Sex-Specific Classification Models for Cognitive Function in Cerebral Small Vessel Disease Based on MRE and Machine Learning.
Authors
Affiliations (3)
Affiliations (3)
- Shuguang Hospital Affiliated with Shanghai University of Traditional Chinese Medicine, Department of Radiology.
- Shuguang Hospital Affiliated with Shanghai University of Traditional Chinese Medicine, Department of Neurology.
- Shuguang Hospital Affiliated with Shanghai University of Traditional Chinese Medicine, Department of Radiology. Electronic address: [email protected].
Abstract
Early identification of cognitive impairment is crucial for preventing dementia progression. Magnetic resonance elastography (MRE), combined with traditional imaging markers, provides a novel dimension for cognitive assessment in cerebral small vessel disease (CSVD). To develop machine learning-based classification models for cognitive function in CSVD patients by integrating MRE brain mechanical parameters, white matter hyperintensity burden, and clinical indicators, and to analyze sex differences in model performance. 100 CSVD subjects underwent MRI scanning including FLAIR for white matter hyperintensities, T1WI for hippocampal atrophy assessment, and MRE for measuring brain mechanical properties across 12 regions. Data were split 8:2 (train:test, stratified); within the training set only, candidate features were ranked by Spearman correlation with MoCA score and a final 13-feature subset was selected by Recursive Feature Elimination (RFE). Subjects were stratified by sex (all-sex: n=100; male: n=42; female: n=58), and 10 machine learning models were compared for each group using 5-fold cross-validation (cvAUC), with SHAP analysis used to interpret feature importance. A 13-feature subset was identified, predominantly MRE stiffness parameters. Naive Bayes performed best in the overall cohort (cvAUC=0.704, 95% CI 0.652-0.756), Extra Trees in males (cvAUC=0.828, 95% CI 0.635-1.000), and XGBoost in females (cvAUC=0.617, 95% CI 0.370-0.863). The female subgroup's confidence interval spanned chance-level discrimination. Using the same feature set, classification accuracy was numerically higher in males than in females. SHAP analysis showed atrophy and white matter hyperintensity burden markers dominated in males, whereas right angular gyrus stiffness and other MRE parameters dominated in the overall cohort and in females. Using an identical feature set and modeling pipeline, machine learning models integrating MRE showed markedly different performance between sexes, indicating that a single sex-agnostic model does not generalize equally to men and women. This performance gap, together with divergent SHAP patterns, underscores the need for sex-stratified modeling in MRE-based cognitive screening for CSVD.