Imaging-Enhanced Prediction of Upper Limb Motor Function After Stroke Using Automated Segmentation and Cluster Analysis Authors: Seungwoo Cha, Wi-Sun Ryu, Bokgi Hwang, Dae Yul Kim, Seung Hak Lee Published: 2026 ABSTRACT AbstractBackgroundStroke is a leading cause of long-term disability, and early prediction of motor outcomes is critical for guiding individualized rehabilitation strategies. Although neuroimaging offers prognostic insights, its clinical integration remains limited. This study aimed to evaluate whether explainable imaging-derived features can improve the prediction of upper limb motor impairment and to identify distinct recovery profiles through cluster analysis.MethodsWe retrospectively analyzed 242 patients with acute ischemic stroke who underwent brain magnetic resonance imaging and motor function assessment. Clinical and imaging variables—including infarct volume, motor tract involvement, and subtype classification—were extracted using automated software. Predictive models for Fugl-Meyer Assessment upper extremity (FMA-UE) scores were developed using hierarchical regression, and K-means clustering was applied to standardized baseline data. Cluster characteristics were compared, and recovery profiles were visualized.ResultsIncorporating imaging variables improved prediction of FMA-UE scores (adjusted R2: 0.595), with motor tract lesion volume emerging as a key predictor. Cluster analysis revealed three distinct patient subgroups: Cluster 1 (mild impairments, younger age), Cluster 2 (older patients with multiple comorbidities), and Cluster 3 (severe neurological deficits, large infarcts). Functional outcomes differed significantly across clusters, with Cluster 3 showing the greatest disability (76% with follow-up modified Rankin Scale 4–6, P < 0.001).ConclusionImaging-derived features from early magnetic resonance imaging improve the prediction of upper limb motor outcomes after stroke. Cluster-based stratification identifies clinically meaningful subgroups, supporting precision rehabilitation through personalized care planning and targeted secondary prevention.Keywords: Stroke, Neuroimaging, Cluster Analysis, Hemiplegia, Rehabilitation INTRODUCTION Stroke is a leading cause of long-term disability, with most survivors experiencing some degree of functional decline.1 Accurately predicting motor deficits in the early stages after stroke is essential for establishing realistic rehabilitation goals and developing individualized treatment strategies. However, outcome prediction remains challenging because of the heterogeneous nature of stroke pathology and patient recovery patterns.2 Neuroimaging provides valuable prognostic insights by delineating the location and extent of brain lesions, which are key determinants of motor outcomes.3 In recent years, automated analysis techniques have emerged as promising tools for integrating complex clinical and imaging data to forecast functional recovery. These methods can handle large datasets and model nonlinear relationships.4 Several studies have demonstrated the utility of imaging-based biomarkers in predicting upper limb recovery. For instance, a multicenter study in patients with chronic stroke employed machine learning models on multimodal data—including demographic, clinical, neurophysiological, and magnetic resonance imaging (MRI) variables—to predict post-therapy Fugl-Meyer Assessment scores for upper extremity function (FMA-UE).5 Similarly, in the acute phase, Kim et al.6 developed a convolutional neural network using early brain MRI and clinical data to predict 6-month motor outcomes in patients with corona radiata infarcts. These findings suggest that incorporating neuroimaging features—such as lesion size, anatomical location, and tract involvement—into predictive models enhances the accuracy of motor outcome predictions compared with clinical measures alone. Nevertheless, despite growing evidence, imaging-based approaches remain insufficiently developed for routine clinical use, with persistent challenges in interpretability, ease of implementation, and practical utility for predicting and guiding upper limb functional recovery after ischemic stroke. The present study aimed to evaluate whether explainable imaging-derived features from automated brain MRI analysis, such as infarct volume or corticospinal tract involvement, can improve the prediction of upper limb motor impairment, as measured by the FMA-UE. Additionally, cluster