Personalized Prediction of Long-Term Renal Function Prognosis Following Nephrectomy Using Interpretable Machine...

Area of Science:

• Nephrology
• Artificial Intelligence
• Medical Informatics

Background:

• Acute kidney injury (AKI) is a common complication after nephrectomy, with potential progression to acute kidney disease (AKD) and chronic kidney disease (CKD).
• Predictive mechanisms for AKI to AKD/CKD transitions are not fully understood.
• Interpretable machine learning (ML) offers insights into clinical features influencing long-term renal outcomes post-nephrectomy.

Purpose of the Study:

• Evaluate postnephrectomy AKI, AKD, and CKD rates and analyze long-term renal outcome trajectories.
• Interpret AKD and CKD predictive models using explainability algorithms (SHAP, LIME).
• Develop a web-based tool for estimating AKD or CKD risk post-nephrectomy.

Main Methods:

• Retrospective cohort study of 1559 patients undergoing nephrectomy (July 2012 - June 2019).
• Eight ML algorithms were used to build predictive models for AKD and CKD, with data split into training, validation, and test sets.
• Explainability plots (SHAP, LIME) and directed acyclic graphs were used for model interpretation; a web-based prediction tool was developed.

Main Results:

• Incidence rates: AKI 21.7%, AKD 15.3%, CKD 10.6%.
• Light Gradient-Boosting Machine (LightGBM) model showed high performance (AUC 0.97 for AKD, 0.96 for CKD).
• Top AKD predictors: operative duration, hemoglobin, blood loss, urine protein, hematocrit. Top CKD predictors: baseline eGFR, pathology, renal function trajectories, age, total bilirubin.

Conclusions:

• Interpretable ML models effectively identify patients at risk of AKD and CKD post-nephrectomy by highlighting critical features.
• A web-based calculator utilizing the LightGBM model can support personalized, evidence-based clinical strategies for renal function management.