Multiclass machine learning vs. conventional calculators for stroke/CVD risk assessment using carotid plaque predictors with coronary angiography scores as gold standard: a 500 participants study.

Machine learning (ML)-based algorithms for cardiovascular disease (CVD) risk assessment have shown promise in clinical decisions. However, they usually predict binary events using only conventional risk factors. Our overall goal was to develop the "multiclass machine learning (MCML)-based algorithms" (labelled as AtheroEdge 3.0) and assess whether considering carotid ultrasound imaging fused with conventional risk factors can provide better CVD/stroke risk prediction than conventional CVD risk calculators (CCVRC). Carotid ultrasound and coronary angiography were performed on 500 participants. Stenosis in the coronary arteries was used to assign participants a coronary angiographic score (CAS). CVD/stroke risk was determined using three types of MCML algorithms: (i) support vector machine (SVM), (ii) random forest (RF), and (iii) extreme gradient boost (XGBoost). The performance of CVD risk assessment using MCML and CCVRC (such as Framingham Risk Score, the Systematic Coronary Risk Evaluation score, and the Atherosclerotic CVD) was evaluated on test patients against the CAS as the gold standard for each class using the area-under-the-curve (AUC) and classification accuracy. The mean percentage improvement in AUC and the mean absolute improvement in accuracy over CCVRC using 90% training and 10% testing protocol (labelled as K10) were ~ 105% and ~ 28%, respectively. Of all the three MCML systems, RF showed the best performance. Further, carotid image phenotypes showed the most effective clinical feature in AtheroEdge 3.0 performance. The AtheroEdge 3.0 using carotid imaging are reliable, accurate, and superior to traditional CVD risk scoring methods for predicting the CVD/stroke risk due to coronary artery disease.