Predictive approach for liberation from acute dialysis in ICU patients using interpretable machine learning

Area of Science:

• Nephrology
• Critical Care Medicine
• Medical Informatics

Background:

• Acute kidney injury requiring dialysis (AKI-D) is a significant clinical challenge in intensive care units.
• Renal recovery after AKI-D is a vital outcome but remains understudied.
• Predicting dialysis liberation is essential for effective patient management and resource allocation.

Purpose of the Study:

• To develop and validate machine learning models for predicting renal recovery and dialysis liberation in AKI-D patients.
• To identify early predictors of renal recovery within the first three days of dialysis initiation.
• To explore the potential of these models in aiding clinical decision-making in the ICU.

Main Methods:

• Retrospective cohort study of 1,381 patients with AKI-D from 2015-2020 in a Taiwanese medical center.
• Utilized 90 routinely collected variables within the first three days of dialysis initiation.
• Developed and temporally tested prediction models using machine learning algorithms (e.g., XGBoost).

Main Results:

• 27.3% of patients experienced renal recovery.
• The XGBoost model achieved an AUC of 0.85 and an AUPRC of 0.69.
• Key predictors included urine volume, Charlson comorbidity index, vital sign trends (respiratory rate, SpO2), and lactate levels.

Conclusions:

• Successfully developed early prediction models for renal recovery in AKI-D patients.
• Models integrate early vital sign changes and fluid balance data for improved accuracy.
• These predictive tools show potential to assist clinical decision-making for dialysis liberation in the ICU.