Multi-modal AI: Integrating electrocardiography with molecular biomarkers for noninvasive detection of cardiac allograft rejection

Noninvasive rejection biomarkers (AlloMap, AlloSure) demonstrate high negative predictive value but positive predictive values of only 10–30%, driving avoidable endomyocardial biopsies. We developed a multi-modal AI model integrating electrocardiography with molecular biomarkers in 389 heart transplant recipients (5,338 ECG–biopsy pairs), performing the first head-to-head comparison of AI-ECG, AlloMap, AlloSure, and their integration against biopsy-proven rejection. Individual modalities showed modest discrimination (AI-ECG AUC 0.65; AlloSure 0.67; AlloMap 0.56), while combining biomarkers alone did not improve performance (AUC 0.65). In the multi-modal test subset of ECGs (n=38, 2 rejection events), the integrated model achieved AUC 0.81 with 100% specificity (0/36 false positives) and 97% negative predictive value, eliminating all 19 false-positive combined-biomarker referrals. These hypothesis-generating results support prospective multi-center validation to determine whether requiring concordance across electrical and biomarker signals reduces unnecessary biopsies while preserving high negative predictive value. Background Noninvasive rejection biomarkers (AlloMap gene-expression profiling (GEP); AlloSure donor-derived cell-free DNA (dd-cfDNA)) deliver high negative predictive value but positive predictive values of only 10–30%, driving avoidable endomyocardial biopsies. Each biomarker interrogates a single biological compartment. Whether integrating AI-enabled electrocardiography (AI-ECG) with these molecular assays improves specificity has not been tested. Methods We retrospectively assembled 5,338 unique ECG–endomyocardial-biopsy pairs from 389 adult heart transplant recipients at a single center (2018–2024). A ResNet-18 convolutional neural network was trained on 8-lead ECG waveforms with clinical covariates. In the 340-patient subset with same encounter AlloMap and AlloSure testing (405 ECG–biomarker–biopsy triplets), an integrated model fused the ECG embedding with biomarker values via late concatenation. We performed the first head-to-head, same-cohort comparison of AI-ECG, AlloMap, AlloSure, and their integration against biopsy-proven rejection (ISHLT ≥2R ACR or pAMR ≥2). Discrimination was evaluated with bootstrap 95% confidence intervals. Results Individual modalities showed modest discrimination (AI-ECG AUC 0.65, 95% CI 0.58–0.72; AlloSure 0.67, 0.56–0.78; AlloMap 0.56, 0.46–0.67); combining biomarkers alone did not improve performance (AUC 0.65). In the multi-modal test subset (n=38 ECG-biopsy pairs, 2 rejection events), the integrated model achieved AUC 0.81 (0.50–1.00), 100% specificity (36/36; Clopper–Pearson 90–100%), and 97% NPV. At the same NPV threshold, all 19 false-positive combined-biomarker referrals in this subset would have been avoided, though this estimate reflects a single operating point in a small sample. AUC gains over single modalities did not reach statistical significance (DeLong p=0.31–0.34) given the limited event count. Conclusions A multi-modal AI model that requires concordance across electrical, immune-transcriptional, and donor-DNA signals showed numerically higher specificity for cardiac allograft rejection than any single modality, although the corresponding AUC differences did not reach statistical significance. These hypothesis-generating findings suggest that cross-compartment concordance may help discriminate true rejection from biomarker noise, but the small multi-modal test set precludes definitive conclusions; a pre-specified prospective multi-center trial powered for ≥50 rejection events is the necessary next step to test this hypothesis.