Fig. 3

Model interpretability assessed using SHapley additive exPlanation. (A) The SuperLearner bee swarm plot depicts each variable’s importance for predicting diabetic kidney disease with type 2 diabetes mellitus and normal renal function (top 20). One dot per patient per feature is coloured according to an attribute value, where orange and purple represent higher and lower values, respectively. Features are sorted in decreasing order of importance, calculated as the average absolute SHAP value per feature. The abbreviations of all analytical variables are detailed in Table S1. (B) SHapley Additive explanation dependence plot of SuperLearner (selected four features), depicting how a single variable affects the prediction. SHapley Additive explanation values greater than zero for specific features suggested an increased risk of diabetic kidney disease. SHapley Additive explanation values below zero for specific features indicate a decreased risk of diabetic kidney disease. The remaining 16 from the top 20 plots are shown in Figure S1. (C) SHAP waterfall plot for patients with predicted low (left) and high (right) risk of developing DKD. SHAP value (left: -0.0789, right: 0.106). The base value at the bottom of the waterfall plot starts at zero. SHAP values shown inside yellow arrows correspond to input variables that ‘push’ the model towards predicting higher risk, whereas those in the magenta ‘push’ the model towards a lower predicted risk