Abstract: PO2339
Usefulness of Machine-Learning-Predicted Probability as a New Risk Index for Prediction of Renal and Life Prognoses of CKD
Session Information
- Reassessing Race in Predicting Progression
November 04, 2021 | Location: On-Demand, Virtual Only
Abstract Time: 10:00 AM - 12:00 PM
Category: CKD (Non-Dialysis)
- 2101 CKD (Non-Dialysis): Epidemiology, Risk Factors, and Prevention
Authors
- Kanda, Eiichiro, Kawasaki Ika Daigaku, Kurashiki, Okayama, Japan
- Epureanu, Bogdan I., University of Michigan, Ann Arbor, Michigan, United States
- Adachi, Taiji, Kyoto Daigaku, Kyoto, Japan
- Sasaki, Tamaki, Kawasaki Ika Daigaku, Kurashiki, Okayama, Japan
- Kashihara, Naoki, Kawasaki Ika Daigaku, Kurashiki, Okayama, Japan
Background
Personalized and accurate prediction is useful for chronic kidney disease (CKD) therapy. Predialysis death is a competitive risk of dialysis in CKD patients and lowers the accuracies of the prediction of their renal and life prognoses. Thus, we determined whether machine-learning-predicted probability works as an index for the risks of predialysis death and dialysis in CKD patients and attempted its application.
Methods
We constructed a database of electronic-medical-record data of CKD patients in Japan, and developed risk prediction machine-learning models using random forest (RF), Gradient Boosting Decision Tree, and eXtreme Gradient Boosting for the prediction of dialysis and death over 1 year. The performances of the probabilities estimated using the models were compared by the bootstrap method with those of clinical indices in a prospective cohort study of CKD patients (n=67,957).
Results
Sixteen models were developed and showed statistically significantly higher C-statistics than clinical indices. Two RF models including 22 or 8 variables showed high C-statistics: 0.932 (95% CI 0.916, 0.948) and 0.93 (0.915, 0.945), respectively, which were higher than estimated glomerular filtration rate and urinary protein levels (p<0.0001). Cox proportional hazards models with the spline term showed the relationship between the high probabilities and the high outcome risks (p<0.0001). We also developed a Web-based risk prediction system using those two models.
Conclusion
This study showed that the machine-learning-based probability is useful as a new risk index for dialysis and death and applicable to clinical practice.