400 Section - Imaging-Enhanced Prediction of Upper Limb Motor Outcomes After Stroke Using Automated Segmentation and Cluster Analysis

Objectives : To determine whether explainable, imaging-derived features from automated brain MRI analysis can improve the prediction of post-stroke upper limb motor impairment and to identify distinct recovery profiles using cluster analysis.

Design: We performed a retrospective cohort study of 242 patients with acute ischemic stroke admitted to Asan Medical Center between July 1, 2022, and July 31, 2023. All patients underwent brain MRI and standardized upper limb motor assessments and were subsequently evaluated in the Department of Rehabilitation Medicine. Patients with incomplete clinical or imaging data were excluded. Automated MRI processing using validated commercial software extracted quantitative biomarkers including infarct volume, motor tract lesion volume, TOAST classification, and white matter hyperintensity burden. Clinical, functional, and imaging variables were incorporated into hierarchical regression models to identify predictors of motor impairment. Unsupervised machine learning with K-means clustering was additionally applied to classify recovery subgroups based on multimodal data patterns. Motor outcomes were assessed using the upper extremity Fugl-Meyer Assessment, and functional status was evaluated with Modified Rankin Scale scores at baseline and follow-up.

Results: Incorporating imaging features improved the prediction of upper limb FMA scores (adjusted R² increased from 0.556 to 0.595 with imaging variables). Motor tract lesion volume emerged as a key imaging predictor. Cluster analysis identified three distinct patient profiles:
Cluster 1 (n=96): Younger, low comorbidity burden, favorable motor and functional outcomes.
Cluster 2 (n=88): Older, high prevalence of atrial fibrillation and previous strokes, moderate outcomes.
Cluster 3 (n=33): Severe deficits, large infarct and motor tract involvement, poorest outcomes with 76% showing follow-up mRS 4–6 (p < 0.001).

Conclusion: Automated extraction of imaging-derived variables, especially motor tract lesion volume, enhances prediction of post-stroke motor impairment. Additionally, unsupervised clustering reveals clinically relevant subgroups with distinct recovery potentials. These findings support a precision rehabilitation framework, enabling stratified therapeutic and preventive strategies based on individualized profiles.