Identification and validation of an explainable early-stage chronic kidney disease prediction model: a multicenter retrospective study

Summary

Background

Chronic Kidney Disease (CKD) has become a significant global public health issue, affecting approximately 10% of adults. Due to the lack of obvious symptoms in the early stages, CKD is often difficult to diagnose in a timely manner, leading to the gradual progression of the disease, which can eventually develop into End-Stage Renal Disease (ESRD). This study applied machine learning (ML) methods to integrate patient clinical data and developed an early CKD prediction model applicable to individuals without diabetes, hypertension, or coronary heart disease. The model aims to enhance the accuracy of early CKD risk assessment, thereby delaying disease progression and improving patient outcomes.

Methods

This study is a retrospective multicenter study conducted in China, including patients with CKD and healthy individuals who underwent physical examinations from February 2021 to April 2024. Six ML methods, including Decision Tree, Multilayer Perceptron, and XGBoost, were used to predict CKD, integrating different combinations of features such as blood routine, urine analysis, and blood biochemistry. Multiple evaluation metrics, including AUC and F1 score, were used to compare the prediction performance. The SHAP interpretability method was applied to assess feature importance and explain the final model's results.

Findings

Data from three hospitals were used in this study, with the dataset divided into training and internal validation sets (CKD: 11,436 cases, non-CKD: 10,004 cases) and an external validation set (CKD: 350 cases, non-CKD: 473 cases). Among the six ML models, XGBoost performed the best. Regarding feature combinations, the "blood routine + urinalysis + basic information" combination yielded the best performance (AUC = 0.9235, external validation AUC=0.8962). Additionally, a web tool was developed in this study to facilitate the application of early CKD risk prediction in clinical practice.

Interpretation

This study applied an interpretable ML model to effectively predict early CKD. Even when using the relatively low-cost "blood routine + urinalysis + basic information" combination, the model still demonstrated high prediction accuracy. This method has potential clinical application prospects and may help identify early CKD, reducing the risk of disease progression.

Funding

This research was supported by the National Key Research and Development Program of China (2023YFC3502903, 2022YFC3502302), National Natural Science Foundation of China (82074580), and Science and Technology Project of Jiangsu Provincial Research Institute of Chinese Medicine Schools (JSZYLP2024011).