Functional connectivity associated with severe upper limb impairment in resting-state electroencephalography among chronic stroke survivors: a machine learning approach Authors: Ji-Yoon Lee, Miseon Shim, Won Kee Chang, Hee-Mun Cho, Ji-Soo Choi, Hyunji Kim, Bongwon Suh, Nam-Jong Paik, Han-Jeong Hwang, Won-Seok Kim Published: 2025 ABSTRACT BackgroundSevere upper limb impairment (ULI) presents a significant challenge in the rehabilitation of chronic stroke survivors and affects their quality of life. Identifying biomarkers and understanding the neural mechanisms associated with severe ULI are essential for evaluating recovery potential and enhancing rehabilitation effectiveness. This study aimed to identify resting-state electroencephalography (EEG) functional connectivity features associated with severe ULI in chronic stroke survivors using machine learning (ML) methods.MethodsEEG data were collected from 34 chronic stroke survivors. Participants were categorized into two groups based on their Fugl–Meyer Assessment for Upper Extremity (FMA-UE) scores: a mild/moderate ULI (FMA-UE ≥ 30; n = 19) and severe ULI (FMA-UE < 30; n = 15). We employed ML algorithms to classify severe ULI, including logistic regression with L1, elastic net regularization, stochastic gradient descent, and support vector machines, along with several feature selection methods. Coherence was evaluated across six frequency bands in both the ipsilesional (affected by the lesion) and contralesional (opposite side of the lesion) hemispheres.ResultsThe logistic regression model with L1 and ReliefF feature selection methods was the most effective, achieving a balanced accuracy of 0.91 (sensitivity = 0.93; specificity = 0.90). This approach identified 14 significant features for distinguishing severe ULI from mild to moderate ULI, including delta interhemispheric and intrahemispheric connectivity in the frontal, parietal, and temporal regions. Additionally, interhemispheric and intrahemispheric theta connectivity was observed in the prefrontal, frontal, temporal, and parietal regions. Low-beta intrahemispheric connectivity was also observed in the contralesional parietal regions.ConclusionsOur research highlights the association between alterations in connectivity within low-frequency bands and severe ULI across widespread brain regions, including areas outside the sensorimotor cortex and bilateral intrahemispheric and interhemispheric regions. Further research utilizing larger longitudinal datasets from early stroke survivors employing ML approaches could contribute to the development of more accurate predictive models for motor recovery and rehabilitation responses.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12984-025-01859-7. Background Stroke is a major health concern due to its high prevalence and substantial impact on disability [1]. According to the World Health Organization, over 60% of chronic stroke survivors continue to experience upper limb impairment (ULI) [2]. Improving upper limb function is a key rehabilitation goal; however, individuals with severe ULI face substantial challenges in achieving optimal outcomes [3]. The impact of severe ULI on survivors’ quality of life is particularly significant [4]. Previous research emphasizes the importance of identifying biomarkers to develop targeted and effective interventions for stroke survivors with severe ULI [5, 6]. Electroencephalography (EEG) is advantageous for biomarker development because of its noninvasive nature, excellent temporal resolution, and cost-effectiveness among neuroimaging methods [7]. In a scoping review conducted by Milani et al. [8], EEG-derived measures―such as power spectral density (PSD), symmetry in brain dynamics, and functional connectivity―proved valuable for evaluating upper limb recovery. However, most studies have concentrated on features related to PSD; a notable gap remains in understanding how different brain regions connect, process, and share information in stroke survivors [9, 10]. Understanding these connections is crucial because functional connectivity can serve as a valuable marker for monitoring motor recovery. Defined as the statistical interdependencies between physiological time series from different brain regions, functional connectivity plays a crucial role in brain function [11]. Furthermore, the brain remains active even during rest, displaying spontaneous activity that is not linked to external stimuli or responses but arises internally [12]. Therefore, resting-state functional connectivity shows promise for developing biomarkers to monitor motor recovery poststroke [6, 13, 14]. Carter et al. [15] emphasized that stroke recovery depends on the reorganization of the brain’s cortical and subcortical networks, not just during active tasks but also in a resting state. Therefore, resting-state connectivity may serve as a potential biomarker for motor impairment following stroke [16–18]. Additionally, the neural mechanisms underlying recovery after stroke, which affect the ipsilesional and contralesional brain hemispheres, are not yet fully understood [19]. Acute stroke survivors exhibit altered EEG interhemispheric connectivity and modifications in the balance between local segregation and global integration [20]. Employing a hemispheric-focused network analysis is crucial for deeply understanding cost-effective neuronal