Abstract: SA-OR017
A Multicenter Artificial Intelligence Prediction Model Identifies Distinct Preoperative Risk Profiles for Postoperative Acute Kidney Disease
Session Information
- AKI Trials, Therapeutics, and Risk Stratification
October 24, 2026 | Location: Room 603, Convention Center
Abstract Time: 05:40 PM - 05:50 PM
Category: Acute Kidney Injury
- 102 AKI: Clinical, Outcomes, and Trials
Authors
- Kwon, Soie, Chung-Ang University, Dongjak-gu, Seoul, Korea (the Republic of)
- Lee, Hajeong, Seoul National University Hospital, Jongno-gu, Seoul, Korea (the Republic of)
Background
While numerous studies have developed prediction models for acute kidney injury (AKI), research on acute kidney disease (AKD) prediction remains scarce despite growing clinical interest in AKD. We developed an artificial intelligence (AI)-based prediction model for postoperative AKD.
Methods
We collected data from 239,642 patients who underwent non-cardiac major surgery lasting more than one hour at five hospitals. AKD was defined according to the KDIGO 2021 as either the occurrence of AKI, or an increase in serum creatinine >50%, a decrease in GFR ≥35%, or an eGFR <60 mL/min/1.73 m2 within 7–90 days postoperatively. A total of 66 preoperative features were used as input variables. Seoul National University Hospital served as the development cohort, with 30% of patients randomly assigned to the internal validation cohort. External validation was performed using data from the remaining four hospitals. Logistic regression, random forest, XGBoost, LightGBM, and CatBoost were employed as prediction models. Model performance was comprehensively evaluated in terms of discrimination, calibration, and clinical utility.
Results
In the overall cohort, AKI occurred in 6.2% and AKD occurred in 11.1% of patients. In internal validation, LightGBM and CatBoost outperformed other models in terms of discrimination. LightGBM was selected as the best-performing model, given its superior clinical utility demonstrated by the highest number of true positives and the fewest false negatives among all models. LightGBM also showed acceptable model performance in external validation. To examine whether AKD with and without AKI have distinct characteristics, we analyzed SHAP values for each outcome: operation duration was the most influential feature for AKI development, whereas baseline eGFR had the greatest impact on AKD without AKI.
Conclusion
Using a LightGBM model, we developed and extensively externally validated a prediction model for postoperative AKD, and demonstrated that the risk factors for AKD with versus without AKI differ.