What question did this study set out to answer?

The research aims to create and validate AI tools for differentiating cognitive conditions in patients with cerebral small vessel disease.

February 2, 2026

Abstract TP070: Early screening framework for cognitive impairment in cerebral small vessel disease: a comparative study of national and single-center data

Key Points

The research aims to create and validate AI tools for differentiating cognitive conditions in patients with cerebral small vessel disease.
Analyzed two cohorts: a nationwide multicenter dataset and a single-center dataset.
Used various AI models including TabPFN, Random Forest, XGBoost, and LightGBM to assess cognitive status.
Cognitive impairment was measured using the Montreal Cognitive Assessment (MoCA) scores.
Performance assessed with area under ROC curve (AUC) through five-fold cross-validation.
Investigated relationships between imaging and physiological features and cognitive performance in a single-center cohort.
TabPFN achieved an AUC of 0.821 in the nationwide cohort, outperforming XGBoost at 0.795.
In the single-center dataset, Random Forest displayed an AUC of 0.841 despite class imbalance.
Multimodal features, especially imaging and physiological, improved diagnostic performance for MCI patients.
For verbal fluency, CTT-1, and Stroop test, AUCs were 0.83, 0.836, and 0.854, respectively.

Abstract

Background: Cerebral small vessel disease (CSVD) is a major cause of vascular cognitive impairment (VCI), yet diagnosis remains difficult due to heterogeneous patient profiles and complex manifestations. This study aimed to develop and validate AI-based tools for distinguishing normal cognition (NC) and mild cognitive impairment (MCI) in CSVD. Methods: Two CSVD cohorts were analyzed: (1) a nationwide multicenter dataset (N=4185) with standardized demographic and clinical variables; and (2) a single-center dataset from Beijing Tiantan Hospital (N=74) including demographic, structural MRI, heart rate variability, EEG, hemodynamic, functional MRI (fMRI), and diffusion tensor imaging (DTI) features. Cognitive status was determined by Montreal Cognitive Assessment (MoCA) scores. Models were developed using the tabular prior-data fitted network (TabPFN) and conventional machine learning methods (Random Forest, XGBoost, LightGBM). Performance was assessed with stratified five-fold cross-validation and independent testing, using area under the ROC curve (AUC). Secondary analyses in the single-center cohort examined associations between multimodal features and domain-specific outcomes: verbal fluency test (VFT-composite), color trails test part 1 (CTT-1), and Stroop test part 3. Results: In the nationwide cohort, TabPFN achieved the best performance (AUC=0.821), slightly surpassing XGBoost (AUC=0.795), while requiring no feature selection. In the single-center multimodal dataset, RF outperformed TabPFN (AUC=0.841, accuracy=68.9%), likely due to small sample size and class imbalance. Adding imaging and physiological features improved classification over demographic-only inputs, especially in MCI patients. For domain-specific outcomes, advanced neuroimaging (fMRI, DTI) contributed most. AUCs were 0.83 for VFT-composite, 0.836 for CTT-1, and 0.854 for Stroop part 3. Conclusions: AI-based models support auxiliary diagnosis of MCI in CSVD. Nationwide TabPFN models provide scalable screening, while multimodal single-center models allow refined evaluation. Advanced imaging and physiological features enhance domain-specific cognitive assessment, informing targeted diagnostic and intervention strategies.

Read Full Paperexternally

Bookmark

View Full Paper

Cite This Study

mao et al. (Thu,) studied this question.

synapsesocial.com/papers/6980fcfcc1c9540dea80eb04 https://doi.org/https://doi.org/10.1161/str.57.suppl_1.tp070

Bookmark

View Full Paper