📄 medicine

Machine Learning-Based Identification of Key Predictors of Estimated Glomerular Filtration Rate in Older Taiwanese Women

This study utilized machine learning on data from 1,344 older Taiwanese women to identify age, serum uric acid, blood pressure, and C-reactive protein as the primary non-linear predictors of estimated glomerular filtration rate, with the XGBoost model demonstrating superior predictive performance compared to other algorithms.

Original authors: Shu-Hua Chen, Yao-Jen Liang, Dee Pei

Published 2026-08-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shu-Hua Chen, Yao-Jen Liang, Dee Pei

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your body is a bustling city, and deep within it lies a massive, high-tech water treatment plant called the kidneys. Their job is to filter out the trash and keep the water clean so you can stay healthy. To check if this plant is working, doctors usually look at a specific chemical in your blood called creatinine. Think of creatinine as a "smoke signal" from your muscles; the more muscle you have, the more smoke you see. But here's the tricky part: as people get older, especially women, they often lose muscle. This means the smoke signal gets weaker, and the old way of checking the water plant might accidentally tell you the plant is working better than it actually is. It's like trying to judge a factory's output by how much smoke comes out of a chimney that's been partially blocked by snow. Because of this, scientists are looking for smarter ways to figure out how well the kidneys are really doing, especially for older women who might be losing muscle without realizing it.

This is where a team of researchers from Taiwan stepped in with a digital magnifying glass called "Machine Learning." Instead of just using a simple formula, they asked a computer to act like a super-detective. They gathered health data from 1,344 older Taiwanese women (aged 65 and up) and fed it into five different computer programs. These programs were tasked with finding the hidden clues that best predict how well the kidneys are filtering, a score known as the estimated Glomerular Filtration Rate, or eGFR. The researchers wanted to see if these smart computer models could spot complex patterns that human doctors or simple math might miss, particularly when dealing with a mix of age, blood pressure, diet, and lifestyle habits.

The study tested five different "detective" models: a basic one called Linear Regression, and four more advanced ones like Random Forest, MARS, XGBoost, and Support Vector Regression. After feeding the computer all the data—including things like age, blood pressure, uric acid levels, and even how much someone smoked or exercised—the models tried to guess the women's kidney scores. The results showed that the most advanced model, a powerhouse called XGBoost, was the best detective. It achieved a score (R²) of 0.60 and an error rate (RMSE) of 9.8, which was better than the other models and much better than the basic math approach.

So, what clues did the computer find most important? The XGBoost model revealed that the biggest factors determining kidney function were age, serum uric acid, blood pressure (both systolic and diastolic), and C-reactive protein (a marker of inflammation). Interestingly, the model found that lifestyle factors like smoking, alcohol, and exercise, as well as socioeconomic factors like income and education, contributed very little to the prediction. The researchers suggest this might be because the data on these lifestyle habits was often missing or unbalanced in their dataset (for example, income data was missing for over 63% of the participants).

The paper concludes that while simple math works okay, machine learning is much better at untangling the messy, non-linear relationships between our body's signals. It suggests that for older women, the state of their kidneys is tightly linked to their age, their uric acid levels, their blood pressure, and how inflamed their body is, rather than just their daily habits or wallet size. The authors note that while their model is promising, it needs to be tested on other groups of people to be sure it works everywhere, and future studies should try to get better, more complete data on lifestyle habits to see if those factors become more important when the numbers are clearer.

Technical Summary: Machine Learning-Based Identification of Key Predictors of Estimated Glomerular Filtration Rate in Older Taiwanese Women

Problem Statement
Estimated glomerular filtration rate (eGFR) is the standard metric for assessing kidney function but is traditionally derived from serum creatinine. This reliance introduces significant limitations, as creatinine levels are influenced by non-renal factors including age, sex, muscle mass, nutrition, inflammation, and fluid status. These factors are particularly problematic in older adults, where age-related sarcopenia (loss of skeletal muscle mass) can lead to the overestimation of kidney function in older women. While previous studies have linked eGFR to various physiological and lifestyle factors, conventional linear statistical models often fail to capture the complex, non-linear interactions among these high-dimensional clinical variables. There is a need to identify key predictors of eGFR specifically within the demographic of older Taiwanese women using advanced analytical approaches that can model these intricate relationships.