analysis was conducted to identify distinct recovery profiles among stroke patients. These findings are intended to support the development of more personalized and effective rehabilitation strategies. METHODS Study population This retrospective cohort study included patients admitted to Asan Medical Center for acute ischemic stroke between July 1, 2022, and July 31, 2023, who underwent brain MRI and were subsequently referred to the Department of Rehabilitation Medicine for functional assessment. Patients with incomplete imaging or clinical data were excluded (Fig. 1). Study population. MRI = magnetic resonance imaging. Clinical variables Clinical data were collected from a total of 242 patients with ischemic stroke and included demographics, comorbidities, history of acute interventions, National Institutes of Health Stroke Scale (NIHSS) scores at admission, FMA-UE scores, and premorbid, initial, and follow-up modified Rankin Scale (mRS) scores. Acute interventions included neurovascular reperfusion procedures such as intravenous thrombolysis and endovascular thrombectomy. The NIHSS is a standardized measure of stroke severity in the acute phase. The Fugl-Meyer Assessment provides a validated, quantitative assessment of motor impairment, and its upper extremity subscale (FMA-UE) specifically measures arm and hand motor function.7 Patients were referred to the Department of Rehabilitation Medicine within 1 day of admission, and the FMA-UE was administered after referral, within approximately 3 days of hospitalization. The mRS is commonly used to evaluate overall functional disability or dependence in daily activities. Premorbid mRS and initial mRS were assessed at the time of admission, whereas follow-up mRS was measured at discharge. Imaging variables MRI scans were obtained within 24 hours of hospital admission using standardized acute stroke imaging protocols. All examinations were performed on 3.0-T clinical MRI systems (Philips Achieva; Philips Healthcare, Eindhoven, The Netherlands; and GE Signa Architect, GE Healthcare, Milwaukee, WI, USA), both equipped with 32-channel phased-array head coils. The imaging protocol consisted of axial diffusion-weighted imaging (DWI), susceptibility-weighted or gradient-echo sequences, fluid-attenuated inversion recovery (FLAIR), time-of-flight intracranial angiography, and contrast-enhanced magnetic resonance angiography of the cervical vessels. DWI was acquired with a repetition time (TR) of 3,000 ms, echo time (TE) of 77.6 ms, field of view (FOV) of 230 mm, flip angle of 90°, matrix size of 256 × 256, and slice thickness of 6 mm. FLAIR images were obtained with a TR of 9,000 ms, TE of 123.4 ms, FOV of 230 mm, flip angle of 160°, matrix size of 512 × 512, and slice thickness of 6 mm. Imaging-derived variables included total infarct volume, quantitative motor tract involvement, stroke subtype classification, and white matter hyperintensity (WMH) volume, extracted from DWI and FLAIR sequences. DWI lesions were semi-automatically segmented using validated commercial software (JLK DWI, JLK Inc., Seoul, Korea), following standardized protocols.8,9 To ensure accuracy, all resulting segmentation masks were visually inspected and manually corrected by a specialist in neuroanatomy (Cha S) to remove artifacts and verify lesion boundaries. Quantification of motor tract involvement was performed by registering a standardized template to each patient’s native DWI space. A standard T1-weighted template10 was aligned to the DWI image using a one-step affine transformation. Although nonlinear registration with cost-function masking is commonly used to correct anatomical variability, a linear approach was deliberately selected to minimize potential distortions related to cerebral edema and mass effect in acute infarction. Previous studies have demonstrated that nonlinear normalization may induce structural deformation or artificial midline shifts in the presence of extensive focal pathology, even when lesion masking is applied. By preserving native anatomical geometry through affine transformation, motor tract involvement measurements more accurately reflected true pathological displacement.11,12 Using the resulting transformation matrix, the template-based motor tract mask representing the corticospinal tract13 was mapped into each patient’s native coordinate system without applying elastic deformation. Finally, the motor tract lesion volume was calculated by computing the