware edge-based graph neural network for predicting language ability in poststroke aphasia. arXiv preprint. 2024. https://doi.org/10.48550/arXiv.2409.02303\u003c/li\u003e\n \u003cli\u003eSmall SL, Llano DA. Biological approaches to the treatment of aphasia. Handb Clin Neurol. 2009;93:459\u0026ndash;70. https://doi.org/10.1016/S0072-9752(09)93036-1\u003c/li\u003e\n \u003cli\u003eSebastianelli L, et al. Low-frequency rTMS of the unaffected hemisphere in stroke: A systematic review. Acta Neurol Scand. 2017;136(6):585\u0026ndash;605. https://doi.org/10.1111/ane.12773\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Aphasia, Transcranial Magnetic Stimulation, Semantic Training, Language Recovery, Functional Connectivity, Neural Circuit Remodeling, Machine Learning","lastPublishedDoi":"10.21203/rs.3.rs-7707819/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7707819/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e: Aphasia is a frequent and debilitating outcome of stroke, often persisting into the chronic stage and significantly affecting communication ability. While transcranial magnetic stimulation (TMS) and semantic training have each demonstrated therapeutic benefits, their synergistic effects and underlying mechanisms remain to be fully elucidated. This study aimed to examine the efficacy of high-frequency TMS combined with semantic training in improving language function in patients with post-stroke aphasia, and to explore neural connectivity changes and predictive modeling of recovery outcomes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e: The TMS plus semantic training group showed significantly greater improvement in WAB-AQ compared to both the sham and control groups. Resting-state fMRI revealed enhanced connectivity between the left IFG and posterior temporal-parietal regions post-intervention. Among the predictive models, linear regression achieved the best performance (R²= 0.47, RMSE = 3.96), followed closely by random forest (R²= 0.44, RMSE = 3.91), while SVM and XGBoost performed less optimally.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e: Combined TMS and semantic therapy effectively enhances language recovery in chronic aphasia, likely through remodeling of left-hemispheric language circuits. Furthermore, regression-based models show promise in predicting treatment outcomes and may inform individualized rehabilitation strategies.\u003c/p\u003e","manuscriptTitle":"Combining Transcranial Magnetic Stimulation and Semantic Training to Promote Language Recovery in Aphasia: Evidence from Neural Circuit Remodeling and Machine Learning Prediction","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-29 14:52:09","doi":"10.21203/rs.3.rs-7707819/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a7b02483-5d15-48d7-943f-707f0e51ae05","owner":[],"postedDate":"October 29th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-12-12T11:08:31+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-29 14:52:09","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7707819","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7707819","identity":"rs-7707819","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}