compensatory mechanisms in stroke survivors [21]. While many studies have concentrated on the connectivity of specific brain regions [22], understanding of comprehensive inter- and intra-hemispheric connectivity patterns across all brain regions remains limited. EEG coherence, which indicates functional connectivity between different electrode sites, reflects the degree of signal synchronization and suggests functional interaction [23]. The complexity of EEG coherence, characterized by the complexity, multichannel, multifrequency, and inherent complexity of brain function and structure, introduces a challenge in comprehensively analyzing resting-state EEG connectivity. For example, Wu et al. [18] demonstrated that beta coherence in the primary motor area was significantly correlated with post-stroke FMA scores (r² = 0.96), suggesting that coherence could serve as a strong biomarker for post-stroke recovery. However, their investigation was limited to specific brain regions, particularly motor areas, due to their hypothesis-driven approach, which restricted the generalizability of the findings to broader aspects of brain function beyond motor control. To overcome these the limitations of the hypothesis-driven approach, recent studies have adopted a data-driven approach, enabling a more comprehensive exploration of the role of EEG coherence in motor impairment after stroke [16, 24, 25]. In particular, ML is widely applied to identify meaningful EEG-based connectivity features while minimizing computational costs and time by mitigating feature redundancies of features caused by the high dimensionality and volume of EEG data [26]. Indeed, Lassi et al. [25] demonstrated the effective use of ML and feature-selection methods to classify ULI severity in acute stroke survivors based on a data-driven approach. Their best model, based on a support vector machine (SVM), achieved over 85% accuracy. Therefore, it is necessary to identify higher-order brain features based on connectivity that emerges during information processing, rather than relying solely on simple brain activation patterns derived from data-driven approaches. This study aimed to identify resting-state EEG connectivity patterns related to ULI severity in chronic stroke survivors, as indicated by the coherence between channels in the ipsilesional and contralesional hemispheres, using a data-driven approach. To this end, we attempted to determine the best ML model for classifying the severity of ULI—severe vs. mild/moderate—and then identify the EEG connectivity features critical for distinguishing ULI severity within the model. We considered that adopting a data-driven approach could further enhance the generalizability of the findings, as it enables the identification of important coherence features without being constrained by prior hypotheses. Methods Participants Thirty-four stroke survivors, aged between 29 and 80, were recruited for the study from two rehabilitation hospitals between June and December 2019. Participants were selected based on the following criteria: (1) diagnosis of first-ever ischemic or hemorrhagic stroke confirmed by brain computed tomography or magnetic resonance imaging, (2) unilateral upper limb weakness, (3) a minimum of six months post-stroke, and (4) the ability to provide written informed consent. Participants were excluded if they met any of the following conditions: (1) a history of central nervous system conditions such as traumatic brain injury, brain tumor, or Parkinson’s disease; (2) inability to wear an EEG cap; or (3) difficulty comprehending and adhering to clinical assessment and EEG study protocols. A clinician reviewed the participant’s structural images and classified their stroke locations into three categories: cortical, subcortical, including corona radiata, internal capsule, or basal ganglia, or a combination of cortical and subcortical. Assessment for upper limb motor impairment The Fugl-Meyer assessment for the upper extremity (FMA-UE) was used to assess upper limb motor function [27]. It is an objective measure used to assess motor recovery and outcomes, with scores ranging from 0 to 66, with higher scores indicating better motor function and recovery. In the study by Coscia et al. [28], ULI severity was classified based on FMA-UE scores (> 45 for mild, 30–45 for moderate, and < 30 for severe), aiming for applicability in neurotechnology trials. In our study, we adopted the threshold of < 30 to classify participants with severe ULI, considering its clinical relevance and practical applicability. Baseline group differences are presented in Table 1. Age and FMA-UE were the only variables that reached statistical significance (p < 0.05). Table 1Demographic characteristics of participantsMild/moderate ULI(n = 19)Severe ULI(n = 15) p Age (years)57.58 ± 13.7766.93 ± 8.510.03FMA-UE47 (14.50)9 (12)< 0.001Time since stroke onset (days)315 (185.50)321 (159.50)0.56SexMale10 (52.60%)5 (33.30%)0.44Female9 (47.40%)10 (66.70%)Hemiplegia sideLeft8 (42.10%)11 (73.30%)0.14Right11 (57.90%)4 (26.70%)Lesion locationCortical6 (31.60%)2 (13.30%)0.46Subcortical11 (57.90%)11 (73.30%)Cortical/subcortical2 (10.50%)2 (13.30%)Continuous variables were analyzed using the Mann-Whitney U test or Student’s t-test based on the Shapiro-Wilk normality test. Categorical variables were assessed using the chi-squared