Improved Glomerular Filtration Rate Estimation by an Artificial Neural Network

Xun Liu, Xiaohua Pei, Ningshan Li, Yunong Zhang, Xiang Zhang, Jinxia Chen, Linsheng Lv, Huijuan Ma, Xiao-Ming Wu, Weihong Zhao, Tanqi LouPLoS 2013 · DOI

Abstract: A new ANN model (the six-variable GABP network) for CKD patients was developed that could provide a simple, more accurate and reliable means for the estimation of GFR and stage of CKD than traditional equations. Further validations are needed to assess the ability of the ANN model in diverse populations.

Introduction

Accurately measuring the Glomerular Filtration Rate (GFR)—the rate at which kidneys filter waste from the blood—is fundamental to diagnosing and managing Chronic Kidney Disease (CKD). While "gold standard" methods like $^{99m}Tc-DTPA$ renal dynamic imaging provide high accuracy, they are invasive, time-consuming, and expensive. Consequently, clinical practice relies on estimation equations (eGFR) derived from serum markers.

For decades, linear regression-based formulas such as the Modification of Diet in Renal Disease (MDRD) and the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) have been the standard. However, these models often fail to account for the complex, non-linear relationships between serum creatinine, muscle mass, and actual kidney function, leading to significant misclassification of disease stages.

The Limitations of Traditional Equations

Traditional equations primarily rely on serum creatinine (SC), a marker heavily influenced by non-renal factors like diet, age, and muscle mass. Because these relationships are inherently non-linear, simple regression models frequently produce biased results, particularly in the early stages of CKD.

Bland-Altman plot showing the significant bias and spread of the traditional MDRD4 equation compared to the gold standard sGFR.

As shown in the Bland-Altman analysis of the MDRD4 equation, there is a wide dispersion of data points and a noticeable trend of bias as GFR increases. This lack of precision can lead to misclassification rates as high as 50%, potentially resulting in inappropriate clinical interventions.

Methodology: The GABP Innovation

To address these shortcomings, Liu et al. (2013) developed a three-layer Back Propagation (BP) neural network optimized by a Genetic Algorithm (GA). Neural networks are "universal approximators," capable of modeling the intricate, non-linear interactions between clinical variables that linear equations miss.

The researchers utilized a Genetic Algorithm to optimize the initial weights and thresholds of the network, preventing the model from becoming trapped in "local minima" during training—a common pitfall for standard BP networks. Using Mean Impact Value (MIV) analysis, they identified six key clinical inputs for the final model (GABP6):

  1. Serum creatinine (SC)
  2. Serum urea nitrogen (SUN)
  3. Age
  4. Weight
  5. Height
  6. Gender

The model was developed and validated using a large cohort of 1,402 Chinese CKD patients across multiple institutions, using $^{99m}Tc-DTPA$ scans as the reference standard (sGFR).

Superior Performance of the GABP-6 Model

The GABP-6 network demonstrated a significant leap in performance over traditional equations across all metrics. In external validation, the model achieved a precision of 46.7 ml/min/1.73 $m^2$, compared to a range of 71.3 to 101.7 for traditional formulas.

Bland-Altman plot for the GABP-6 neural network model, demonstrating much tighter agreement with the gold standard sGFR.

The visual evidence from the Bland-Altman plots confirms that the GABP-6 model provides much tighter agreement with the gold standard. Key findings include:

  • Accuracy: The 30% accuracy rate (estimates within 30% of the true GFR) reached 75.1% for the GABP network, whereas traditional equations hovered between 46% and 55%.
  • Staging Accuracy: Misclassification of CKD stages dropped to 32.4%, a marked improvement over the ~50% error rate seen in MDRD and CKD-EPI models.
  • Early Detection: The model was exceptionally effective at identifying Stage 1 CKD, with a 90.9% correct classification ratio, compared to roughly 40% for traditional methods.

Discussion and Conclusion

The study highlights that kidney function is a multi-dimensional system better served by machine learning than simple linear logic. By incorporating anthropometric data like height and weight, the ANN effectively filters out the "noise" associated with creatinine-based estimates.

While the results are promising, the authors note that the model was trained on a Chinese population and requires further validation in diverse ethnic groups to ensure global generalizability. Additionally, to overcome the "black box" nature of neural networks, the researchers have provided an Excel-based tool to facilitate bedside application.

In conclusion, the GABP-6 neural network offers a more precise, reliable, and clinically actionable tool for staging CKD, potentially transforming how kidney health is monitored and treated.

Summary updated

Abstract

A new ANN model (the six-variable GABP network) for CKD patients was developed that could provide a simple, more accurate and reliable means for the estimation of GFR and stage of CKD than traditional equations. Further validations are needed to assess the ability of the ANN model in diverse populations.
BibTeX
@inproceedings{liu2013improved,
  author = {Xun Liu and Xiaohua Pei and Ningshan Li and Yunong Zhang and Xiang Zhang and Jinxia Chen and Linsheng Lv and Huijuan Ma and Xiao-Ming Wu and Weihong Zhao and Tanqi Lou},
  title = {{Improved Glomerular Filtration Rate Estimation by an Artificial Neural Network}},
  booktitle = {PLoS},
  year = {2013},
  doi = {10.1371/journal.pone.0058242},
}
Figure