Machine learning-enhanced echocardiography for screening coronary artery disease

Fig. 1. Study flowchart. All patients who underwent coronary angiography were divided into a data set of 818 patients for model training and classifier testing and a validation data set of 115 patients for validating the efficiency of the predictive model. AUC Area under the curve, ROC Receiver operator characteristics

Fig. 2. Framework of the proposed model for diagnosing CAD. A stepwise workflow to optimize the predictive learning model by screening and recomposing features, selecting superior classifiers, and identifying the convergent training data set. AUC Area under the curve, ROC Receiver operator characteristics

Fig. 3. Changes in the curve of ROC AUC with different feature combinations, including the ten best classifiers for each feature combination. Different feature panels screened by MDA (27 features), r coefficient (33 features), collinearity diagnostics (14 features), and the importance across all the tested classifiers (39 features). Stepwise combinations of 5 (overlapped features: GWI, G peak SL Full (Avg), Systolic BP., Diastolic BP, LASr R-Wave (Avg)), 30, 44, 59, 69, 79, 106, and 235 features were tested by classifiers

Fig. 4. Importance of 59 selected features. BMI body mass index, BP blood pressure, DM diabetes mellitus, FPG fasting plasma glucose, GCW global constructive work, GLS global longitudinal strain, GNW global negative work, GPW global positive work, GSCW global systolic constructive work, GSWW global systolic wasted work, GWE global work efficiency, GWI global work index, GWW global wasted work, LAScd left atrial longitudinal strain during conduit phase, LASct left atrial longitudinal strain during contraction phase, LASr left atrial longitudinal strain during reservoir phase, LDL_C low-density lipoprotein cholesterol, LVEF left ventricular ejection fraction, NegativeW negative work, PositiveW positive work, PSD peak strain dispersion, PSS peak systolic strain

Fig. 5. Testing accuracy of 18 classifiers using different feature combinations. The increased accuracy and superior performance (over 80% accuracy) were observed by the methods of catBoost, extra tree, gradient boosting, light gradient boosting, random forest, and Xgboost

Fig. 6. ROC curves of the proposed six classifiers with ROC AUC > 80%. Gradient boosting (light green) demonstrated the highest diagnostic efficacy among the six classifiers (ROC AUC: 0.852)

Fig. 7. Key items of echocardiographic measurements in the machine learning model for screening CAD