test. Continuous data are presented as mean ± standard deviation or median (interquartile range), whereas categorical data are presented as numbers (percentages). Statistical significance was set at p < 0.05. significant ULI upper limb impairment, FMA-UE the Fugl-Meyer assessment for upper extremity Table 1Demographic characteristics of participantsMild/moderate ULI(n = 19)Severe ULI(n = 15) p Age (years)57.58 ± 13.7766.93 ± 8.510.03FMA-UE47 (14.50)9 (12)< 0.001Time since stroke onset (days)315 (185.50)321 (159.50)0.56SexMale10 (52.60%)5 (33.30%)0.44Female9 (47.40%)10 (66.70%)Hemiplegia sideLeft8 (42.10%)11 (73.30%)0.14Right11 (57.90%)4 (26.70%)Lesion locationCortical6 (31.60%)2 (13.30%)0.46Subcortical11 (57.90%)11 (73.30%)Cortical/subcortical2 (10.50%)2 (13.30%)Continuous variables were analyzed using the Mann-Whitney U test or Student’s t-test based on the Shapiro-Wilk normality test. Categorical variables were assessed using the chi-squared test. Continuous data are presented as mean ± standard deviation or median (interquartile range), whereas categorical data are presented as numbers (percentages). Statistical significance was set at p < 0.05. significant ULI upper limb impairment, FMA-UE the Fugl-Meyer assessment for upper extremity Demographic characteristics of participants Continuous variables were analyzed using the Mann-Whitney U test or Student’s t-test based on the Shapiro-Wilk normality test. Categorical variables were assessed using the chi-squared test. Continuous data are presented as mean ± standard deviation or median (interquartile range), whereas categorical data are presented as numbers (percentages). Statistical significance was set at p < 0.05. significant ULI upper limb impairment, FMA-UE the Fugl-Meyer assessment for upper extremity EEG recording and preprocessing Participants were seated comfortably in an armchair in front of a monitor to facilitate EEG recordings. To minimize eye movement artifacts and ensure the quality of the EEG recordings, the participants were instructed to focus their gaze on a fixation cross displayed on the monitor. EEG data were collected during two resting-state sessions for a total of 4 min (1 min with eyes closed, followed by 1 min with eyes open, and repeated twice). A movement-related EEG session was conducted between resting-state sessions. EEG data were recorded at a sampling rate of 1000 Hz using 32 Ag/AgCl scalp electrodes, following the extended international 10–10 system (Brain Products LiveAmp, GmbH, Gilching, Germany) (Fig. 1A). The ground electrode was placed between Fp1 and Fp2, and the reference electrode was placed at Cz. All electrode impedances were maintained below 20 kΩ during EEG recording. The EEG data were band-pass filtered from 0.1 to 55 Hz, and eye-related artifacts were removed using mathematical procedures based on the first principal component analysis via Curry 7 (Compumedics, USA) software. All EEG electrodes were used without interpolation after artifact reduction. We analyzed only eyes-closed EEG data, as EEG signals recorded during eyes-closed conditions are inherently less susceptible to noise from eye-related artifacts. The data were down-sampled to 500 Hz and re-referenced using a common average reference. Two sets of artifact-free eyes-closed EEG data were divided into 4-s epochs. Epochs were rejected if they contained significant physiological artifacts (± 100 µV) at any electrode. The first 56 s were used, excluding the last 4 s due to data quality issues, resulting in 28 epochs for analysis. The EEG channels were reorganized so that coherence analyses consistently reflected the ipsilesional hemisphere (lesioned side) and the contralesional hemisphere (intact side), regardless of the side of hemiplegia. Fig. 1Configuration of electrode positions and flowchart of the proposed machine-learning approach, including feature selection. A Configuration of the electrodes used for coherence analysis in this study. B Flowchart illustrating the evaluation of the models using various algorithms and feature selection methods with leave-one-out cross-validation Configuration of electrode positions and flowchart of the proposed machine-learning approach, including feature selection. A Configuration of the electrodes used for coherence analysis in this study. B Flowchart illustrating the evaluation of the models using various algorithms and feature selection methods with leave-one-out cross-validation Feature extraction For feature extraction, we used EEG electrodes from the left and right hemispheres, excluding the midline electrodes (Fz, Cz, Pz, and Oz), as our focus was on investigating inter- and intrahemispheric brain connectivity. To calculate functional connectivity, the imaginary part of coherency (iCOH) was computed between the 28 electrodes for each EEG epoch and then averaged across all epochs. iCOH is widely used to avoid volume conduction effects caused by the differing conductivities of brain layers (e.g., cerebrospinal fluid, skull, and scalp), thereby preserving the original signals from the cortical surface in the functional connectivity pattern [29]. The iCOH features were computed in six frequency bands: delta (1–-4 Hz), theta (4–8 Hz), alpha (8–12 Hz), low-beta (12–20 Hz), high-beta (20–30 Hz), and gamma (30–55 Hz). Finally, a total of 2,268 iCOH values (28C2 × 6 frequency bands) were used for further