Abstract: TH-OR08
Machine Learning for Development of a Real Time AKI Risk Prediction Model in ICU With External Validation and Federated Learning at Five Medical Centers: From Model Development to Clinical Application
Session Information
- AKI Research: Prediction, Diagnosis, Recovery
November 03, 2022 | Location: W230, Orange County Convention Center‚ West Building
Abstract Time: 05:33 PM - 05:42 PM
Category: Acute Kidney Injury
- 102 AKI: Clinical‚ Outcomes‚ and Trials
Author
- Huang, Chun-Te, Taichung Veterans General Hospital, Taichung, Taiwan
Background
Early prediction of AKI ahead of 24 to 48 hours window has now been proposed as a possible strategy to prevent or alleviate AKI by providing clinicians sufficient time for timely intervention. We aim to develop a parsimonious AKI prediction model and build a platform using federated learning for easy data transport and integration of the model between different hospitals.
Methods
All adult ICU admissions from 2015-2020 in Taichung Veterans General Hospital (TCVGH) were used as the derivation cohort. Adult ICU admissions from 2018-2020 in four other medical centers were used as external validation cohort. AKI labelling is based on 2012 KDIGO AKI definition. A parsimonious prediction model was selected by LASSO and deployed to the 2018-2020 external validation cohort at 4 other medical centers. A federated learning platform is built retrain and tune the AKI prediction model.
Results
A total of 16785 adult ICU admissions were included in the derivation cohort with median age of 68, 62% male, and 30.8% AKI incidence. Total of 60 predictors were included in the model. The machine learning algorithm for AKI prediction revealed XGBoost with better performance metrics in terms of sensitivity:0.795, specificity:0.866, AUROC:0.911, precision:0.726, and accuracy:0.844, and decision curve analysis. After applying LASSO for parsimonious prediction model for external validation at these 4 hospitals showed the prediction performance with AUROC ranging from 0.785 to 0.864. Federated learning using neuron network algorithm improved the prediction performance.
Conclusion
AKI is common in ICU with poor prognosis. Early prediction ahead of 24-48 hours may help clinicians for timely intervention to prevent it from happening. We demonstrate the concept of artificial intelligence as medical assistance to clinicians in a feasible way.
Funding
- Government Support – Non-U.S.