Cardiovascular disease/stroke risk stratification in deep learning framework: a review

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Cardiovascular disease (CVD) remains a leading global cause of mortality, necessitating advanced, non-invasive risk identification methods.
• Conventional CVD risk prediction models struggle with the complex, non-linear relationships in multi-ethnic populations.
• Existing reviews on machine learning for CVD risk lack comprehensive deep learning (DL) integration.

Purpose of the Study:

• To review and analyze solo deep learning (SDL) and hybrid deep learning (HDL) techniques for cardiovascular disease (CVD) risk stratification.
• To evaluate the characteristics, applications, and validation of various DL architectures in CVD risk assessment.
• To identify and discuss the risk of bias in AI systems used for CVD risk stratification.

Main Methods:

• A systematic review of 286 DL-based CVD studies was conducted using the PRISMA model across major scientific databases.
• Analysis focused on SDL and HDL architectures, plaque tissue characterization, and Electrocardiogram (ECG)-based solutions.
• Risk of bias was assessed using tools such as PROBAST and ROBINS-I, with a focus on ground truth selection in UNet-based frameworks.

Main Results:

• Convolutional Neural Network (CNN) algorithms are prevalent due to automated feature extraction.
• Ensemble-based DL techniques are emerging as a superior paradigm for CVD risk stratification over SDL and HDL.
• DL methods demonstrate high accuracy, reliability, and efficiency for CVD risk assessment.

Conclusions:

• Deep learning offers a powerful and promising approach for non-invasive CVD risk assessment and stratification.
• Ensemble DL methods are poised to become the leading techniques for CVD risk stratification.
• Mitigating risk of bias in DL requires multicenter data and rigorous clinical validation.