Abstract: FR-PO0003
Temporal Trends in Predicted CKD Risk and Clinical Implications After Disease Onset: Machine Learning-Based Investigation
Session Information
- Artificial Intelligence and Digital Health at the Bedside
November 07, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Yoon, Soo-Young, Kyung Hee University Hospital, Dongdaemun-gu, Seoul, Korea (the Republic of)
- Lee, Jeong-Yeun, Kyung Hee University Hospital, Dongdaemun-gu, Seoul, Korea (the Republic of)
- Moon, Youngyoon, Kyung Hee University Hospital at Gangdong, Gangdong-gu, Seoul, Korea (the Republic of)
- Joo, Yoosun, Kyung Hee University Hospital at Gangdong, Gangdong-gu, Seoul, Korea (the Republic of)
- Lee, Sangho, Kyung Hee University Hospital at Gangdong, Gangdong-gu, Seoul, Korea (the Republic of)
- Kim, Jin Sug, Kyung Hee University Hospital, Dongdaemun-gu, Seoul, Korea (the Republic of)
- Jeong, Kyunghwan, Kyung Hee University Hospital, Dongdaemun-gu, Seoul, Korea (the Republic of)
- Hwang, Hyeon Seok, Kyung Hee University Hospital, Dongdaemun-gu, Seoul, Korea (the Republic of)
Background
Chronic kidney disease (CKD) is a progressive condition where risk accumulates before clinically apparent onset, potentially accelerating disease onset and progression. However, strategies for evaluating risk of CKD occurrence remains highly limited, and the trajectories of CKD risk over time are not elucidated.
Methods
We enrolled 51,156 individuals with 128,395 visits from the University of California Health Data Warehouse between 2012 and 2024. Debiased XGBoost with binary cross-entropy (D-XGBoost) incorporated demographic, comorbidity, and laboratory data to predict the risk of CKD occurrence. Latent class growth mixture modeling identified risk trajectories, and kidney and cardiovascular outcomes after CKD onset were evaluated in relation to CKD risk patterns.
Results
During a mean of 2.95 ± 2.87 visits with a mean visit interval of 1.22 ± 1.27 years, 5,530 patients were identified with CKD. The D-XGBoost model achieved time-dependent areas under the receiver-operating characteristic curve of 0.95 (95% confidence interval [CI], 0.94 – 0.96), C-index of 0.93 (0.92 – 0.94), and Brier score of 0.027 (0.026 – 0.028) for predicting 5-year CKD occurrence. Similar predictive performance was also observed across 1 to 4 years of CKD occurrence. When the D-XGBoost model calculated 5-year CKD risk at all visits, three distinct trajectories of CKD risk were identified: gradual (75.9%), progressive (10.1%), and rapid increase (14.0%). Patients with a progressive increase in CKD risk had an adjusted hazard ratio (HR) of 1.31 (95% CI, 1.06–1.61) for an eGFR decline <25% and HR of 1.59 (1.09–2.31) for a decline <50% from baseline after CKD onset. Patients with a rapid increase were associated with HR of 1.84 (1.17–2.91) for cardiovascular events and HR of 2.36 (1.49–3.73) for all-cause mortality.
Conclusion
Our D-XGBoost model accurately predicted risk of CKD occurrence and categorized patients into three distinct risk trajectories. These trajectories were associated with differing post-CKD outcomes, emphasizing the importance of dynamic risk monitoring for earlier, targeted interventions.