analyses. To clearly reflect the effects associated with stroke lesion locations in each hemisphere, the EEG channels used for coherence analysis were reorganized by consistently defining the ipsilesional hemisphere as the lesioned side and the contralesional hemisphere as the intact side, irrespective of hemiplegic laterality. For patients with right-hemisphere lesions, EEG channel positions were inverted to consistently align the ipsilesional hemisphere to the left side [21]. The Matlab-based Fieldtrip toolbox was used to calculate the iCOH values [30]. Feature selection and ML-based classification A total of 2,268 iCOH values were used to classify groups with different levels of ULI severity; however, a high number of features relative to the number of samples (participants) could lead to overfitting. To mitigate this, we adopted feature-selection methods to optimize the features for classifying the two groups. Three feature-selection methods widely used in EEG studies―analysis of variance (ANOVA), SVM-recursive feature elimination (RFE), and ReliefF―were used [31, 32]. ANOVA employs F-values to select features by considering the relationship between the label and each feature [33]. SVM-RFE handles many features without requiring statistical corrections and considers individual differences by eliminating irrelevant or extreme data points [34, 35]. ReliefF, a category of relief-based feature selection methods, assesses the importance of features by their ability to distinguish between neighboring data points [36, 37]. All features were normalized prior to feature selection. Feature selection was performed in each iteration of the cross-validation process to prevent overfitting and reduce potential selection bias [38]. Preliminary analyses with feature subsets ranging from 1 to 250 revealed that classification performance initially increased with the number of features, peaked at approximately 50 features, and then slightly declined. Consequently, we focused on a feature subset ranging between 1 and 100 for further analysis. The classification performance was evaluated using four ML algorithms: logistic regression with L1, elastic net regularization, stochastic gradient descent (SGD), and SVM. Hyperparameters for each algorithm are listed in Supplementary Table S1. Leave-one-out cross-validation (LOOCV) was used to mitigate overfitting and reduce biases due to the small sample size (n = 34). Important features were identified based on their occurrence rate, calculated as the ratio of their appearances to the total number of LOOCV iterations. The performance of the ML approach was evaluated using balanced accuracy, sensitivity, and specificity, each reported with 95% confidence intervals (CIs) calculated using the Wilson score interval. The ML analysis was conducted using Python 3.11 (https://www.python.org/) and its compatible open-source libraries. A flowchart of the proposed ML approach is shown in Fig. 1B. Results Model performance The performances of all possible combinations of the three-feature selection and classification methods are detailed in Table 2; Fig. 2. Logistic regression with the L1 model, combined with the ReliefF feature selection method, showed superior performance in classifying ULI severity in chronic stroke survivors (accuracy = 0.91, sensitivity = 0.93, and specificity = 0.90 when using 14 features). Table 2Model performance and the number of selected features for each algorithm and feature selection methodANOVASVM-RFEReliefFN.Bal. Accu.Sens.Spec.N.Bal. Accu.Sens.Spec.N.Bal. Accu.Sens.Spec.Logistic regression with L120.82(0.66–0.92)0.80(0.55–0.93)0.84(0.62–0.94)530.71(0.48–0.79)0.80(0.42–0.85)0.63(0.41–0.81) 14 0.91 (0.77–0.97) 0.93 (0.70–0.99) 0.90 (0.69–0.97) Logistic regression with elastic net20.79(0.63–0.90)0.73(0.48–0.89)0.84(0.62–0.94)60.68(0.51–0.81)0.60(0.36–0.80)0.74(0.51–0.88)20.85(0.70–0.94)0.80(0.55–0.93)0.90(0.69–0.97)Logistic regression with SGD20.74(0.57–0.85)0.67(0.42–0.85)0.79(0.57–0.91)90.65(0.48–0.79)0.53(0.30–0.75)0.74(0.51–0.88)20.79(0.63–0.90)0.73(0.48–0.89)0.84(0.62–0.94)SVM20.82(0.66–0.92)0.80(0.55–0.93)0.84(0.62–0.94)60.65(0.48–0.79)0.53(0.30–0.75)0.74(0.51–0.88)110.88(0.73–0.95)0.80(0.55–0.93)0.95(0.75–0.99)Bold font indicates the highest model performance across all algorithms and feature selection methods. Values in parentheses indicate the 95% confidence intervals of model performance assessments ANOVA analysis of variance, SVM support vector machine, RFE recursive feature elimination, N. the number of features, Bal. Accu. balanced accuracy, Sens sensitivity, Spec specificity, SGD stochastic gradient descent Table 2Model performance and the number of selected features for each algorithm and feature selection methodANOVASVM-RFEReliefFN.Bal. Accu.Sens.Spec.N.Bal. Accu.Sens.Spec.N.Bal. Accu.Sens.Spec.Logistic regression with L120.82(0.66–0.92)0.80(0.55–0.93)0.84(0.62–0.94)530.71(0.48–0.79)0.80(0.42–0.85)0.63(0.41–0.81) 14 0.91 (0.77–0.97) 0.93 (0.70–0.99) 0.90 (0.69–0.97) Logistic regression with elastic net20.79(0.63–0.90)0.73(0.48–0.89)0.84(0.62–0.94)60.68(0.51–0.81)0.60(0.36–0.80)0.74(0.51–0.88)20.85(0.70–0.94)0.80(0.55–0.93)0.90(0.69–0.97)Logistic regression with SGD20.74(0.57–0.85)0.67(0.42–0.85)0.79(0.57–0.91)90.65(0.48–0.79)0.53(0.30–0.75)0.74(0.51–0.88)20.79(0.63–0.90)0.73(0.48–0.89)0.84(0.62–0.94)SVM20.82(0.66–0.92)0.80(0.55–0.93)0.84(0.62–0.94)60.65(0.48–0.79)0.53(0.30–0.75)0.74(0.51–0.88)110.88(0.73–0.95)0.80(0.55–0.93)0.95(0.75–0.99)Bold