Data-driven identification of gait phenotypes in cerebral small vessel disease
He-Jiao Mao, Yan-Yan Wang, Ning Su, Fei Han, Tian-Qi Bao, Zi-Ang Pan, Zi-Yue Liu, Jiang-Xia Zhang, Xiang-Min Fan, Jun Ni, Li-Xin Zhou, Ming Yao, Feng Tian, Yi-Cheng ZhuBackground
Cerebral small vessel disease (CSVD) often manifests with gait disturbance, yet patients with similar imaging burden may present heterogeneous gait phenotypes. Systematic, data-driven subtyping of CSVD-related gait impairments remains lacking.
Objectives
To identify CSVD gait subtypes using quantitative gait analysis and unsupervised clustering and to explore their associations with clinical features, imaging markers, and prognosis.
Methods
We enrolled 118 patients with severe CSVD burden in an observational cohort study. Forty-six quantitative gait parameters were assessed using a vision-based motion-capture system and reduced to 28 variables and 8 principal components. K-means identified three gait subtypes; robustness was assessed using random seeds, bootstrapping, and alternative algorithms. Clinical and imaging characteristics were compared across subtypes. Cox regression and Kaplan–Meier analyses examined short-term prognostic associations, with proportional-hazards testing and Firth-corrected sensitivity analyses.
Results
Three subtypes were identified. Stable gait type (n=92) showed relatively preserved performance; temporal dysfunction type (n=10) exhibited prolonged stance and double-support phases; and rhythmic variable type (n=16) demonstrated increased rhythm variability. The three-cluster solution was generally supported by sensitivity and resampling analyses, although bootstrap stability was lower for the rhythmic variable type. The temporal dysfunction type had worse motor, cognitive, and functional measures and, over a short median follow-up of 5 months, a higher risk of disability (HR=13.46, P=0.003) and composite outcomes (HR=5.62, P=0.023), with similar estimates after Firth correction. Exploratory imaging analyses suggested that periventricular white matter hyperintensities were linked to phasic abnormalities and impaired postural control; basal ganglia lacunes to reduced velocity; lobar cerebral microbleeds to impaired postural control; and corpus callosum/brainstem atrophy to phasic abnormalities.
Conclusion
Machine learning identified three distinct CSVD gait subtypes with differing clinical profiles and imaging correlates. These subtypes showed preliminary short-term prognostic associations that require external and longitudinal validation, supporting the potential utility of gait phenotyping for functional assessment and risk stratification.