Predicting CKD Progression Using Machine Learning

Background: Chronic Kidney Disease (CKD) increases morbidity and mortality. With multiple new interventions available, there will be a need for optimal combination therapies to slow or reverse the progression of CKD. We present a first step in building a computational framework for achieving this goal. We develop an Artificial Neural Network (ANN) model to predict CKD trajectory based on current patient status and selected therapeutic interventions. Methods: Using data from Chronic Renal Insufficiency Cohort (CRIC) we predicted GFR (CKD-EPI) trajectory, based on age, BMI, smoking status, hypertension, proteinuria, diabetes, and the prescription of antidiabetic medication, beta blockers, ACE inhibitors, and ARB’s. We used a transformer ANN combining Multi-Head Self-Attention with Multi-Layer Perceptron. Training was performed using 10-fold Cross-Validation. Predictive performance was assessed using Root Mean Square Error (RMSE) and R2. Results: The mean Cross-Validation RMSE achieved was 6.98 +- 0.37 and the R2 was 0.83 +- 0.02 (Table 1). Figure 1 shows an example of GFR trajectory prediction for an individual subject (left) and the regression plot for one of the Cross-Validation folds. Conclusion: Modern Machine Learning techniques facilitate sequential prediction tasks, such as modeling CKD progression. A transformer ANN model predicts individual CKD trajectories with high accuracy. The model will be used to discover optimal treatment combinations. As more data become available about the use of new classes of therapeutic agents for CKD, those data can be incorporated into the model to enhance its utility. Acknowledgment: CRIC data were provided by NIDDK Central Repository, a program of the National Institute of Diabetes and Digestive and Kidney DiseasesFigure 1: GFR trajectory prediction for an individual subject (left) and the regression plot on data from a single validation fold (right). Prediction quality metrics - CV Fold 1 2 3 4 5 6 7 8 9 10 RMSE 6.8 7.1 6.5 7.4 6.9 7.5 7.1 7.2 6.4 6.7 R2 0.83 0.82 0.86 0.80 0.83 0.86 0.84 0.80 0.84 0.82