font indicates the highest model performance across all algorithms and feature selection methods. Values in parentheses indicate the 95% confidence intervals of model performance assessments ANOVA analysis of variance, SVM support vector machine, RFE recursive feature elimination, N. the number of features, Bal. Accu. balanced accuracy, Sens sensitivity, Spec specificity, SGD stochastic gradient descent Model performance and the number of selected features for each algorithm and feature selection method 0.82 (0.66–0.92) 0.80 (0.55–0.93) 0.84 (0.62–0.94) 0.71 (0.48–0.79) 0.80 (0.42–0.85) 0.63 (0.41–0.81) 0.91 (0.77–0.97) 0.93 (0.70–0.99) 0.90 (0.69–0.97) 0.79 (0.63–0.90) 0.73 (0.48–0.89) 0.84 (0.62–0.94) 0.68 (0.51–0.81) 0.60 (0.36–0.80) 0.74 (0.51–0.88) 0.85 (0.70–0.94) 0.80 (0.55–0.93) 0.90 (0.69–0.97) 0.74 (0.57–0.85) 0.67 (0.42–0.85) 0.79 (0.57–0.91) 0.65 (0.48–0.79) 0.53 (0.30–0.75) 0.74 (0.51–0.88) 0.79 (0.63–0.90) 0.73 (0.48–0.89) 0.84 (0.62–0.94) 0.82 (0.66–0.92) 0.80 (0.55–0.93) 0.84 (0.62–0.94) 0.65 (0.48–0.79) 0.53 (0.30–0.75) 0.74 (0.51–0.88) 0.88 (0.73–0.95) 0.80 (0.55–0.93) 0.95 (0.75–0.99) Bold font indicates the highest model performance across all algorithms and feature selection methods. Values in parentheses indicate the 95% confidence intervals of model performance assessments ANOVA analysis of variance, SVM support vector machine, RFE recursive feature elimination, N. the number of features, Bal. Accu. balanced accuracy, Sens sensitivity, Spec specificity, SGD stochastic gradient descent Fig. 2Balanced classification accuracies of the four classification methods. Logistic regression with A L1, B elastic net, C SGD, and D SVM for the three feature selection methods (ANOVA, SVM-RFE, and ReliefF) Balanced classification accuracies of the four classification methods. Logistic regression with A L1, B elastic net, C SGD, and D SVM for the three feature selection methods (ANOVA, SVM-RFE, and ReliefF) Important EEG features Figure S1 illustrates the group-level iCOH patterns between the mild/moderate and severe ULI groups across six frequency bands. Logistic regression with L1 and the ReliefF method identified 30 features from 14 input variables within LOOCV, as shown in Supplementary Table S2. Among these, we focused on important features with empirically determined occurrence rates exceeding 0.5. Important features were identified, including three delta, ten theta, and one low-beta coherences, whereas alpha and high-beta coherences were not selected (Fig. 3). Table 3 presents 14 important features with statistical differences (p-values) between the two groups. Fig. 3Topographic maps of the top fourteen functional connectivity between the mild/moderate ULI and severe ULI groups. A Delta band, B theta band, and C low beta band. The intensity of the red line indicates the occurrence rate, which represents the number of features appearing in a cross-validation iteration Topographic maps of the top fourteen functional connectivity between the mild/moderate ULI and severe ULI groups. A Delta band, B theta band, and C low beta band. The intensity of the red line indicates the occurrence rate, which represents the number of features appearing in a cross-validation iteration Table 3Occurrence rates of the top fourteen features and their corresponding coherence values in the optimal model distinguishing severe from mild/moderate ULISelected featuresOccurrence rateMild/moderate ULI(n = 19)Severe ULI(n = 15) p DeltaTP.ipsi - F.con1.000.070 ± 0.0270.029 ± 0.022< 0.001DeltaCP.ipsi-P.ipsi1.000.047 (0.027)0.017 (0.022)< 0.01Low-betaP.con-TP.con1.000.068 (0.016)0.035 (0.025)< 0.01ThetaTP.con-FP.con1.000.078 ± 0.0450.035 ± 0.017< 0.01ThetaF.con-FP.con1.000.063 (0.074)0.033 (0.033)< 0.01ThetaFC.ipsi-CP.ipsi1.000.085 ± 0.0490.038 ± 0.023< 0.01ThetaCP.ipsi-FP.con1.000.080 ± 0.0390.037 ± 0.020< 0.001ThetaF.ipsi-P.ipsi1.000.091 (0.097)0.042 (0.031)< 0.001ThetaFC.ipsi-P.ipsi1.000.095 (0.073)0.035 (0.026)< 0.01ThetaFC.ipsi-CP.ipsi1.000.071 (0.075)0.034 (0.023)< 0.01ThetaFC.ipsi-P.ipsi0.910.087 (0.076)0.046 (0.031)0.04DeltaT.ipsi-CP.ipsi0.790.078 ± 0.0330.044 ± 0.026< 0.01ThetaFC.ipsi-P.con0.790.060 (0.037)0.023 (0.026)< 0.01ThetaP.ipsi-FP.con0.620.081 (0.052)0.034 (0.022)< 0.01The results were obtained using the best-performing ML model (Logistic regression with L1 regularization and the relieff feature selection method)The occurrence rate is computed as the number of appearances divided by the total number of cross-validation iterations. Coherence values were analyzed using the Mann–Whitney U test or Student’s t-test based on the Shapiro–Wilk normality test and were presented as mean ± standard deviation or median (interquartile range)ULI upper limb impairment. FMA-UE the Fugl-Meyer assessment for upper extremity; channel number; ipsi, ipsilesional, con contralesional, FP prefrontal, F frontal, FC frontocentral, FT frontotemporal, T temporal, TP temporoparietal, C central, CP centroparietal, P parietal, O occipitalipsi ipsilesional, con contralesional, EEG electroencephalography, num the number of features, n the number of participants, ANOVA analysis of variance; SVM, support vector machine; RFE recursive feature elimination, SGD stochastic gradient descentANOVA analysis of variance, SGD stochastic