ORAKLE: Optimal Risk prediction for mAke30 in patients with sepsis associated AKI using deep LEarning.

Area of Science:

• Critical Care Medicine
• Nephrology
• Artificial Intelligence in Healthcare

Background:

• Major Adverse Kidney Events within 30 days (MAKE30) is a crucial patient-centered outcome for acute kidney injury (AKI).
• Existing MAKE30 prediction models are static and do not account for dynamic clinical changes.
• There is a need for advanced models that can adapt to evolving patient data.

Purpose of the Study:

• To introduce ORAKLE, a novel deep-learning model for predicting MAKE30.
• To utilize evolving time-series data for more accurate MAKE30 prediction.
• To enable personalized, patient-centered approaches to AKI management.

Main Methods:

• Retrospective study using MIMIC-IV, SiCdb, and eICU-CRD databases.
• Development of ORAKLE using the Dynamic DeepHit framework for time-series survival analysis.
• Comparison of ORAKLE against Cox and XGBoost models, with calibration assessed by Brier score.

Main Results:

• ORAKLE demonstrated superior performance in predicting MAKE30 across all cohorts, with AUROCs ranging from 0.83 to 0.85.
• ORAKLE outperformed XGBoost and Cox models in both AUROC and AUPRC.
• Good model calibration was observed with a Brier score of 0.21.

Conclusions:

• ORAKLE is a robust deep-learning model for predicting MAKE30 in critically ill patients with AKI.
• The model's use of dynamic time-series data captures evolving patient trajectories and treatment effects.
• ORAKLE facilitates tailored risk assessments and personalized management strategies for AKI.