Machine Learning for Selecting Important Clinical Markers of Imaging Subgroups of Cerebral Small Vessel Disease Based on a Common Data Model

Differences in the imaging subgroups of cerebral small vessel disease (CSVD) need to be further explored. First, we use propensity score matching to obtain balanced datasets. Then random forest (RF) is adopted to classify the subgroups compared with support vector machine (SVM) and extreme gradient boosting (XGBoost), and to select the features. The top 10 important features are included in the stepwise logistic regression, and the odds ratio (OR) and 95% confidence interval (Cl) are obtained. There are 41 290 adult inpatient records diagnosed with CSVD. Accuracy and area under curve (AUC) of RF are close to 0.7, which performs best in classification compared to SVM and XGBoost. OR and 95% Cl of hematocrit for white matter lesions (WMLs), lacunes, microbleeds, atrophy, and enlarged perivascular space (EPVS) are 0.9875 (0.9857–0.9893), 0.9728 (0.9705–0.9752), 0.9782 (0.9740–0.9824), 1.0093 (1.0081–1.0106), and 0.9716 (0.9597–0.9832). OR and 95% Cl of red cell distribution width for WMLs, lacunes, atrophy, and EPVS are 0.9600 (0.9538–0.9662), 0.9630 (0.9559–0.9702), 1.0751 (1.0686–1.0817), and 0.9304 (0.8864–0.9755). OR and 95% Cl of platelet distribution width for WMLs, lacunes, and microbleeds are 1.1796 (1.1636–1.1958), 1.1663 (1.1476–1.1853), and 1.0416 (1.0152–1.0687). This study proposes a new analytical framework to select important clinical markers for CSVD with machine learning based on a common data model, which has low cost, fast speed, large sample size, and continuous data sources.