gradient descent, SVM support vector machineipsi ipsilesional, con contralesional. Table 3Occurrence rates of the top fourteen features and their corresponding coherence values in the optimal model distinguishing severe from mild/moderate ULISelected featuresOccurrence rateMild/moderate ULI(n = 19)Severe ULI(n = 15) p DeltaTP.ipsi - F.con1.000.070 ± 0.0270.029 ± 0.022< 0.001DeltaCP.ipsi-P.ipsi1.000.047 (0.027)0.017 (0.022)< 0.01Low-betaP.con-TP.con1.000.068 (0.016)0.035 (0.025)< 0.01ThetaTP.con-FP.con1.000.078 ± 0.0450.035 ± 0.017< 0.01ThetaF.con-FP.con1.000.063 (0.074)0.033 (0.033)< 0.01ThetaFC.ipsi-CP.ipsi1.000.085 ± 0.0490.038 ± 0.023< 0.01ThetaCP.ipsi-FP.con1.000.080 ± 0.0390.037 ± 0.020< 0.001ThetaF.ipsi-P.ipsi1.000.091 (0.097)0.042 (0.031)< 0.001ThetaFC.ipsi-P.ipsi1.000.095 (0.073)0.035 (0.026)< 0.01ThetaFC.ipsi-CP.ipsi1.000.071 (0.075)0.034 (0.023)< 0.01ThetaFC.ipsi-P.ipsi0.910.087 (0.076)0.046 (0.031)0.04DeltaT.ipsi-CP.ipsi0.790.078 ± 0.0330.044 ± 0.026< 0.01ThetaFC.ipsi-P.con0.790.060 (0.037)0.023 (0.026)< 0.01ThetaP.ipsi-FP.con0.620.081 (0.052)0.034 (0.022)< 0.01The results were obtained using the best-performing ML model (Logistic regression with L1 regularization and the relieff feature selection method)The occurrence rate is computed as the number of appearances divided by the total number of cross-validation iterations. Coherence values were analyzed using the Mann–Whitney U test or Student’s t-test based on the Shapiro–Wilk normality test and were presented as mean ± standard deviation or median (interquartile range)ULI upper limb impairment. FMA-UE the Fugl-Meyer assessment for upper extremity; channel number; ipsi, ipsilesional, con contralesional, FP prefrontal, F frontal, FC frontocentral, FT frontotemporal, T temporal, TP temporoparietal, C central, CP centroparietal, P parietal, O occipitalipsi ipsilesional, con contralesional, EEG electroencephalography, num the number of features, n the number of participants, ANOVA analysis of variance; SVM, support vector machine; RFE recursive feature elimination, SGD stochastic gradient descentANOVA analysis of variance, SGD stochastic gradient descent, SVM support vector machineipsi ipsilesional, con contralesional. Occurrence rates of the top fourteen features and their corresponding coherence values in the optimal model distinguishing severe from mild/moderate ULI The results were obtained using the best-performing ML model (Logistic regression with L1 regularization and the relieff feature selection method) The occurrence rate is computed as the number of appearances divided by the total number of cross-validation iterations. Coherence values were analyzed using the Mann–Whitney U test or Student’s t-test based on the Shapiro–Wilk normality test and were presented as mean ± standard deviation or median (interquartile range) ULI upper limb impairment. FMA-UE the Fugl-Meyer assessment for upper extremity; channel number; ipsi, ipsilesional, con contralesional, FP prefrontal, F frontal, FC frontocentral, FT frontotemporal, T temporal, TP temporoparietal, C central, CP centroparietal, P parietal, O occipital ipsi ipsilesional, con contralesional, EEG electroencephalography, num the number of features, n the number of participants, ANOVA analysis of variance; SVM, support vector machine; RFE recursive feature elimination, SGD stochastic gradient descent ANOVA analysis of variance, SGD stochastic gradient descent, SVM support vector machine ipsi ipsilesional, con contralesional. We identified significant features, including the delta interhemispheric connectivity between the ipsilesional temporoparietal and contralesional frontal regions. Additionally, delta connectivity was observed with intrahemispheric connectivity, specifically between the ipsilesional temporal and centroparietal regions as well as the ipsilesional centroparietal and parietal areas. Furthermore, theta interhemispheric connectivity was found between the ipsilesional centroparietal and contralesional prefrontal regions, the ipsilesional frontocentral and contralesional parietal regions, and the ipsilesional parietal and contralesional prefrontal regions. Regarding intrahemispheric connectivity, theta coherence was noted between the ipsilesional frontocentral and centroparietal regions, ipsilesional frontal and parietal regions, ipsilesional frontocentral and parietal regions, and ipsilesional frontal and centroparietal regions. In the contralesional hemisphere, theta coherence was observed between the temporoparietal and prefrontal regions, as well as between the frontal and prefrontal regions. Intrahemispheric coherence in the low beta band was also found between the contralesional parietal and temporoparietal regions. Discussion Achieving a balanced accuracy of approximately 91%, the ML approach in this study achieved a higher model performance in classifying severe ULI in chronic stroke survivors, notably by utilizing whole-brain resting-state EEG coherence features from all channels. Fourteen significant coherence features were identified across the delta, theta, and low-beta frequency bands, most frequently selected in LOOCV. Most resting-state EEG coherence features that substantially contribute to distinguishing severe ULI are distributed in low-frequency bands, specifically the delta and theta bands. Logistic regression with L1 significantly selected delta