Table 1 Reinforcement learning and similar artificial intelligence (AI) approaches for improving cardiovascular care

Deep learning

Neural network-based, can be supervised or unsupervised

Requires large, high-quality datasets for training

Highly flexible; excels at complex tasks like imaging analysis

“Black box” effect, large computational demands, risk of overfitting

Automated interpretation of echocardiograms or coronary CT angiography

Hybrid or ensemble methods

Combination of multiple algorithms (e.g., supervised + RL)

Integrates diverse data sources; can switch algorithms dynamically

Balances strengths of different models, potentially increasing accuracy & robustness

Complexity, potential for overfitting, and higher computational cost

Multi-stage CAD risk assessment where an RL agent refines a supervised model treatment recommendations

Reinforcement Learning

Iterative decision-making process

Learns optimal actions from sequential patient data (e.g., CAD registry)

Actively improves treatment strategies by continuously updating policies based on outcomes

Requires careful offline validation and interpretability; may differ from standard physician patterns

Offline RL models for revascularization decisions, balancing PCI vs. CABG (Ghasemi et al.)

Supervised Learning

Predictive modeling

Relies on labeled datasets (e.g., mortality or MACE outcomes)

Excellent for classification (e.g., risk stratification), quick to train and deploy

Limited adaptation to changing data; focuses on static predictions rather than ongoing decision-making

Predicting likelihood of major adverse cardiac events post revascularization

Unsupervised learning

Clustering or dimensionality reduction

Uses unlabeled patient data (e.g., imaging, laboratory results)

Identifies hidden patterns or phenotypes without prior assumptions

Results can be harder to interpret; no direct linkage to specific outcomes

Segmenting heart failure subtypes or clustering CAD phenotypes