Prediction of post-donation renal function using machine learning techniques and conventional regression models in living kidney donors

Abstract

Background: Accurate prediction of renal function following kidney donation and careful selection of living donors are essential for living-kidney donation programs. We aimed to develop a prediction model for post-donation renal function following living kidney donation using machine learning.

Methods: This retrospective cohort study was conducted with 823 living kidney donors between 2009 and 2020. The dataset was randomly split into training (80%) and test sets (20%). The main outcome was the post-donation estimated glomerular filtration rate (eGFR) 12 months after nephrectomy. We compared the performance of machine learning techniques, traditional regression models, and models from previous studies. The best-performing model was selected based on the mean absolute error (MAE) and root mean square error (RMSE).

Results: The mean age of the participants was 45.2 ± 12.3 years, and 48.4% were males. The mean pre-donation and post-donation eGFRs were 101.3 ± 13.0 and 68.8 ± 12.7 mL/min/1.73 m2, respectively. The XGBoost model with the eGFR, age, serum creatinine, 24-h urine creatinine, 24-h urine sodium, creatinine clearance, cystatin C, cystatin C-based eGFR, computed tomography volume of the remaining kidney/body weight, normalized GFR of the remaining kidney measured through a diethylenetriaminepentaacetic acid scan, and sex, showed the best performance with a mean absolute error of 6.23 and root mean square error of 8.06. An easy-to-use web application titled Kidney Donation with Nephrologic Intelligence (KDNI) was developed.

Conclusions: The prediction model using XGBoost accurately predicted the post-donation eGFR after living kidney donation. This model can be applied in clinical practice using KDNI, the developed web application.