Frequency-specific alterations of spontaneous brain activity in stroke patients with upper limb motor dysfunction: A multi-metric rs-fMRI and machine learning study.

Authors

Abstract

Post-stroke upper limb motor dysfunction is associated with complex alterations in brain function, but the frequency-specific characteristics of spontaneous local neural activity remain incompletely understood. This study investigated multi-metric resting-state functional magnetic resonance imaging (rs-fMRI) alterations in stroke patients with upper limb motor dysfunction and explored their discriminative value using machine-learning classifiers. A total of 70 stroke patients and 45 healthy controls (HCs) underwent rs-fMRI. Fractional amplitude of low-frequency fluctuations (fALFF), percent amplitude of fluctuation (PerAF), and wavelet transform-based amplitude of low-frequency fluctuations (Wavelet-ALFF) were calculated across the conventional band (0.01-0.08 Hz), Slow-4 band (0.027-0.073 Hz), and Slow-5 band (0.01-0.027 Hz). Voxel-wise group comparisons were performed with age, sex, and mean framewise displacement as covariates. Lesion mapping, lesion-volume analysis, and a subcortical-lesion subgroup sensitivity analysis were conducted to assess lesion-related effects. Regional features showing significant group differences were used to construct support vector machine (SVM), Random Forest, and XGBoost classifiers, and model performance was evaluated on an independent testing set. Compared with HCs, stroke patients showed widespread frequency-specific alterations in cerebellar, visual, cingulo-motor, temporal, insular, subcortical, and frontal regions. Core abnormalities in cerebellar, visual, cingulate, supplementary motor, and frontal regions were largely retained in the subcortical-lesion subgroup. Lesion volume was negatively correlated with Fugl-Meyer Assessment for Upper Extremity scores. Several nominal associations were observed between Slow-5 fALFF features in visual regions and motor impairment severity, but none survived false discovery rate correction. Among all testing-set models, the SVM classifier using combined multi-metric Slow-4 and Slow-5 features achieved the highest numerical performance, with an AUC of 0.956, accuracy of 0.912, sensitivity of 0.905, and specificity of 0.923. However, DeLong tests showed no statistically significant superiority of sub-frequency features over conventional-band features. Stroke patients with upper limb motor dysfunction exhibit distributed, frequency-specific alterations in spontaneous local brain activity. Multi-metric rs-fMRI features, particularly combined sub-frequency features, may have exploratory discriminative value, but external validation is needed before they can be considered robust imaging biomarkers.

Tags

Topics