interhemispheric connectivity between the ipsilesional temporoparietal and contralesional frontal regions as well as intrahemispheric connectivity within the ipsilesional temporal, centroparietal, and parietal areas. Significant theta interhemispheric and intrahemispheric connectivity was observed in the frontal and parietal regions. These findings align with previous studies reporting that poorer outcomes after stroke are associated with low-frequency oscillations [39, 40]. Low-frequency oscillations are associated with cognitive control, attentional processing, movement, and skilled motor control [41–44]. Brain damage following stroke, brain tumors, or traumatic brain injury exhibits altered activity in the delta and theta frequency bands [40, 45, 46], with more pronounced slowing observed in cases of more severe brain damage [47]. Alterations in low-frequency bands are also associated with general stroke-related outcomes, such as the modified Rankin Scale (mRS) and the National Institutes of Health Stroke Scale (NIHSS). Higher delta power in the ipsilateral hemisphere [48], a higher delta power ratio [49], and increased delta power during the acute stroke phase [50] were associated with higher NIHSS scores, indicating poorer functional outcomes. Furthermore, a higher delta-theta/alpha-beta ratio was linked to poorer functional outcomes, as reflected by higher mRS scores [51, 52]. Therefore, changes in delta- or theta-band frequencies on EEG are recognized as consistent and important biomarkers of stroke severity and overall functional outcomes. Recent studies investigating EEG parameters related to motor impairment have consistently shown that low-frequency bands can serve as significant correlates or predictors of motor deficits, especially in the context of ULI following stroke, as measured by functional assessments such as FMA-UE and Actional Research Arm Test (ARAT) [16, 17, 53–56]. Saes et al. [17, 53] demonstrated that the brain symmetry index in the delta and theta frequency bands is predominantly associated with FMA-UE in chronic stroke and serves as a prognostic biomarker for FMA-UE 6 months post-stroke. In stroke survivors with poor motor function (ARAT < 10 and no motor-evoked potential in the paretic hand), delta and theta band power in the ipsilesional hemisphere, as well as theta band power in the contralesional hemisphere, were found to be increased [56]. Low node strength of the delta band in the ipsilesional hemisphere is correlated with poor motor function (low FMA-UE score) [54]. These findings align with those of our study, which further support the role of low-frequency band activity as a significant biomarker of motor impairment after stroke. Altered EEG power and coherence in low-frequency bands may serve as correlates of compensatory mechanisms following stroke arising from anatomical or functional disruptions. In addition, our data-driven approach underscores the importance of low-frequency bands across various regions―not only in motor areas but also in non-motor regions and across different hemispheres (ipsilesional, contralesional, and interhemispheric)―as biomarkers reflecting motor impairments after stroke. This finding may reflect various aspects of the injury- or recovery-related process [16] and aligns with previous studies reporting contributions from widespread network modulation in non-motor regions [57–59]. The present study found a significant association between low-beta intrahemispheric connectivity in the contralesional parietal and temporoparietal regions and severe ULI, although only one of the 14 significant features was identified. Alterations in the low-beta frequency band during motor tasks following a stroke are well known. Low-beta event-related desynchronization (ERD) observed when moving the hemiparetic hand showed stronger ERD in the contralesional hemisphere of the stroke survivors compared to that of healthy controls [60]. In addition, more efficient connectivity parameters in the low-beta band during the task in the ipsilesional hemisphere correlated with better upper limb function in chronic stroke [21]. However, few studies have investigated the association between low beta band activity and motor recovery after stroke, and results regarding beta activity in motor recovery have been inconsistent [16]. Therefore, further research is required to reveal low-beta connectivity and motor recovery. Coherences in the contralesional hemisphere between temporoparietal and frontoparietal regions were also identified as important features associated with motor impairment (Table 3). Although the role of the contralesional hemisphere in motor recovery after stroke remains controversial, several studies suggest that altered connectivity may reflect maladaptive reorganization rather than beneficial compensation, potentially leading to poorer functional outcomes [61–63]. Since our analysis focused solely on data from chronic stroke surviors, we could not determine whether the observed alterations in connectivity within the contralesional hemisphere are de novo phenomena emerging during the recovery process or represent maladaptive changes that should be addressed during rehabilitation. Therefore, longitudinal follow-up studies are necessary to clarify these associations. Additionally, the spatial distribution of selected features spanning motor and