A neural network approach to glomerular filtration rate estimation: a single-centre retrospective audit

The 2009 Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation without race correction factor is frequently used for an estimate of glomerular filtration rate (eGFR) and to support a single-sample GFR regime. This study examines whether neural networks offer a potential means to improve the accuracy of GFR estimates using the same initial inputs as eGFR.

An audit of 865 adult GFR examinations and serum creatinine measurements between January 2010 and 2024 was undertaken. Patient sex, age, creatinine, and measured GFR were used to train a neural network (NN) model with an 80 : 20 train-test split, with test set root mean square error (RMSE), accuracy, median bias, and sensitivity calculated and compared against the 2009 CKD-EPI equation eGFR.

NN GFR showed an improved performance against the 2009 CKD-EPI equation in RMSE: 12.0 vs. 16.6 mL/min/1.73 m 2 ( P  < 0.001), median bias: -2.50 vs. 7.86 mL/min/1.73 m 2 ( P  < 0.001) and accuracy: 94.2 vs. 83.2% ( P  < 0.001). Both NN GFR and the eGFR equation had poor sensitivity across the British Nuclear Medicine Society single-sample ranges of 25-50, 50-70, 70-100, and >100 mL/min/1.73 m 2 : 57.9 vs. 57.9%, 50.0 vs. 26.9%, 84.4 vs. 54.2%, 10.0 vs. 70.0%.

This study has suggested that locally trained NNs can offer a potential avenue to improve GFR predictions, even on small and diverse datasets.

Although the model is not sufficiently sensitive to predict the optimum time-sample point for a single-sample regime, this work can serve as a proof of concept for UK-specific NN GFR models.