Deep learning algorithms for predicting renal replacement therapy initiation in CKD patients: a retrospective cohort...

Area of Science:

• Nephrology
• Artificial Intelligence in Medicine
• Predictive Analytics

Background:

• Accurate prediction of renal replacement therapy (RRT) initiation is crucial for managing chronic kidney disease (CKD).
• Current prediction methods may not fully leverage comprehensive patient data.

Purpose of the Study:

• To develop and evaluate deep learning algorithms (DLAs) for predicting RRT risk in CKD patients.
• To compare the performance of DLAs against the established Kidney Failure Risk Equation (KFRE).

Main Methods:

• A multi-centre retrospective cohort study involving CKD patients with eGFR < 30 ml/min/1.73m².
• Development and training of various deep learning algorithm structures using patient medical history, prescriptions, and biochemical data.
• Comparative analysis of DLA predictive performance against KFRE using a dedicated test set.

Main Results:

• DLAs demonstrated superior predictive accuracy for RRT initiation risk compared to KFRE (e.g., CNN ROC-AUC = 0.91 vs. KFRE ROC-AUC = 0.84).
• DLAs successfully predicted outcomes like renal transplants and dialysis initiation up to 5 years post-assessment.
• The algorithms effectively captured complex, non-linear relationships within patient data.

Conclusions:

• Deep learning algorithms offer a more accurate approach to predicting RRT risk in CKD patients than traditional methods.
• Integrating diverse patient data, including medical history and prescriptions, significantly improves prediction performance.
• DLAs show promise for enhancing CKD patient care and resource allocation, though further validation is needed.