Abstract
Abstract
Background
Cardiovascular toxicity (CVT) is a major concern after radiotherapy (RT), contributing to morbidity and mortality among cancer survivors. Artificial intelligence (AI) may improve risk prediction and RT planning; however, its role in RT-associated CVT remains unclear. This study aimed to systematically evaluate AI applications and study quality in this field.
Methods
A PRISMA-guided systematic review of PubMed, Ovid EMBASE, Cochrane Library, and Web of Science was conducted through October 1, 2025. Eligible studies included original human research in English applying AI to CVT or imaging in cancer populations receiving RT. Predictive and imaging studies were evaluated using TRIPOD + AI/PROBAST and CLAIM/QUADAS-2, respectively. Exploratory meta-analysis of performance metrics was conducted where feasible.
Results
Sixty-five studies were included, comprising AI prediction models (
n
= 31, 48%) and cardiovascular imaging applications (
n
= 34, 52%). Deep learning was the most common approach (45/65, 69%) and demonstrated the highest predictive performance (median AUC = 0.82; median sensitivity = 0.83). Calibration assessment (3/31, 10%) and external validation (6/31, 19%) were limited. Meta-analysis demonstrated an overall predictive model accuracy of 0.83 (95% CI: 0.77–0.87). Imaging models performed well for larger cardiac structures (overall median DSC = 0.85, range: 0.76–0.94), while coronary artery segmentation remained challenging. Average TRIPOD + AI and CLAIM adherence were 79% and 71%, respectively. Most predictive (97%) and imaging (82%) studies were rated at high risk-of-bias.
Conclusion
AI shows promise for RT-associated CVT prediction and imaging but is underdeveloped for routine clinical implementation. Heterogeneity, limited validation, and methodological limitations highlight the need for standardized endpoints, external validation, and prospective clinical evaluation.