Abstract: TH-PO032
Development of Artificial Intelligence Model for the Prediction of AKI, Acute Kidney Disease, and CKD After General Anesthesia Surgery
Session Information
- AI, Digital Health, Data Science - I
November 02, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Eum, Sang Hun, Incheon St. Mary's Hospital, Incheon, Korea (the Republic of)
- Min, Ji Won, Bucheon St. Mary's Hospital, Bucheon, Korea (the Republic of)
- Koh, Eun Sil, Yeouido St. Mary's Hospital, Seoul, Korea (the Republic of)
- Ban, Tae Hyun, Eunpyeong St. Mary's Hospital, Seoul, Korea (the Republic of)
- Hong, Suyeon, Uijeongbu St. Mary's Hospital, Uijeongbu, Korea (the Republic of)
- Chung, Byung ha, Seoul St. Mary's Hospital, Seoul, Korea (the Republic of)
- Kim, Yong Kyun, St. Vincent's Hospital, Suwon, Korea (the Republic of)
- Shin, Seok Joon, Incheon St. Mary's Hospital, Incheon, Korea (the Republic of)
- Yoon, Hye Eun, Incheon St. Mary's Hospital, Incheon, Korea (the Republic of)
Background
After the occurrence of AKI, ongoing renal pathophysiologic process leads to acute kidney disease (AKD) or chronic kidney disease (CKD). Postoperative AKI is associated with increased mortality and morbidity in patients undergoing surgeries performed under general anesthesia. A machine learning-based prediction model for postoperative AKI, AKD, and CKD is needed.
Methods
In this retrospective cohort analysis, noncardiac surgeries performed between 2009 and 2019 at 7 university hospitals in South Korea were included. According to the creatinine-criteria of KDIGO, postoperative AKI was defined as an increase of serum creatinine at least 1.5 times the baseline value or initiation of renal replacement therapy occurring over 7 days or less after surgery; AKD was defined as persistent AKI over 7 and 90 days; CKD was defined as persistent AKD beyond 90 days. Data imbalance was adjusted using the SMOTE algorithm, and four machine learning prediction models were tested: deep neural networks, decision tree, random forest, and light gradient boosting machine (GBM). Model performance was compared using the area under the curve (AUC) of the receiver-operating characteristic, accuracy, and F1 score.
Results
Among 239,267 surgeries, 2,716 postoperative AKI (1.14%), 97 AKD (0.04%), and 1,203 CKD (0.5%) events occurred. Four machine learning methods were run on 32 independent preoperative predictors. While the model run on random forest exhibited a higher AUC (0.80) than light GBM (0.77), accuracy (0.95), weighted F1 score (0.94), micro-average F1 score (0.95), and macro-average F1 score (0.31) were highest in the model run on light GBM.
Conclusion
In our comprehensive comparison of machine learning approaches, light GBM demonstrated the best performance to predict postoperative AKI, AKD and CKD. The current model may be implemented in clinical practice to predict short- and long-term kidney outcomes after surgery.