Machine learning combined with resting-state functional MRI to characterize functional brain differences in post-stroke depression.
Authors
Affiliations (5)
Affiliations (5)
- First Clinical Medical College, Nanjing University of Chinese Medicine, Nanjing, Jiangsu, China.
- Liaocheng People's Hospital, Liaocheng, Shandong, China.
- Taicang TCM Hospital Affiliated to Nanjing University of Chinese Medicine, Taicang, Jiangsu, China.
- Dalian Medical University, Dalian, China.
- Vietnam National Hospital of Acupuncture, Hanoi, Vietnam.
Abstract
Post-stroke depression (PSD) is a common neuropsychiatric condition after stroke, but its resting-state functional imaging correlates remain incompletely characterized. This study examined multi-level resting-state functional differences between patients with PSD and healthy controls and evaluated whether interpretable machine learning could identify candidate imaging features associated with PSD. Fifty patients with PSD and 50 age- and sex-matched healthy controls underwent resting-state functional MRI. Four complementary imaging indices were extracted using the AAL atlas: amplitude of low-frequency fluctuation (ALFF), regional homogeneity (ReHo), degree centrality (DC), and ROI-to-ROI functional connectivity (FC). Imaging features were adjusted for demographic variables, head motion, medication use, and stroke-related factors where appropriate. Candidate features with significant between-group differences were further reduced using LASSO regression. Nine machine-learning classifiers were trained and compared under the same feature-selection framework. Model performance was assessed primarily by ROC-AUC. SHapley Additive exPlanations (SHAP) were used to examine the contribution of individual features to the best-performing model. Patients with PSD showed distributed resting-state functional differences involving cingulate, thalamic, prefrontal, insular, posterior default-mode, and visual-associated regions. Twenty-nine candidate features differed between groups, including 7 ReHo, 8 ALFF, 6 DC, and 8 FC features. LASSO retained 10 core features, with a cross-validated AUC of 0.878. Among the nine classifiers, the Extra Trees model achieved the highest independent test-set performance, with an AUC of 0.889. SHAP analysis indicated that the most influential features included DC in the left anterior cingulate and paracingulate gyri, ReHo in the left thalamus, FC between the left precuneus and left calcarine cortex, ALFF in the right precuneus, and ALFF in the left angular gyrus. Within the PSD group, moderate depression was associated with higher ALFF in the left insula and lower precuneus-calcarine connectivity compared with mild depression. Patients with PSD showed multi-level resting-state functional differences compared with healthy controls. Interpretable machine learning identified a set of candidate rs-fMRI features with plausible neurobiological relevance. These findings require validation in larger longitudinal cohorts that include post-stroke patients without depression to clarify their specificity and clinical utility.