Machine learning combined with fMRI identifies dynamic brain network alterations in post-stroke depression: A dual-center cross-sectional study.
Authors
Affiliations (7)
Affiliations (7)
- Pathophysiology Teaching and Research Section, Jining Medical University, No. 133 Hehua Road, Taibaihu New District, Jining City, Shandong Province, China. Electronic address: [email protected].
- Pathophysiology Teaching and Research Section, Jining Medical University, No. 133 Hehua Road, Taibaihu New District, Jining City, Shandong Province, China. Electronic address: [email protected].
- Department of Medical Imaging, The Affiliated Hospital of Northwest University, No. 10 East Section of Fengcheng 3rd Road, Weiyang District, Xi'an City, Shaanxi Province, China. Electronic address: [email protected].
- Pathophysiology Teaching and Research Section, Jining Medical University, No. 133 Hehua Road, Taibaihu New District, Jining City, Shandong Province, China. Electronic address: [email protected].
- Pathophysiology Teaching and Research Section, Jining Medical University, No. 133 Hehua Road, Taibaihu New District, Jining City, Shandong Province, China. Electronic address: [email protected].
- Pathophysiology Teaching and Research Section, Jining Medical University, No. 133 Hehua Road, Taibaihu New District, Jining City, Shandong Province, China. Electronic address: [email protected].
- Rehabilitation Department, Jining First People's Hospital, No. 6 Jiankang Road, Rencheng District, Jining City, Shandong Province, China. Electronic address: [email protected].
Abstract
Post-stroke depression (PSD) is under-recognized and associated with poorer rehabilitation outcomes and quality of life. We characterized PSD-related alterations in large-scale brain network dynamics and evaluated their cross-center classification performance. This dual-center cross-sectional study included 300 patients with post-stroke motor impairment: 100 with and 100 without PSD in the development cohort, and 50 per group in the external validation cohort. Resting-state functional MRI, sliding-window dynamic functional connectivity, and clustering across K = 2-8 state resolutions were combined with nested cross-validation, nine machine-learning classifiers, and SHAP interpretation. Sensitivity analysis excluded users of antidepressants, anxiolytics, or sedative-hypnotics. The five-state solution provided the best balance of discrimination and interpretability. The support vector machine achieved internally cross-validated and external validation AUCs of 0.838 and 0.784, respectively. Higher proportion and longer dwell time of a globally integrated state, lower proportion and dwell time of a stable modular state, and higher self-transition probability of a low-integration state shifted classification toward PSD. After medication users were excluded, the artificial neural network under the five-state solution achieved internally cross-validated and external validation AUCs of 0.798 and 0.793, respectively; a higher proportion of the globally integrated state also remained among the most informative features. PSD was associated with altered temporal organization and reduced stability of large-scale functional networks. Dynamic state features may serve as interpretable candidate imaging markers for PSD classification, but longitudinal multicenter validation is required before clinical application.