Methodology
This retrospective study utilized data from 1,344 community-dwelling Taiwanese women aged 65 years and older, sourced from the MJ Health Database. The dataset comprised 31 variables: one dependent variable (eGFR) and 30 independent variables spanning demographics (e.g., age, education, income), physiological measurements (e.g., blood pressure, hemoglobin, uric acid, C-reactive protein, lipid profiles), and lifestyle indicators (alcohol consumption, smoking exposure, physical activity, sleep duration).

Lifestyle variables were quantified using composite metrics:

Drink-area: Weekly drinking days × volume × alcohol concentration.
Smoke-area: Cigarettes/day × days/week × years.
Sport-area: Minutes × days/week × intensity score.

Data preprocessing involved median imputation for continuous variables and mode imputation for categorical variables, followed by one-hot encoding for categorical features and standardization (z-score scaling) for continuous predictors. The dataset was split into an 80% training set and a 20% testing set.

Five machine learning models were trained and compared for predictive performance:

Linear Regression (LR): Served as a baseline.
Random Forest (RF): An ensemble of decision trees.
Multivariate Adaptive Regression Splines (MARS): A non-parametric method using piecewise linear basis functions to capture non-linearities and interactions while maintaining interpretability.
XGBoost (eXtreme Gradient Boosting): A high-performance gradient-boosted decision tree algorithm.
Support Vector Regression (SVR): With a Radial Basis Function (RBF) kernel.

Model performance was evaluated using Root Mean Squared Error (RMSE) and the coefficient of determination ( $R^2$ ). Feature importance was derived primarily from the Random Forest model to identify the most significant predictors.

Key Results
The study population exhibited a mean eGFR of 68.9 ± 15.5 mL/min/1.73 m², with a right-skewed distribution and outliers ranging from 7 to 169 mL/min/1.73 m².

Model Performance: XGBoost achieved the highest predictive accuracy with an RMSE of 9.8 and an $R^2$ of 0.60. It was followed by MARS (RMSE = 10.1, $R^2$ = 0.58) and Random Forest (RMSE = 10.2, $R^2$ = 0.55). Both tree-based ensemble methods outperformed SVR (RMSE = 11.8, $R^2$ = 0.40) and Linear Regression (RMSE = 12.5, $R^2$ = 0.35).
Feature Importance: The analysis identified age (importance = 0.25) as the dominant predictor, followed by serum uric acid (0.20), systolic blood pressure (0.15), diastolic blood pressure (0.12), C-reactive protein (CRP) (0.10), and hemoglobin (0.08).
Low-Impact Variables: Socioeconomic factors (income, education) and lifestyle variables (smoking, alcohol, physical activity) contributed minimally to the predictive models. The authors attribute this to substantial data missingness (e.g., 63.5% missing for income, 46.4% for sport-area) and class imbalance (81–85% zero values for alcohol and smoking metrics).

Significance and Claims
The paper claims that machine learning, particularly tree-based ensemble methods like XGBoost and MARS, effectively captures the complex, non-linear relationships among clinical variables that traditional linear models miss. The study highlights that in older Taiwanese women, renal function is primarily determined by age, uric acid levels, blood pressure, and inflammatory status (CRP), rather than socioeconomic or lifestyle factors within this specific dataset.

The authors emphasize the utility of MARS for providing clinically meaningful insights into these complex interactions (such as the interplay between age and hyperuricemia) while maintaining a degree of interpretability often lacking in "black-box" models. The findings suggest that interpretable AI approaches could play a role in precision nephrology and clinical decision support.

Limitations and Future Directions
The authors acknowledge several limitations, including the cross-sectional design which precludes causal inference, and the significant missingness and imbalance in lifestyle and socioeconomic data which may limit generalizability. They note that the skewed distribution of eGFR and the presence of outliers present challenges for regression analysis.

Future work proposed by the authors includes integrating longitudinal data, improving the completeness of lifestyle data, and exploring classification-based approaches (e.g., distinguishing CKD from non-CKD) potentially utilizing techniques like SMOTE to address class imbalance in high-risk subgroups. The study also suggests the potential for combining ensemble machine learning with explainable AI techniques to further enhance clinical applicability.

Technical Summary: Machine Learning-Based Identification of Key Predictors of Estimated Glomerular Filtration Rate in Older Taiwanese Women

More like this