Machine Learning-Based Cardiovascular Risk Classification Using Dynamic Time-Series Features From Carotid Duplex Sonography

Abstract

Although carotid duplex sonography provides comprehensive hemodynamic and morphological information, clinical assessments typically rely on limited static features and flow velocity measurements. This study investigated the feasibility of applying machine learning to classify cardiovascular disease (CVD) risk using dynamic arterial time-series data derived from carotid diameter, flow velocity, and brachial pulse pressure waveforms. Signal-derived and shapelet-based features were extracted and selected using the minimum redundancy maximum relevance algorithm, and eight classifiers were evaluated. Integrating signal-derived and shapelet-based features improved classification accuracy to 0.90, compared with 0.77 and 0.80 for each feature set alone. Among the models, the random forest classifier achieved the best performance (accuracy 0.90, AUC 0.95, F1-score 0.80). The combined use of diameter, velocity, and pressure waveforms yielded the highest accuracy, demonstrating that dynamic carotid time-series data provide complementary information for CVD risk classification. These findings highlight the potential of machine learning-based ultrasound time-series analysis as a complementary, data-driven approach to conventional static imaging for CVD assessment.