intersection (overlap volume) between the patient’s segmented infarct mask and the transformed motor tract mask (Fig. 2). Stroke subtypes were classified based on lesion characteristics and the presence of atrial fibrillation, resulting in categories of cardioembolism, large artery atherosclerosis, or other etiologies.14 WMHs were segmented from FLAIR images using JLK WMH (JLK Inc.), a validated automated software package for WMH quantification.15 The calculated WMH volume represented the total burden, inclusive of both periventricular and deep WMHs. Schematic workflow for quantifying motor tract lesion volume. (A) The acute infarct lesion (red) is segmented on the patient’s native DWI. (B) A standard template containing the motor tract mask (blue) is prepared. (C) The standard template is registered to the patient's native space using an affine transformation matrix. (D) The motor tract lesion volume is quantified as the overlap (purple) between the segmented infarct and the transformed motor tract mask. DWI = diffusion-weighted imaging. Statistical analysis Upper limb motor impairment severity was dichotomized using FMA-UE scores into two groups: severe (≤ 33) and mild-to-moderate (> 33). Clinical variables were compared between groups using independent t-tests or Wilcoxon rank-sum tests for continuous variables and χ2 or Fisher’s exact tests for categorical data, as appropriate. To predict continuous FMA-UE scores, hierarchical multiple regression models were constructed, incorporating clinical, functional, and imaging variables sequentially. Model 1 included demographic and vascular risk factors (age, sex, body mass index, hypertension, diabetes mellitus, atrial fibrillation, hyperlipidemia, coronary artery disease, peripheral artery disease, cancer, previous stroke, family history of stroke, and smoking). Model 2 additionally incorporated clinical severity and functional variables (acute intervention, NIHSS score, premorbid mRS, and initial mRS). Model 3 further included imaging-derived variables (motor tract lesion volume, WMH volume, and stroke subtype classification). Multicollinearity was assessed using variance inflation factors (VIFs), and variables with VIF values greater than 10 were excluded from the models. Due to high collinearity between total infarct volume and motor tract lesion volume, only motor tract lesion volume was retained in the final model. Model performance was assessed using adjusted R2 to determine the incremental benefit of adding imaging-derived predictors. Regression coefficients (β) and corresponding standard errors were estimated for each predictor. Model generalizability was evaluated using k-fold cross-validation with five folds, and cross-validated root mean square error (CV RMSE) was calculated for performance comparison across models. Additionally, cluster analysis was performed to identify subgroups with distinct recovery profiles. Clustering was performed using age, sex, body mass index, hypertension, diabetes mellitus, atrial fibrillation, hyperlipidemia, coronary artery disease, peripheral artery disease, cancer, previous stroke, family history of stroke, smoking, acute intervention, NIHSS score, premorbid and initial mRS scores, total lesion volume, motor tract lesion volume, WMH volume, and stroke subtype classification. The optimal number of clusters was determined to be 3 based on internal validation using the Silhouette index calculated for both K-means and agglomerative clustering across candidate solutions (K = 2–15). In both algorithms, Silhouette values showed an early plateau beginning at K = 3, which was selected as the stable solution (Supplementary Fig. 1). K-means clustering was then conducted, and clusters were compared using analysis of variance or Kruskal-Wallis tests for continuous variables and chi-square or Fisher’s exact tests for categorical variables. Cluster structure and separation were visualized in both two- and three-dimensional space using principal component analysis. All statistical analyses were conducted using R 4.5.0 (R Foundation for Statistical Computing, Vienna, Austria). Ethics statement The study was approved by the Institutional Review Board (IRB) of Asan Medical Center, University of Ulsan College of Medicine (IRB No. 2023-1269). Informed consent was waived because of the retrospective nature of the study. RESULTS Clinical characteristics by FMA-UE severity Among the 242 patients included in the analysis, 45 (18.6%) were