AI-integrated multi-omics fusion for precision cardiovascular diagnostics

Cardiovascular disease remains one of the leading causes of mortality worldwide, requiring accurate and early diagnostic systems for effective intervention. Traditional prediction approaches relying only on clinical data often fail to capture complex biological interactions underlying disease progression. This study proposes CardioPrecision-BioFusion, an AI-driven framework that combines clinical, genomic, transcriptomic, and simulated proteomic representations. The clinical and omics datasets were obtained from independent cohorts and were not patient matched; consequently, the framework performs cohort-level latent representation integration and does not claim patient-level multi-omics fusion. Modality-specific encoders, target-blind prototype retrieval, multi-head attention, transformer-based learning, an autoencoder, and stacking were used to model heterogeneous feature interactions, while masked attention handled structurally unavailable modalities without imputing unobserved patient-level omics measurements. On the internal held-out constructed fusion test set, the model achieved 97.96% accuracy (95% CI, 97.14–98.55%), 98.05% precision, 98.05% recall, and a 98.05% F1-score. In a source-disjoint, transcriptomics-only external evaluation using 332 samples from GSE141910, the frozen transcriptomics branch achieved 84.64% accuracy and an AUC-ROC of 91.20%. SHAP analysis identified influential clinical and molecular features. These findings support the framework as an exploratory multimodal methodology, but prospective validation in patient-matched multi-omics cohorts is required before clinical use.