Machine learning analysis of contrast-enhanced ultrasound (CEUS) for the diagnosis of acute graft dysfunction in kidney transplant recipients

Abstract

Aim: The aim of the study was to develop machine learning algorithms (MLA) for diagnosing acute graft dysfunction (AGD) in kidney transplant recipients based on contrast-enhanced ultrasound (CEUS) analysis of the graft.

Materials and methods: This prospective study involved 71 patients with kidney transplant undergoing CEUS during follow-up. AGD was
defined as an increase in serum creatinine levels of at least 25% compared to the baseline of the last three months. The control group consisted of patients with stable kidney graft function (SGF). The top five CEUS parameters that achieved the best discrimination between the AGD and SGF groups were selected based on ANOVA testing and then employed as input for training MLA (naïve Bayes (NB), k-nearest neighbors (k-NN), and logistic regression (LR)). The models were validated by leave-one-out cross-validation.

Results: Among the 111 CEUS analyses, 21 corresponded to the AGD group and 90 to the SGF group. CEUS analyses yielded 44 parameters, from which five were selected: the wash out rate in segmental arteries,
time to peak in segmental arteries, medullary mean transit time, renal mean transit time, and medullary time to fall. These five parameters were employed as input for MLA, yielding an AUROC of 0.68 for NB and k-NN and 0.72 for LR. The inclusion of graft survival in the MLA significantly improved discrimination accuracy, yielding an AUROC of 0.79 for NB, 0.76 for k-NN,
and 0.81 for LR.

Conclusions: The use of MLA represents a promising strategy for analyzing CEUS-derived parameters in the setting AGD.