Personalized prediction model generated with machine learning for kidney function one year after living kidney...

Area of Science:

• Nephrology
• Medical Informatics
• Biomedical Engineering

Background:

• Living kidney donation typically results in a ~30% reduction in kidney function.
• Individual variation in post-donation kidney function necessitates personalized monitoring.
• Accurate prediction of post-donation kidney function is crucial for donor well-being.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting serum creatinine (Cre) levels one year post-kidney donation.
• To utilize preoperative clinical data, including volumetric measurements, for prediction.
• To enhance personalized care and monitoring for living kidney donors.

Main Methods:

• Developed an ML model using symbolic regression via genetic programming.
• Included 204 living kidney donors with preoperative clinical data and CT-derived volumetry.
• Validated the model using a 7:3 training-to-test data split.

Main Results:

• The ML model achieved a median absolute error of 0.079 mg/dL for predicting Cre.
• Outperformed conventional methods in a validation cohort (higher R², lower RMSE and MAE).
• Key predictive variables identified were preoperative Cre and remnant kidney volume.

Conclusions:

• The developed ML model accurately predicts post-donation kidney function.
• The model can aid in monitoring donor outcomes and enhancing personalized care.
• Deployed as a web application for practical clinical utility.