Risk factor mining and prediction of urine protein progression in chronic kidney disease: a machine learning- based study

Yufei Lu, Yichun Ning, Yang Li, Bowen Zhu, Jian Zhang, Yan Yang, Weize Chen, Zhixin Yan, Annan Chen, Bo Shen, Yi Fang, Dong Wang, Nana Song, Xiaoqiang DingBMC Medical Informatics Decis. Mak 2023 · DOI

TL;DR

This study developed a machine learning framework to predict the progression of chronic kidney disease (CKD) by analyzing urine protein levels. By combining feature selection, ensemble modeling, and advanced interpretation techniques (SHAP), the researchers achieved a highly accurate prediction model that also provides insights into the key risk factors, offering valuable clinical decision support.

What Problem Does This Paper Address?

Chronic kidney disease (CKD) is a major global health concern, and its progression can lead to severe health consequences and high healthcare costs. Early and accurate prediction of CKD progression is crucial for timely intervention and resource allocation. Traditional methods for assessing CKD progression, such as 24-hour urine protein tests, are often cumbersome, requiring significant patient compliance and lengthy measurement cycles. While machine learning (ML) has shown promise in medical prediction, achieving both high accuracy and interpretability in CKD prediction remains a challenge. Specifically, understanding the detailed relationships between various risk factors and CKD progression has been less explored. This paper aims to address this by developing an ML model that not only predicts CKD progression but also identifies and explains the most influential risk factors.

Methods

The study employed a multi-stage machine learning approach:

  • Data Construction and Preprocessing: Initial data from 1,358 CKD patients was collected, with 100 clinical variables considered. Missing values were handled using imputation techniques (mode for categorical, random forest for continuous). To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training set.
  • Feature Selection: To identify the most informative variables, the researchers used two methods: Recursive Feature Elimination with Logistic Regression (RFE-LR) and Least Absolute Shrinkage and Selection Operator (LASSO). RFE-LR was particularly effective in reducing the feature set to 17 key predictors.
  • Model Development: Several ML classifiers were trained, including Extreme Gradient Boosting (XGBoost), Gaussian Naive Bayes (GNB), Neural Network (NN), Ridge Regression, and Logistic Regression (LR). These models were trained using stratified 10-fold cross-validation.
  • Ensemble Modeling: To further enhance predictive performance, an ensemble model was created by combining the two best-performing single models (LR and XGBoost) using a soft voting integration method.
  • Feature Interpretation: The SHapley Additive exPlanations (SHAP) algorithm was utilized to provide both global and local interpretations of the model's predictions. This allowed for the identification of the most important predictors and an understanding of their impact on CKD risk.

Datasets

The study utilized a retrospective dataset of 1,358 patients with pathologically confirmed CKD treated at Zhongshan Hospital between December 2017 and September 2020. The data included 100 clinical variables. The dataset was split into a training cohort (80%) and a test cohort (20%). No new benchmark datasets were introduced.

Key Insights and Ideas

  • Ensemble Power: Combining multiple ML models (ensemble learning) significantly boosted predictive performance compared to individual models, highlighting the benefit of diverse algorithmic approaches.
  • Interpretability is Key: Simply achieving high accuracy isn't enough. The use of SHAP values was crucial for understanding why the model makes certain predictions, providing actionable insights for clinicians.
  • Beyond Traditional Metrics: The study demonstrated that ML models, incorporating a broader range of clinical variables, significantly outperform traditional kidney function indicators (eGFR, Scr) in predicting CKD progression.
  • Vitamin D's Role: Lower levels of 25-hydroxyvitamin D (25OHD) emerged as a highly significant predictor of moderate-to-severe CKD, underscoring the importance of vitamin D metabolism in kidney health.
  • Albumin as a Crucial Indicator: Reduced serum albumin levels were identified as a strong indicator of CKD risk, suggesting that even values within the lower end of the normal range might warrant closer monitoring in CKD patients.

Key Results

  • The Logistic Regression (LR) model achieved the highest AUC of 0.850 among single machine learning models.
  • The ensemble model, combining LR and XGBoost, further improved the performance with an AUC of 0.856.
  • The RFE-LR feature selection method reduced the number of features to 17, achieving the highest accuracy and AUC compared to using all features.
  • Key predictors of moderate-to-severe CKD progression included lower levels of 25-OH-vitamin, albumin, and transferrin (in males), along with higher levels of cystatin C.
  • The ML model improved prediction accuracy by 17.6% and 24.6% compared to traditional indicators like eGFR and Scr, respectively.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single-Center Study: The study was conducted at a single center, which may limit the generalizability of the findings. Validation in external cohorts is recommended.
  • Correlation vs. Causation: The study identified correlations between predictive factors and CKD progression but did not establish causality.
  • Feature Selection Methods: While RFE-LR and LASSO are effective, the authors suggest that more advanced techniques could potentially identify even more reliable risk factors.
  • Limited Data Types: The dataset primarily included blood-related indicators and excluded medical prescriptions and imaging examinations, which could provide additional predictive power.

Future work could involve validating the model in diverse populations, exploring causal relationships, incorporating advanced feature selection methods, and integrating a wider range of data types. The developed framework is also suggested to be extendable to predicting and interpreting other diseases.

Summary updated

BibTeX
@article{lu2023risk,
  author = {Yufei Lu and Yichun Ning and Yang Li and Bowen Zhu and Jian Zhang and Yan Yang and Weize Chen and Zhixin Yan and Annan Chen and Bo Shen and Yi Fang and Dong Wang and Nana Song and Xiaoqiang Ding},
  title = {{Risk factor mining and prediction of urine protein progression in chronic kidney disease: a machine learning- based study}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2023},
  doi = {10.1186/s12911-023-02269-2},
}
Figure