Artificial Intelligence Assisted Early Warning System for Acute Kidney Injury Driven by Multi-Center ICU Database

BackgroundTo improve the performance of early acute kidney injury (AKI) prediction in intensive care unit (ICU), we developed and externally validated machine learning algorithms in two large ICU databases. MethodsUsing eICU(R) Collaborative Research Database (eICU) and MIMIC-III databases, we selected all adult patients (age [≥] 18). The detection of AKI was based on both the oliguric and serum creatinine criteria of the KDIGO (Kidney Disease Improving Global Outcomes). We developed an early warning system for forecasting the onset of AKI within the first week of ICU stay, by using 6- or 12-hours as the data extraction window and make a prediction within a 1-hour window after a gap window of 6- or 12-hours. We used 52 features which are routinely available ICU data as predictors. eICU was used for model development, and MIMIC-III was used for externally validation. We applied and experimented on eight machine learning algorithms for the prediction task. Results3,816 unique admissions in multi-center eICU database were selected for model development, and 5,975 unique admissions in single-center MIMIC-III database were selected for external validation. The incidence of AKI within the first week of ICU stay in eICU and MIMIC-III cohorts was 52.1% (n=1,988) and 31.3% (n=1,870), respectively. In eICU cohort, the performance of AKI prediction is better with shorter extraction window and gap window. We found that the AdaBoost algorithm yielded the highest AUC (0.8859) on the model with 6-hours data extraction window and 6-hours gap window (model 6-6) rather than other prediction models. In MIMIC-III cohort, AdaBoost also performed well. ConclusionsWe developed the machine learning-based early AKI prediction model, which considered clinical important features and has been validated in two datasets.