An interpretable multimodal fusion model for improving diagnosis and prognosis prediction of kidney allograft rejection

Background: Kidney allograft rejection exhibits highly complex features due to the intricate immune mechanisms involved.The early and accurate diagnosis of rejection remains challenging, particularly when differentiating between/among subtypes.Herein, the authors present "RenalTransPredNet," an interpretable model that integrates whole-slide images with routine clinicopathological variables to diagnose kidney allograft rejection.Methods: This study included kidney allograft biopsies performed at the Headquarters and Lingnan Hospital of the Third Affiliated Hospital of Sun Yat-sen University between January 2015 and August 2024.After screening, 1086 digital whole-slide images from 362 biopsies and their corresponding clinicopathological information were analyzed, comprising 209 other lesions and 153 rejections.By integrating multiple instance learning-based pathological image analysis with random forest-based clinicopathological information analysis, RenalTransPredNet was used to efficiently construct models for kidney allograft rejection-related multiclass and binary classification tasks.Results: RenalTransPredNet achieved a higher area under the curve (AUC) value than both the pathological imaging and clinicopathological information models in detecting and subtyping kidney allograft rejection, yielding an AUC of 0.886 using data from the internal testing set.For the prediction of treatment response to rejection, RenalTransPredNet yielded an AUC of 0.743.In the task predicting graft loss after rejection, RenalTransPredNet demonstrated predictive capabilities for graft loss at 1, 2, 3, and 5 years post-rejection, with AUCs of 0.876, 0.942, 0.904, and 0.978, respectively.Conclusion: These findings support the potential of RenalTransPredNet in improving the diagnostic accuracy of pathologists in detecting and subtyping kidney allograft rejection and in contributing to prognostic assessments.