categorized as having severe upper limb motor impairment (FMA-UE ≤ 33), while 197 (81.4%) exhibited mild-to-moderate impairment (FMA-UE > 33). Compared with the mild-to-moderate group, patients in the severe group had a significantly lower prevalence of hypertension (56% vs. 75%, P = 0.015) but a higher rate of acute intervention (24% vs. 8.6%, P = 0.006) (Table 1). Baseline stroke severity, as measured by NIHSS, was substantially greater in the severe group (10.6 ± 4.5 vs. 3.4 ± 2.8, P < 0.001). Both total lesion volume and motor tract lesion volume were significantly larger in the severe group (35.7 ± 59.9 mL and 1.6 ± 1.7 mL, respectively; both P < 0.001). Premorbid, initial, and follow-up mRS scores were also significantly worse in the severe group (P < 0.01 for all comparisons). There were no significant differences between groups in stroke subtype classification or WMH volume. Table 1Clinical characteristics by severity, defined by the Fugl-Meyer Assessment upper extremity scoreVariablesAll (N = 242)Severe (≤ 33) (n = 45)Mild-to-moderate (> 33) (n = 197)P valueAge69.5 ± 12.467.5 ± 13.470.0 ± 12.10.253Sex, male142 (59.0)31 (69.0)111 (56.0)0.169Body mass index24.1 ± 3.523.3 ± 3.624.3 ± 3.40.171ComorbiditiesHypertension173 (71.0)25 (56.0)148 (75.0)0.015Diabetes mellitus92 (38.0)19 (42.0)73 (37.0)0.636Atrial fibrillation87 (36.0)18 (40.0)69 (35.0)0.649Hyperlipidemia105 (43.0)20 (44.0)85 (43.0)1.000Coronary artery disease32 (13.0)8 (18.0)24 (12.0)0.450Peripheral artery disease3 (1.2)1 (2.2)2 (1.0)1.000Cancer18 (7.4)4 (8.9)14 (7.1)0.923Previous stroke53 (22.0)10 (22.0)43 (22.0)1.000Family history of stroke21 (8.7)3 (6.7)18 (9.1)0.812Smoking50 (21.0)10 (22.0)40 (20.0)0.934Acute intervention28 (12.0)11 (24.0)17 (8.6)0.006NIHSS4.7 ± 4.210.6 ± 4.53.4 ± 2.8< 0.001Premorbid mRS0.0090–1191 (80.0)32 (71.0)159 (82.0)2–333 (14.0)6 (13.0)27 (14.0)4–614 (6.0)7 (16.0)7 (4.0)Initial mRS< 0.0010–178 (32.0)2 (4.0)76 (39.0)2–387 (36.0)2 (4.0)85 (43.0)4–677 (32.0)41 (92.0)36 (18.0)Follow-up mRS< 0.0010–184 (35.0)2 (4.0)82 (42.0)2–387 (36.0)4 (9.0)83 (43.0)4–668 (28.0)39 (87.0)29 (15.0)Lesion volume12.7 ± 31.035.7 ± 59.97.4 ± 15.0< 0.001Motor tract lesion volume0.5 ± 1.01.6 ± 1.70.2 ± 0.3< 0.001White matter hyperintensity volume16.9 ± 16.717.5 ± 14.816.8 ± 17.10.632Stroke subtype0.199CE83 (34.0)21 (47.0)62 (31.0)LAA108 (45.0)17 (38.0)91 (46.0)Others51 (21.0)7 (15.0)44 (22.0)Values are presented as mean ± standard deviation or number (%).NIHSS = National Institutes of Health Stroke Scale, mRS = modified Rankin Scale, CE = cardioembolism, LAA = large artery atherosclerosis. Clinical characteristics by severity, defined by the Fugl-Meyer Assessment upper extremity score Values are presented as mean ± standard deviation or number (%). NIHSS = National Institutes of Health Stroke Scale, mRS = modified Rankin Scale, CE = cardioembolism, LAA = large artery atherosclerosis. Predictive modeling for FMA-UE scores To identify predictors of upper limb motor outcome, three hierarchical multiple regression models were constructed. Model 1 demonstrated limited predictive power (adjusted R2 = 0.010; CV RMSE = 21.41, Akaike information criterion [AIC] = 1,952.7). Model 2 showed significant improvement in predictive performance (adjusted R2 = 0.556, CV RMSE = 14.57, AIC = 1,781.7). Model 3 yielded the highest predictive performance (adjusted R2 = 0.595, CV RMSE = 14.49, AIC = 1,764.4), highlighting the additional value of imaging in predicting upper limb motor recovery (Table 2). In Model 3, higher NIHSS scores, larger motor tract lesion volumes, and higher initial mRS scores were significantly associated with lower FMA-UE scores, while increasing age was also associated with worse FMA-UE scores (Supplementary Table 1). Table 2Comparison of predictive models for Fugl-Meyer Assessment upper extremity scores using clinical, functional, and imaging variablesModelAdjusted R2CV RMSEAICModel 1: Clinical information only0.01021.411,952.7Model 2: Model 1 plus functional evaluation0.55614.571,781.7Model 3: Model 2 plus imaging variables0.59514.491,764.4Model 1: Age, sex, body mass index, hypertension, diabetes mellitus, atrial fibrillation, hyperlipidemia, coronary artery disease, peripheral artery disease, cancer, previous stroke, family history of stroke, and smoking; Model 2: Model 1 variables plus acute intervention, National Institutes of Health Stroke Scale, premorbid mRS, and initial mRS; Model 3: Model 2 variables plus motor tract