Machine Learning-Based Study on Risk Classification of Cardiovascular Diseases Using Dynamic Temporal Features of Carotid Ultrasound

This article introduces an innovative medical AI study that achieves accurate prediction of cardiovascular disease risk by analyzing dynamic waveform data from carotid ultrasound, using time-series feature extraction and machine learning classifiers. The study integrates multi-modal physiological signals and demonstrates their great potential in medical diagnosis.

Cardiovascular disease is the leading cause of death globally, and early risk identification is crucial. Traditional assessments rely on static indicators (such as plaque area, peak Doppler velocity) and cannot capture the dynamic change characteristics of blood vessels. The innovation of this study lies in introducing dynamic arterial waveform analysis, collecting three synchronized temporal signals (carotid artery diameter, Doppler blood flow velocity, brachial artery pressure), and combining machine learning to build a risk prediction model.

Data was collected from 150 patients over 50 years old, including carotid ultrasound examination and ankle-brachial pulse wave velocity (PWV) measurement data. Three waveform extraction methods: diameter waveform using the CAROLAB speckle tracking algorithm, blood flow velocity waveform using the gray threshold method to extract the envelope, and pressure waveform recorded via a plethysmography system. Preprocessing steps: remove slow drift, low-pass filtering (8-10Hz), min-max normalization, synchronization to ECG R waves and resampling to 128Hz, and extraction of data for a single cardiac cycle.

Two strategies were used for feature extraction: 1. TSFEL to extract 201 statistical, time-domain, and frequency-domain features; 2. Shapelet transformation to identify discriminative short subsequences. The mRMR algorithm was used to select 15 optimal features. Eight classifiers were evaluated: the fused feature set + Random Forest performed best with an accuracy of 90%, AUC of 0.95, and F1 score of 0.80; followed by LightGBM (87%) and XGBoost (83%), while traditional SVM (66%) and k-NN (80%) had poorer performance.

Ablation experiments showed: accuracy of 77% for pressure waveform alone, 83% for diameter waveform alone, and 77% for velocity waveform alone; after fusing the three waveforms, the accuracy increased to 90%. This proves that multi-modal signals carry complementary information, which can improve prediction accuracy and robustness.

Prospects: Can be integrated into routine physical examinations to realize automatic identification and early warning of high-risk groups, assisting primary care doctors in decision-making. Challenges: Need for standardized collection protocols and quality control; model interpretability needs to be improved; multi-center large-scale datasets are required to verify generalization ability.

The code is open-sourced on GitHub, based on Python 3.10+, and depends on libraries such as NumPy, Pandas, and Scikit-learn. It includes complete process scripts: low-pass filtering, normalization, feature extraction, selection, and model training. The code structure is clear and highly scalable; some Windows paths need to be modified by users.

This study demonstrates the potential of machine learning in medical image analysis, constructing a high-precision risk classification system through multi-modal dynamic temporal features. In the future, it can be combined with deep learning technology and larger-scale clinical data to improve the system's accuracy, robustness, and real-time performance, contributing to the prevention and control of cardiovascular diseases.