Extracting Surrogate Decision Trees from Black-Box Models to Explain the Temporal Importance of Clinical Features in Predicting Kidney Graft Survival

Jaber Rad, Karthik K. Tennankore, Amanda Vinson, Syed Sibte Raza Abidi

Abstract: Prognostic modelling using machine learning techniques has been used to predict the risk of kidney graft failure after transplantation. Despite the clinically suitable prediction performance of the models, their decision logic cannot be interpreted by physicians, hindering clinical adoption. eXplainable Artificial Intelligence (XAI) is an emerging research discipline to investigate methods for explaining machine learning models which are regarded as ‘black-box’ models. In this paper, we present a novel XAI approach to study the influence of time on information gain of donor and recipient factors in kidney graft survival prediction. We trained the most accurate models regardless of their transparency level on subsequent non-overlapping temporal cohorts and extracted faithful decision trees from the models as global surrogate explanations. Comparative exploration of the decision trees reveals insightful information about how the information gain of the input features changes over time.