Construct a classification decision tree model to select the optimal equation for estimating glomerular filtration rate and estimate it more accurately

Zhenliang Fan, Qiaorui Yang, Zhuohan Xu, Ke Sun, Mengfan Yang, Riping Yin, Dongxue Zhao, Junfen Fan, Hongzhen Ma, Yiwei Shen, Xia HongScientific 2022 · DOI

Abstract

Accurate assessment of renal function is critical for managing Chronic Kidney Disease (CKD). While the glomerular filtration rate (GFR) is the gold standard for evaluation, direct measurement is clinically cumbersome. Consequently, various estimation equations (eGFR) are used, though their accuracy varies across patient populations. This study developed a classification decision tree model to dynamically select the most suitable eGFR equation for individual patients. Using data from 518 CKD patients, the model selects between the BIS-2, CKD-EPI(CysC), CKD-EPI(Cr-CysC), and Ruijin equations based on clinical variables. The resulting model significantly outperformed 13 traditional equations, achieving a Mean Absolute Error (MAE) of 8.5 and a Root Mean Square Error (RMSE) of 12.2 in the test set.

The Challenge of GFR Estimation

The Glomerular Filtration Rate (GFR) represents the volume of fluid filtered from the renal glomerular capillaries into the Bowman's capsule per unit time. Direct measurement using exogenous markers like $^{99m}$Tc-DTPA (Gate’s method) is precise but invasive and expensive. Most clinicians rely on eGFR formulas based on endogenous markers like creatinine or cystatin C. However, the performance of these formulas is often inconsistent, influenced by factors such as age, body mass index (BMI), and underlying pathologies like diabetes.

Methodology: A "Meta-Model" Approach

Rather than deriving a new mathematical formula, researchers built a classification decision tree using the CART algorithm to act as a clinical "traffic controller." The model evaluates a patient's specific clinical profile to assign the most accurate existing equation.

The study analyzed 518 patients from the First Affiliated Hospital of Zhejiang Chinese Medicine University, divided into a training set (n=362) and a test set (n=156). Out of 28 initial clinical variables, the tree identified six key predictors:

  1. Diabetic Nephropathy (DN) status
  2. Body Surface Area (BSA)
  3. Body Mass Index (BMI)
  4. 24-hour urine protein quantity
  5. Age
  6. Use of RASi (Renin-Angiotensin System inhibitors)
The classification decision tree model selects the optimal eGFR equation based on clinical parameters such as Diabetic Nephropathy status, BMI, and BSA.

Performance and Results

The decision tree model demonstrated a marked improvement in accuracy over traditional "one-size-fits-all" equations. In the validation phase, the model's predictions were compared against 13 standard formulas, including MDRD and various CKD-EPI iterations.

The model achieved the lowest error rates (MAE 8.5, RMSE 12.2). For comparison, the widely used MDRD equation showed significantly higher deviation. Bland–Altman analysis confirmed that the decision tree approach had the smallest degree of variation and the most centralized bias relative to the measured GFR (sGFR).

Comparison of eGFR-sGFR across different equations. The Decision Tree Classifier (far right) shows the most stable performance with the narrowest distribution around the zero-error line.

Clinical Implications

The primary advantage of this model is its interpretability. Unlike "black-box" neural networks, the decision tree allows clinicians to follow the logic of the selection process. For instance, the model prioritizes the BIS-2 equation for patients with diabetic nephropathy, while favoring cystatin C-based equations for patients with higher BMI/BSA, where creatinine-based measures are often less reliable due to muscle mass variations.

While the study is limited by its specific geographic cohort (mainland China), it proves that ensemble-style machine learning can effectively bridge the gap between complex direct measurements and simplified estimation formulas. Future work will focus on multi-center external validation to ensure global applicability.

Summary updated

Abstract

Chronic kidney disease (CKD) has become a worldwide public health problem and accurate assessment of renal function in CKD patients is important for the treatment. Although the glomerular filtration rate (GFR) can accurately evaluate the renal function, the procedure of measurement is complicated. Therefore, endogenous markers are often chosen to estimate GFR indirectly. However, the accuracy of the equations for estimating GFR is not optimistic. To estimate GFR more precisely, we constructed a classification decision tree model to select the most befitting GFR estimation equation for CKD patients. By searching the HIS system of the First Affiliated Hospital of Zhejiang Chinese Medicine University for all CKD patients who visited the hospital from December 1, 2018 to December 1, 2021 and underwent Gate's method of <sup>99m</sup>Tc-DTPA renal dynamic imaging to detect GFR, we eventually collected 518 eligible subjects, who were randomly divided into a training set (70%, 362) and a test set (30%, 156). Then, we used the training set data to build a classification decision tree model that would choose the most accurate equation from the four equations of BIS-2, CKD-EPI(CysC), CKD-EPI(Cr-CysC) and Ruijin, and the equation was selected by the model to estimate GFR. Next, we utilized the test set data to verify our tree model, and compared the GFR estimated by the tree model with other 13 equations. Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Bland-Altman plot were used to evaluate the accuracy of the estimates by different methods. A classification decision tree model, including BSA, BMI, 24-hour Urine protein quantity, diabetic nephropathy, age and RASi, was eventually retrieved. In the test set, the RMSE and MAE of GFR estimated by the classification decision tree model were 12.2 and 8.5 respectively, which were lower than other GFR estimation equations. According to Bland-Altman plot of patients in the test set, the eGFR was calculated based on this model and had the smallest degree of variation. We applied the classification decision tree model to select an appropriate GFR estimation equation for CKD patients, and the final GFR estimation was based on the model selection results, which provided us with greater accuracy in GFR estimation.
BibTeX
@inproceedings{fan2022construct,
  author = {Zhenliang Fan and Qiaorui Yang and Zhuohan Xu and Ke Sun and Mengfan Yang and Riping Yin and Dongxue Zhao and Junfen Fan and Hongzhen Ma and Yiwei Shen and Xia Hong},
  title = {{Construct a classification decision tree model to select the optimal equation for estimating glomerular filtration rate and estimate it more accurately}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-19185-6},
}
Figure