lesion volume, white matter hyperintensity volume, and stroke subtype classification.CV RMSE = cross-validated root mean square error, AIC = Akaike information criterion, mRS = modified Rankin Scale. Comparison of predictive models for Fugl-Meyer Assessment upper extremity scores using clinical, functional, and imaging variables Model 1: Age, sex, body mass index, hypertension, diabetes mellitus, atrial fibrillation, hyperlipidemia, coronary artery disease, peripheral artery disease, cancer, previous stroke, family history of stroke, and smoking; Model 2: Model 1 variables plus acute intervention, National Institutes of Health Stroke Scale, premorbid mRS, and initial mRS; Model 3: Model 2 variables plus motor tract lesion volume, white matter hyperintensity volume, and stroke subtype classification. CV RMSE = cross-validated root mean square error, AIC = Akaike information criterion, mRS = modified Rankin Scale. To evaluate the robustness of our findings, we performed a sensitivity analysis excluding patients with previous stroke or a premorbid mRS score ≥ 2 (Supplementary Table 2). The results were consistent with the primary analysis, with Model 3 demonstrating the best predictive performance (adjusted R2 = 0.624, CV RMSE = 13.82, AIC = 1,157.0). Cluster-based stratification of patient profiles K-means clustering identified three distinct patient subgroups based on standardized demographic, clinical, and imaging data (Fig. 3). Cluster 1 (n = 96) was characterized by younger age, fewer comorbidities, and favorable clinical and functional outcomes; Cluster 2 (n = 88) comprised predominantly older patients with a high prevalence of atrial fibrillation (68%) and prior stroke history (51%); and Cluster 3 (n = 33) represented the most severely affected group, with the highest baseline NIHSS scores (11.7 ± 4.1), largest lesion volumes (55.9 ± 59.8 mL), and greatest motor tract involvement (2.0 ± 1.8 mL), despite a relatively lower burden of comorbidities (Table 3). The distribution of stroke subtypes also varied significantly across clusters, with cardioembolic strokes predominating in Clusters 2 and 3. Post-clustering comparisons of functional outcomes revealed marked differences across the three groups. Approximately 70% of patients in Cluster 3 had severe upper limb motor impairment (FMA-UE ≤ 33). Follow-up mRS scores were notably worse in Cluster 3, with 76% of patients exhibiting a score of 4–6 (P < 0.001). Visualization of K-means clustering results in (A) 2-dimensional and (B) 3-dimensional spaces. Table 3Clinical characteristics by clusterVariablesCluster 1 (n = 96)Cluster 2 (n = 88)Cluster 3 (n = 33)P valueAge65.4 ± 11.576.9 ± 8.068.4 ± 12.0< 0.001Sex, male57 (59.0)49 (56.0)22 (67.0)0.547Body mass index24.7 ± 3.523.7 ± 3.123.1 ± 3.00.057Hypertension73 (76.0)70 (80.0)12 (36.0)< 0.001Diabetes mellitus33 (34.0)44 (50.0)6 (18.0)0.003Atrial fibrillation9 (9.4)60 (68.0)12 (36.0)< 0.001Hyperlipidemia41 (43.0)40 (45.0)12 (36.0)0.666Coronary artery disease6 (6.3)22 (25.0)4 (12.0)0.001Peripheral artery disease0 (0.0)2 (2.3)1 (3.0)0.284Cancer3 (3.1)7 (8.0)6 (18.0)0.016Intervention3 (3.1)11 (13.0)12 (36.0)< 0.001Previous stroke4 (4.2)45 (51.0)2 (6.1)< 0.001Family history9 (9.4)5 (5.7)3 (9.1)0.621Smoking28 (29.0)7 (8.0)8 (24.0)0.001NIHSS2.9 ± 2.54.0 ± 3.211.7 ± 4.1< 0.001Lesion volume6.7 ± 12.53.5 ± 6.255.9 ± 59.8< 0.001Motor tract lesion volume0.2 ± 0.30.2 ± 0.32.0 ± 1.8< 0.001White matter hyperintensity12.5 ± 11.324.1 ± 20.910.1 ± 8.5< 0.001Stroke subtype< 0.001CE9 (10.0)48 (55.0)18 (55.0)LAA50 (52.0)29 (33.0)14 (42.0)Others37 (38.0)11 (12.0)1 (3.0)Premorbid mRS< 0.0010–188 (92.0)58 (66.0)28 (85.0)2–35 (5.0)21 (24.0)4 (12.0)4–63 (3.0)9 (10.0)1 (3.0)Initial mRS< 0.0010–148 (50.0)23 (26.0)0 (0.0)2–338 (40.0)36 (41.0)3 (9.0)4–610 (10.0)29 (33.0)30 (91.0)Follow-up mRS< 0.0010–150 (52.0)28 (32.0)0 (0.0)2–337 (39.0)33 (37.0)8 (24.0)4–69 (9.0)27 (31.0)25 (76.0)FMA-UE severity< 0.001Severe (≤ 33)5 (5.2)13 (15.0)23 (70.0)Mild to moderate (> 33)91 (94.8)75 (85.0)10 (30.0)Values are presented as mean ± standard deviation or number (%).NIHSS = National Institutes of Health Stroke Scale, CE = cardioembolism, LAA = large artery atherosclerosis, mRS = modified Rankin Scale, FMA-UE = Fugl-Meyer Assessment upper extremity. Clinical characteristics by cluster Values are presented as mean ± standard deviation or number (%). NIHSS = National Institutes of Health Stroke Scale, CE = cardioembolism, LAA = large artery