Stroke prediction in elderly patients with atrial fibrillation using machine learning combined clinical and left atrial appendage imaging phenotypic features

Figures

• 

  Fig. 1. (A) K-means clustering heatmap of clinical phenotype characteristics in patients with atrial fibrillation (AF). Z-score normalization was applied to the data for each component, with red representing relatively high values and blue representing relatively low values. (B) Scatter plot of K-means clustering for clinical phenotype characteristics. (C) Incidence rates of stroke events in AF patients in Groups A, B, and C. Group A: the phenotype with cardiac enlargement and decreased cardiac function, Group B: the normal phenotype, Group C: the phenotype with systemic metabolic and coagulation abnormalities

• 

  Fig. 2. (A) K-means clustering heatmap of left atrial appendage (LAA) imaging features in patients with atrial fibrillation (AF). (B) Scatter plot of K-means clustering for LAA imaging features in AF patients. (C) Incidence rates of stroke events in AF patients in Groups A, B, and C. Group A: the relatively normal LAA phenotype, Group B: the phenotype with enlarged LAA structure, Group C: the phenotype with complex LAA morphology

• 

  Fig. 3. Feature selection process for five machine learning models. The shaded area represents the feature set with the maximum area under the receiver operating characteristic curve (AUROC), and the illustration only displays the top 30 features. AF: atrial fibrillation; LDL: low density lipoprotein; LAA: Left Atrial Appendage; INR: International Normalized Ratio; LVEF: Left Ventricular Ejection Fraction; LA: Left Atrial Diameter

• 

  Fig. 4. (A) The AUROC of the machine learning model and the traditional score of the derivation group. (B) The stroke prediction model of the derivation group and the H-L test chart of CHA2DS2-VASc. (C) The derivation group supervised the decision curve analysis of the machine learning model and CHA2DS2-VASc score. NB: Naive Bayes; LR: logical regression; RF: random forest; SVM: support vector machine; XGBoost: extreme gradient boosting

• 

  Fig. 5. (A) The AUROC of the machine learning model and the traditional score of the validation group. (B) The stroke prediction model of the validation group and the H-L test chart of CHA2DS2-VASc. (C) The validation group supervised the decision curve analysis of the machine learning model and CHA2DS2-VASc score. NB: Naive Bayes; LR: logical regression; RF: random forest; SVM: support vector machine; XGBoost: extreme gradient boosting