Deep Learning Identifies Intelligible Predictors of Poor Prognosis in Chronic Kidney Disease

Area of Science:

• Nephrology
• Artificial Intelligence
• Biomedical Informatics

Background:

• Early prediction of chronic kidney disease (CKD) progression is vital for personalized patient treatment and improved outcomes.
• Identifying key clinical and laboratory features that predict progression to end-stage renal disease (ESRD) remains a challenge.

Purpose of the Study:

• To evaluate the intelligibility and predictive performance of machine learning (ML) and deep learning (DL) models for ESRD prediction in CKD patients.
• To identify significant clinical and laboratory markers associated with CKD progression using ML and DL interpretability techniques.

Main Methods:

• Eight ML models were trained to predict ESRD within three years using patient demographics, clinical data, and comorbidities.
• LASSO, random forest, and XGBoost were employed for feature selection.
• Four advanced attribution methods were applied to a DL model to enhance its interpretability.

Main Results:

• The DL model achieved a high AUC-ROC of 0.8991, outperforming baseline ML models.
• Interpretations from DL with attribution methods, random forest, and XGBoost aligned with clinical knowledge.
• Key predictors of CKD progression included hematuria, proteinuria, potassium, and urine albumin-to-creatinine ratio (positively associated), and eGFR and urine creatinine (negatively associated).

Conclusions:

• Deep learning models, enhanced with attribution algorithms, offer intelligible insights into CKD progression.
• The study identified potentially novel markers for CKD progression, supported by data-driven evidence.
• These findings support the integration of ML in clinical decision-making for CKD management.