Fig. 4: AI model to longitudinally predict rejection episodes.

Supervised learning was used to build a diagnostic predictive model based on artificial intelligence to detect rejection episodes exploiting EV surface antigen profiling performed after membrane-sensing peptide (MSP)-capturing (285 samples included in the analysis). a A random forest regressor (rRF) model was trained, tuned, and validated through a leave-one-patient-out strategy (see methods). The rRF model combines levels of expression of EV antigens in a biomolecular fingerprint, based on normalized fluorescence intensity after correction for median levels of each antigen in correspondence with non-rejecting episodes (G0) for each single patient, and reported as a percentage of variation. The AI model considers, at each subsequent visit, only the G0 episodes encountered during the patient’s follow-up until the time point at which the patient is evaluated, evolving, and dynamically adapting to that specific patient, and continuously re-defining the threshold of variation associated with a high probability of rejection. b, c Likelihood of rejection after stratification of patients for rRF coefficients. d ROC curve analysis; area under the curve (AUC) together with 95% confidence interval (CI) is reported for rRF coefficient discriminating ACR grade 2–3A from G0-1A/B episodes. e Median values of rRF, distribution, and likelihood of rejection of patient stratified according to rejection grade (from 0 to 3A; n = 285). f Box plot and interquartile range for rRF coefficients in patients stratified according to the time point of evaluation and rejection grade (n = 285): after surgery (first sampling after heart transplant); grade 0 (non-rejecting patients); ACR grade 1A/B; pre 1A/B (time point of evaluation before a diagnosis of ACR grade 1A/B); ACR grade 2–3A; pre 2–3A (time point of evaluation before a diagnosis of ACR grade 2–3A). g Diagnostic performance (sensitivity, specificity, accuracy, positive and negative predictive values) at rRF model validation. Source data and statistics are reported in Supplementary Data S4 and Table S13. Part of (a) was produced using the Servier Medical Art public domain (https://smart.servier.com).