Machine Learning-Predicted Risk Trajectories for Incident Chronic Kidney Disease and Associations with Post-CKD Outcomes.

Chronic kidney disease (CKD) is a progressive condition where risk accumulates before clinical onset. However, strategies for evaluating CKD risk remain limited, and the trajectories of risk over time are poorly understood. We enrolled 51,156 individuals without CKD in the University of California health system between 2012 and 2024. A debiased XGBoost model was developed to predict CKD risk, and the predicted probabilities from this model at each visit were used to derive risk trajectories through latent class growth mixture modeling. The model achieved a time-dependent area under the curve of 0.95 (95% confidence interval (CI), 0.94-0.96), a C-index of 0.93 (0.92-0.94), and a Brier score of 0.027 (0.026-0.028) for predicting 5-year CKD occurrence. Similar predictive performance was observed for 1 to 4 years. Longitudinal patterns of predicted CKD risk revealed three distinct trajectories: gradual (75.9%), progressive (10.1%), and rapid increase (14.0%). Compared to the gradual trajectory, the progressive trajectory was associated with a higher risk of estimated glomerular filtration rate (eGFR) decline of ≥ 25% (adjusted hazard ratio (HR) 1.31; 95% CI, 1.06-1.61) and ≥ 50% (HR 1.59; 95% CI, 1.09-2.31) following CKD onset. The progressive and rapid trajectory was linked to an increased risk of cardiovascular events (HR 1.84; 95% CI 1.18-2.91) and all-cause mortality (HR 2.36; 95% CI 1.49-3.73). Our machine learning model demonstrated strong predictive performance for incident CKD risk and identified three distinct risk trajectories that were associated with different post-CKD outcomes, highlighting the importance of early risk monitoring.