Abstract
Exosomes and other extracellular vesicles (EVs) carry microRNAs, proteins, and lipids that reflect cardiovascular pathophysiology and can enable minimally invasive biomarker discovery. However, EV datasets are highly dimensional and heterogeneous, strongly influenced by pre-analytic variables and non-standardized isolation/characterization workflows, limiting reproducibility across studies. Artificial intelligence (AI), including machine learning (ML), deep learning (DL), and network-based approaches, can support EV biomarker development by integrating multi-omics profiles with clinical metadata. These approaches enable feature selection, disease subtyping, and interpretable model development. Among the AI approaches evaluated, ensemble methods (Random Forest, gradient boosting) demonstrate the most consistent performance for EV biomarker classification (AUC 0.80-0.92), while graph neural networks (GNNs) are particularly promising for path integration but require larger validation cohorts. Evolutionary neural networks applied to EV morphological features yield comparable discrimination but face interpretability challenges for clinical use. Current studies report promising discrimination performance for selected EV-derived panels in acute myocardial infarction and heart failure. However, most evidence remains exploratory, based on small cohorts (n < 50) and limited external validation. For clinical implementation, EV biomarkers need direct comparison against established standards (high-sensitivity troponin and natriuretic peptides), supported by locked-in assay plans, and validation in multicenter cohorts using MISEV-aligned protocols and transparent AI reporting practices. Through a comprehensive, integrative, and comparative analysis of AI methodologies for EV biomarker discovery, together with explicit criteria for reproducibility and translational readiness, this review establishes a practical framework to advance exosomal diagnostics from exploratory research toward clinical implementation.