Comprehensive Summary
Heart transplant recipients need regular monitoring for rejection, typically through endomyocardial biopsy. A Mayo Clinic study instructed a deep-learning algorithm to analyze routine 12-lead ECGs and detect moderate-to-severe acute cellular rejection (ACR). The authors used 7,590 ECG pairs in 1,427 patients, each ECG obtained within 30 days prior to biopsy. The independent test set contained 140 patients and 758 ECG–biopsy pairs, of which 20 had ACR events. In the test set, the model had AUC of 0.84 [95% confidence interval (CI): 0.78–0.90] and a sensitivity of 95% (19/20; 95% CI: 75–100%). A prospective proof-of-concept (56 patients, 97 pairs, 2 ACR events) had AUC 0.78 and sensitivity 100% (2/2, 95% CI 16–100%). This proves potential application to detect ACR. Performance declined for ECGs > 7 days before biopsy (AUC 0.77). This study is limited by a relatively small sample size used for model derivation (n = 1427) and the prospective cohort (n = 56), limiting finer analysis of multiple subgroups and explainability of the AI model, common shortcomings of deep-learning models.
Outcomes and Implications
ACR is the most common type of transplant rejection, and AI systems to project potential immune system attacks are crucial in improving transplant success. This deep learning model is a sensitive detector that can screen patients for quicker examination or biopsy and can enable remote triage. It can assist in deciding who should receive earlier treatment, but it cannot yet replace biopsy today due to the patient dataset size. Before bedside use, the model requires larger external validation, calibration of predictive risk, and more to assess trade-offs with biopsy-based monitoring.