atherosclerosis, mRS = modified Rankin Scale, FMA-UE = Fugl-Meyer Assessment upper extremity. DISCUSSION This study demonstrated that incorporating both clinical and imaging variables improves the prediction of upper limb function after stroke. The inclusion of imaging features, particularly motor tract lesion volume, yielded the highest predictive performance (Table 2), underscoring the added prognostic value of lesion characteristics captured on early MRI. These findings are consistent with previous studies highlighting the pivotal role of corticospinal tract integrity and involvement in motor recovery after stroke.16,17 Additionally, cluster analysis identified three distinct patient subgroups, suggesting the potential for individualized rehabilitation strategies based on functional impairment patterns. Despite the well-recognized prognostic utility of imaging-derived biomarkers—particularly those reflecting lesion location and corticospinal tract integrity—their adoption in routine clinical practice has been limited. This is primarily due to the technical demands, time requirements, and expertise needed for manual image processing and quantification. Consequently, the use of imaging features for individualized prognostication has remained largely confined to research settings. Recent advances in artificial intelligence and automated image analysis, including deep learning-based segmentation, now enable efficient and accurate extraction of clinically relevant imaging biomarkers. In this study, we utilized automatically derived imaging variables—including motor tract lesion volume, WMH volume, and stroke subtype classification—to predict upper limb motor impairment using the FMA-UE score. Importantly, these structured imaging features can aid clinical decision-making by providing insight into the anatomical and etiological factors most strongly associated with motor outcomes. Although fully automated artificial intelligence models that directly predict functional outcomes from raw imaging data represent an exciting and evolving frontier, they often operate as “black boxes,” offering limited transparency into the rationale behind their predictions.18 By incorporating clinically guided feature selection, models can prioritize pathophysiologically meaningful inputs, thereby enhancing both the accuracy and interpretability of imaging-based prognostic tools.19 Clustering based on clinical and imaging data revealed distinct patient subgroups with varying profiles of motor impairment and recovery potential. Among the three clusters identified, Cluster 3 exhibited the most severe neurological deficits, the largest lesion volumes, and the greatest motor tract involvement. These patients also demonstrated the poorest functional outcomes, with most showing high mRS scores at follow-up, indicating substantial disability. Given this profile, individuals in Cluster 3 are likely to require the most intensive and targeted rehabilitation interventions. Cluster 2 primarily comprised older patients with a high burden of comorbidities, including atrial fibrillation and prior stroke. The characteristics of this subgroup suggest an elevated risk of recurrent vascular events, highlighting the need for targeted secondary prevention strategies, such as rigorous blood pressure management, anticoagulation, and structured patient education on lifestyle modification and treatment adherence. Previous clustering studies in stroke populations have also identified distinct subgroups and emphasized the potential benefits of tailored preventive interventions, particularly for older patients with multiple comorbidities, to reduce the risk of recurrence.20,21 In contrast, Cluster 1 included younger patients with milder impairments and more favorable functional trajectories, suggesting that intensive rehabilitation may be less critical in this group. For these individuals, interventions prioritizing self-management education, discharge planning, and home-based exercise may be sufficient to support continued recovery and maintain independence. By aligning therapy intensity and content with phenotypic group characteristics, rehabilitation can be more effectively individualized, enhancing resource efficiency while supporting optimal outcomes across a heterogeneous patient population. Clustering offers clinical and interpretive advantages over traditional severity classification by capturing multidimensional patterns—integrating comorbidities, lesion