non-motor regions across the hemispheres suggests that severe impairment is accompanied by systemic disruption of large-scale brain networks rather than being confined to primary motor areas [64]. This aligns with growing evidence that stroke recovery engages not only motor execution systems but also broader networks responsible for attention, sensorimotor integration, and cognitive control, with the involvement of non-motor regions potentially reflecting compensatory recruitment of associative systems in individuals with stroke [18, 59, 65–67]. The performance of ML-based classification may vary depending on the choice of functional connectivity metrics, such as phase-locking value (PLV) [68], partial directed coherence (PDC) [69], and iCOH. Therefore, selecting an appropriate functional connectivity metric according to EEG signal characteristics and analytical objectives is crucial. For example, PLV is suited for task-induced signals, such as event-related potentials, where neural responses are time-locked to specific events [70–72]. However, PLV has increasingly been avoided due to its susceptibility to spurious connections caused by volume conduction effects. Similarly, directed metrics like PDC are preferable for investigating causal relationships based on explicit hypotheses [73, 74] rather than for exploratory, data-driven analyses of resting-state EEG. In contrast, iCOH exclusively captures phase-lagged interactions, making it robust against volume conduction artifacts and especially suitable for unbiased exploration of resting-state brain connectivity. Consequently, careful selection of functional connectivity metrics aligned with research objectives is essential for achieving optimal classification outcomes. In the present study, among various functional connectivity metrics, iCOH was adopted to capture resting-state EEG characteristics while minimizing volume conduction effects, thereby leading to reliable results. There are several limitations in the present study, and we propose future directions to address them. First, the relatively small sample size and age imbalance might limit the generalizability of the findings to the broader population of individuals with chronic stroke. To investigate potential confounding effects due to age imbalances, we conducted additional analyses using age-matched subgroups generated via a nearest-neighbor matching approach. The classification performance obtained from these age-matched subgroups showed no substantial differences compared to the original groups (Tables S3 and S4), suggesting that age is unlikely to be a major confounding factor in our findings. Nevertheless, further studies with larger and more balanced samples are required. Second, source-level analysis was not conducted due to the limited number of EEG channels (fewer than 62). Therefore, region-specific cortical activity related to motor impairments could not be directly examined in this study. To overcome this limitation, future studies could incorporate cortical source-level analysis, providing more precise, region-specific neural information and enabling the development of more targeted and personalized predictive models. Third, the absence of raw subscale scores from the FMA limited our ability to explore more detailed distinctions in motor impairments, such as differentiating between distal and proximal functions. Recognizing the clinical significance of these finer-grained distinctions, developing a prediction model based on detailed FMA subscale scores would provide a valuable direction for future research. Finally, we cannot guarantee that the selected EEG coherence features will serve as reliable biomarkers for motor recovery and responses to rehabilitative interventions, as the analysis was based solely on data from chronic stroke survivors [13]. Therefore, future research using larger longitudinal datasets that include repeated EEG and clinical measurements following early stroke is necessary [75]. With these larger longitudinal datasets, applying data-driven approaches using ML methods can reveal more comprehensive EEG features related to motor recovery after stroke that hypothesis-driven studies may overlook [16]. This approach also has the potential to develop predictive models for motor recovery or impairment that can be applied to new subjects outside the initial dataset, thereby enhancing the accuracy of predictions in clinical settings [76]. Conclusions The present data-driven analysis, employing ML methods, utilized whole-brain resting-state EEG connectivity features to classify severe ULI from mild-to-moderate ULI in chronic stroke survivors. Our research highlights the association between alterations in connectivity within low-frequency bands (delta and theta) and severe ULI across widespread brain regions, including areas outside the sensorimotor cortex, as well as bilateral intrahemispheric and interhemispheric regions. Further research utilizing larger longitudinal datasets from patients with early strokes and applying data-driven approaches with ML methods, including both classification and regression techniques, could help identify more comprehensive EEG features related to motor recovery and contribute to the development of more accurate predictive models for motor recovery and rehabilitation responses. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1.