A machine learning approach using echocardiographic and hematological parameters to classify cerebral small vessel...

Cerebral small vessel disease (CSVD) is a leading cause of stroke and dementia and is associated with cardiac and hematological biomarkers. Although brain magnetic resonance imaging (MRI) remains the reference standard for assessing CSVD burden, its cost and contraindications limit widespread clinical application.

To develop machine learning (ML) models based on routine echocardiographic and hematological parameters for CSVD burden stratification.

In this retrospective study, 446 patients with MRI-confirmed CSVD were classified into low-burden (CSVD-L, n = 314) and high-burden (CSVD-H, n = 132) groups using the validated total CSVD score (0-2 vs. 3-4). Clinical characteristics, echocardiographic parameters, and hematological biomarkers were compared between the two groups. Variables significant on univariate analysis were entered into eight ML algorithms for binary classification of CSVD burden. Model performance and feature importance were assessed.

Patients with high CSVD burden exhibited distinct echocardiographic alterations suggestive of structural remodeling and subclinical diastolic dysfunction, as well as significantly lower fibrin(ogen) degradation products (FDP), D-dimer, and fibrin formation rate (FbFR). The Random Forest (RF) model achieved the highest discriminative performance (AUC 0.885). Shapley Additive exPlanations (SHAP) analysis identified FbFR as the most important feature.

Integration of routine echocardiographic and hematological parameters within a machine learning framework can effectively stratify CSVD burden. The RF model shows promise as a supplementary tool for CSVD risk stratification in settings where MRI is unavailable. Future prospective external validation is warranted to confirm the generalizability and clinical utility of this approach.