Photo Credit: iStock.com/EvgeniyShkolenko

The following is a summary of “ORAKLE: Optimal Risk prediction for mAke30 in patients with sepsis associated AKI using deep LEarning,” published in the May 2025 issue of Critical Care by Oh et al. 

Major Adverse Kidney Events within 30 days (MAKE30) were recognized as a key patient-centered outcome for evaluating acute kidney injury (AKI), while existing prediction models remained static and failed to capture dynamic clinical changes.  

Researchers conducted a retrospective study to develop and estimate Optimal risk prediction for mAke30 in patients with acute Kidney injury using deep LEarning (ORAKLE), a deep-learning model using evolving time-series data to predict MAKE30.  

They used 3 critical care databases: Medical information mart for intensive care (MIMIC-IV) for development and Salzburg intensive care database (SiCdb) and EICU collaborative research database (eICU-CRD) for external validation. Patients meeting Third international consensus definitions for sepsis and septic shock (sepsis-3) criteria who developed AKI within 48 hours of intensive care unit (ICU) admission were selected. The primary outcome was MAKE30, defined as death, new dialysis, or persistent kidney dysfunction within 30 days of admission and ORAKLE was developed using the Dynamic DeepHit framework for time-series survival analysis. Its performance was compared to Cox and XGBoost models. Model calibration was evaluated using the Brier score.  

The results showed that data were analyzed from 16,671 individuals in MIMIC-IV, 2,665 in SICdb, and 11,447 in eICU-CRD and ORAKLE demonstrated stronger predictive accuracy for MAKE30 compared to the XGBoost and Cox models. In the MIMIC-IV test set, the area under the receiver operating characteristic curve (AUROC) was 0.84 (95% CI: 0.83–0.86) for ORAKLE, 0.81 (95% CI: 0.79–0.83) for XGBoost, and 0.80 (95% CI: 0.78–0.82) for Cox. In SICdb, AUROCs were 0.83 (95% CI: 0.81–0.85), 0.80 (95% CI: 0.78–0.83), and 0.79 (95% CI: 0.77–0.81), respectively. For eICU-CRD, ORAKLE reached 0.85 (95% CI: 0.84–0.85), XGBoost 0.83 (95% CI: 0.83–0.84), and Cox 0.81 (95% CI: 0.80–0.82). Precision-recall performance was also higher for ORAKLE. The Brier score for ORAKLE was 0.21 across all datasets, indicating favorable calibration.  

Investigators concluded that ORAKLE effectively leveraged evolving time-series data to improve the prediction of MAKE 30 in AKI, facilitating more personalized and dynamic risk assessment. 

Source: ccforum.biomedcentral.com/articles/10.1186/s13054-025-05457-w