characteristics, and functional status—that conventional single-score systems cannot fully represent, thereby providing a more nuanced framework for prognostication and personalized rehabilitation planning.22,23 This study had several limitations. First, it was conducted at a single tertiary care center with a relatively small sample size, which may limit the generalizability of the findings to broader stroke populations. Second, the retrospective study design introduces inherent risks of selection bias and unmeasured confounding, particularly as only patients who underwent both brain MRI and functional assessment were included. Third, motor function was evaluated exclusively using the FMA-UE score, which, although validated, provides a limited perspective on functional status and does not capture longitudinal recovery or broader outcomes such as quality of life and social participation. Lastly, the predictive models did not incorporate neurophysiological or advanced neuroimaging measures—such as transcranial magnetic stimulation, diffusion tensor imaging, or functional connectivity analyses—which have been shown in prior studies to enhance the accuracy of motor outcome prediction. Future research should address these limitations through prospective, multicenter studies incorporating more comprehensive outcome measures and multimodal data integration. This study demonstrated that combining clinical and imaging data through explainable, automated analysis significantly improves the prediction of post-stroke motor impairment. In particular, incorporating motor tract lesion volume from early MRI enhanced model performance, underscoring the importance of this metric for functional prognostication. Additionally, cluster analysis identified distinct patient subgroups with divergent recovery trajectories, highlighting the need for tailored rehabilitation strategies. By identifying patients according to their risk of poor outcomes or recurrent events, this approach may inform the intensity and direction of both therapeutic interventions and secondary prevention. Together, these findings highlight the growing value of quantitative imaging and unsupervised learning techniques in advancing precision stroke rehabilitation. Footnotes Disclosure: Wi-Sun Ryu and Bokgi Hwang are employees of JLK Inc. Author Contributions: Conceptualization: Cha S, Lee SH.Data curation: Cha S, Ryu WS, Hwang B.Formal analysis: Ryu WS, Hwang B.Investigation: Cha S.Methodology: Cha S, Lee SH, Kim DY.Project administration: Kim DY, Lee SH.Supervision: Kim DY, Lee SH.Visualization: Ryu WS, Hwang B.Writing - original draft: Cha S.Writing - review & editing: Ryu WS, Hwang B, Kim DY, Lee SH. Conceptualization: Cha S, Lee SH. Conceptualization: Cha S, Lee SH. Data curation: Cha S, Ryu WS, Hwang B. Data curation: Cha S, Ryu WS, Hwang B. Formal analysis: Ryu WS, Hwang B. Formal analysis: Ryu WS, Hwang B. Investigation: Cha S. Investigation: Cha S. Methodology: Cha S, Lee SH, Kim DY. Methodology: Cha S, Lee SH, Kim DY. Project administration: Kim DY, Lee SH. Project administration: Kim DY, Lee SH. Supervision: Kim DY, Lee SH. Supervision: Kim DY, Lee SH. Visualization: Ryu WS, Hwang B. Visualization: Ryu WS, Hwang B. Writing - original draft: Cha S. Writing - original draft: Cha S. Writing - review & editing: Ryu WS, Hwang B, Kim DY, Lee SH. Writing - review & editing: Ryu WS, Hwang B, Kim DY, Lee SH. SUPPLEMENTARY MATERIALS Supplementary Table 1 Full multivariable regression coefficients from Model 3 predicting Fugl–Meyer Assessment upper extremity scores Supplementary Table 2 Comparison of predictive models for Fugl–Meyer Assessment upper extremity scores using clinical, functional, and imaging variables in patients without a history of stroke or premorbid mRS ≥ 2 (n = 167) Supplementary Fig. 1 Silhouette index values for K-means and agglomerative clustering across candidate numbers of clusters (K = 2–15). References Associated Data Supplementary Materials Supplementary Table 1 Full multivariable regression coefficients from Model 3 predicting Fugl–Meyer Assessment upper extremity scores Supplementary Table 2 Comparison of predictive models for Fugl–Meyer Assessment upper extremity scores using clinical, functional, and imaging variables in patients without a history of stroke or premorbid mRS ≥ 2 (n = 167) Supplementary Fig. 1 Silhouette index values for K-means and agglomerative clustering across candidate numbers of clusters (K = 2–15).