Under Pressure: Using a Machine-Learning Program to Stratify Risk Factors Associated with Incident CKD

Background: The initial stages of chronic kidney disease (CKD) are often asymptomatic, but targeted screening can help prevent early disease progression and improve outcomes. Neural networks are a machine learning program capable of combing through large databases to attribute statistical importance to multiple risk factors associated with CKD. Methods: Using data from the CDC’s Kidney Disease Surveillance System and Atlas of Heart Disease and Stroke, 10-year trends in several risk factors were calculated from 2009 to 2019 on a county-wide basis across the United States. The quintile-based trend for each was then input to a neural network tasked with predicting an increase or decrease in the prevalence of CKD over the 10-year window. Statistical importance was calculated to determine which risk factors were most important to control incident CKD. Results: 3,021 counties in the United States were included. 50 iterations of the optimized neural network were generated with each model utilizing a random 30% of input data for training and 70% for performance challenge. Mean area under the curve during these performance challenges was 0.954 (95% CI: 0.946 - 0.963) (Figure 1). Accuracy was similar in predictions of increased or decreased CKD prevalence. Hypertension trends were the most statistically important risk factor in correctly adjudicated forecasting (Figure 2). Conclusion: These findings underscore the importance of early detection and timely management of hypertension in preventing the onset and progression of CKD.