Machine learning-supported interpretation of kidney graft elementary lesions in combination with clinical data

Background The Banff classification standardizes the diagnoses of kidney transplant rejection based on histological criteria. Clinical decisions are generally made after integration of the Banff diagnoses in the clinical context. However, interpretation of the biopsy cases is still heterogeneous among pathologists or clinicians. Machine Learning (ML) algorithms may be trained from expertly assessed cases to provide clinical decision support. Methods The ML technique of Extreme Gradient Boosting learned from two large training datasets from the European programs BIOMARGIN and ROCKET (n= 631 and 304), in which biopsies were read centrally and consensually interpreted by a group of experts and used as a reference for untargeted biomarker screenings. The model was then externally validated in three independent datasets (n= 3744, 589 and 360). Results In the three validation datasets, the algorithm yielded a ROC curve AUC of mean (95% CI) 0.97 (0.92-1.00), 0.97 (0.96-0.97) and 0.95 (0.93-0.97) for antibody-mediated rejection (ABMR); 0.94 (0.91-0.96), 0.94 (0.92-0.95) and 0.91 (0.88-0.95) for T cell-mediated rejection; >0.96 (0.90-1.00) in all three for interstitial fibrosis - tubular atrophy (IFTA). Finally, using the largest validation cohort, we developed an additional algorithm to discriminate active and chronic active ABMR with an accuracy of 0.95. Conclusion We built an Artificial Intelligence algorithm able to interpret histological lesions together with a few routine clinical data with very high sensitivity and specificity. This algorithm should be useful in routine or clinical trials to help pathologists and clinicians and increase biopsy interpretation homogeneity.