Artificial intelligence optimizes immune rejection prediction and management in heart transplantation: a structured narrative review

Background: Heart transplantation remains the definitive therapy for end-stage heart failure, yet long-term outcomes are limited by three core clinical bottlenecks in immune rejection management: imprecise preoperative donor-recipient matching, overreliance on invasive endomyocardial biopsy (EMB) for postoperative rejection surveillance, and high inter-observer variability in manual pathological diagnosis of rejection. Artificial intelligence (AI) has emerged as a promising tool to address these gaps, but the methodological quality and clinical translation readiness of supporting evidence have not been comprehensively synthesized. Methods: This structured narrative review synthesized original research published between October 1, 2020, and October 1, 2025, identified via a targeted PubMed search and manual reference screening. Two independent reviewers performed study selection and data extraction, with discrepancies resolved by consensus. Common methodological limitations across included studies were synthesized qualitatively. Results: A total of 42 studies were included in the final qualitative synthesis. Preoperatively, 3D-Convolutional Neural Networks (3D-CNNs) enabled automated, accurate total cardiac volume (TCV) measurement for anatomical matching, while machine learning models identified non-linear synergistic risk factors for postoperative adverse events, outperforming traditional regression models. Postoperatively, AI models integrating non-invasive biomarkers (gene expression profiles, extracellular vesicles, donor-derived cell-free DNA) showed high diagnostic accuracy for rejection, with one single-center retrospective study estimating a 56.8% reduction in unnecessary EMB procedures (prospective clinical validation is still required). For pathological diagnosis, AI models improved the sensitivity of high-grade acute cellular rejection (ACR) detection from 39.5% to 74.4% compared with manual assessment, generative adversarial networks (GANs) addressed rare rejection sample scarcity with a rejection region detection AUROC of 98.84%, and explainable AI tools aligned model decisions with pathologists' judgment. The overall methodological quality of included studies was suboptimal, with most studies limited by single-center retrospective design, small sample size, and lack of independent external validation. Conclusions: AI has demonstrated promising potential to optimize donor-recipient matching, enable non-invasive rejection surveillance, and standardize pathological diagnosis in heart transplantation. However, most current evidence comes from exploratory, single-center retrospective studies with important methodological limitations that restrict their immediate clinical translation. Future research should prioritize prospective, multi-center clinical validation, standardized biomarker and model reporting, and federated learning data ecosystems to translate AI innovations into routine clinical practice.