The data for this study were collected from the CKD screening population in Jing’an District from January 1, 2015, to December 22, 2020. Information obtained included demographic and sociological characteristics, height, weight, diastolic and systolic blood pressure, health insurance type, screening date, urinary protein and urinary albumin-creatinine ratio (UACR), blood creatinine, eGFR, and screening results. In total, 103,960 records were initially screened and CKD diagnoses were categorized based on ICD-10 (International Statistical Classification of Diseases, Tenth Revision) criteria. Records with incomplete or duplicate data were excluded, resulting in a final sample size of 40,686 cases for analysis. These data are considered credible and authentic.

The participants were categorized based on dichotomous variables: 1 for the nonmanagement population (indicating the absence of CKD) and 2 for the management population (indicating the presence of CKD).

We used the 11 factors identified in the univariate analysis as explanatory variables for the LR model. The grouping and assignment of the dependent and independent variables are listed in Table 1.

A database was established using Excel (Microsoft Corp) 2010, and SAS (version 9.4; SAS Institute Inc) statistical software was used for data analysis. The chi-square test was performed for 1-way analysis to select variables for inclusion in the model, with the threshold for statistical significance set at P<.05. Based on the GFR stage, albuminuria (Alb) grouping, and the distribution of data, the study categorized participants for CKD screening into management (suspected and diagnosed patients) and nonmanagement (healthy individuals) populations. The resulting dichotomous LR model was then used for subsequent analysis.

RF is a classification algorithm that uses multiple decision trees to train and predict samples. Specifically, the algorithm samples the training data set N times with replacement and selects a random subset of training samples each time. The remaining undrawn samples are subsequently used to evaluate the prediction error of the model.

The data set of 40,686 participants was randomly split into the following 2 subsets using simple random sampling in Python 3.6: one for validation sample set A including 13,549 cases (or 33.3% of the total data set), and the other for then training sample set B including 27,139 cases (or 66.7% of the total data set). The first subset A constituted the external validation sample set with 3000 cases (accounting for 7.4% of the total data set). The RF algorithm was subsequently applied to the training sample set to evaluate the importance of each variable and construct a CKD risk factor model. This model was used to predict the test sample set, with a minimum prediction accuracy threshold of 70%.

The mean number of feature selections was used for each random tree (mtry) in the model.

For a set with predictors, a typical number is the rounded square root of mtry [12]. Only 11 features were used in this study. We did not use the square root method to calculate mtry. However, we randomly selected a certain number of features each time and fixed ntree to adjust mtry to determine the values that minimized generalization errors as the optimal value of mtry.

The mean number of random trees was used in the RF algorithm (ntree) in the model. (1) Using bootstrap resampling, 20% of the B set was randomly split and was used as an internal validation set and 80% was used as the training set. (2) Assuming that the number of the decision tree was ntree, for each node, mtry features were randomly selected. These mtry features were used to divide the sample set, and the index Gini was used to determine the best partitioning method. (3) For determining the mean error of the test set, steps (1), (2), and (3) were repeated. With each iteration of step (2), the ntree was increased by 1. ntree gradually increased from 1 to 200. We obtained the set for average generalization error, and observed the variation in the average generalization error with ntree. When the optimal model was achieved, we obtained the number of ntrees.

After establishing the RF model, it was used for prediction. Given the abundance of trees in the forest, determining which variables have the most significant impact on predictions can be challenging. Fortunately, an important method was used to assess the significance of variables in the model. Specifically, for each variable, in each decision tree of an RF, the decrease in the splitting criterion function (residual squared or Gini index) caused by that variable was measured. The decrease in magnitude for each decision tree was then averaged to determine the importance of the variable. The importance of each feature variable was ranked and plotted in order, resulting in a variable importance plot.

The Institutional Review Committee Board at Shanghai Changzheng Hospital affiliated with the Naval Medical University approved this study with written consent (No.2016SL020). This observational study analyzed existing data sources, which did not contain any patient-identifiable information. This study did not involve the collection, use, or transmission of individually identifiable data.

The area under the receiver operating characteristic curve (AUC) of the RF model based on the training and test sets was 93.15% (Figure 4). The RF algorithm outputs voting results (0s and 1s), whereas the receiver operating characteristic curve requires voting probability data. Converting probabilities to voting results can lead to error because of extreme probabilities, such as 0.01515526 and 0.98484474. Therefore, we calculated the AUC to assess model performance and the classification prediction rate to indicate the accuracy of the model. Herein, the RF algorithm achieved an accuracy rate of 92.67%, with some degree of error. These results suggest that the model exhibited good predictive power and accurately classified new data samples.

Four possible predicted results were as follows: true positives, false positives, true negatives, and false negatives. Table 5 shows the confusion matrix of the RF model. The precision, recall, and F₁-score were 0.951, 0.984, and 0.967, respectively.

A risk assessment model for CKD was developed in this study using dichotomous LR and RF models. Our results indicate that gender, older age, BMI beyond the normal range, abnormal index eGFR, retirement status, and urban employee medical insurance were significantly associated with a higher risk of CKD. By leveraging the RF model, the most important factors for CKD development were older age, abnormal urinary test results (eg, Alb, UACR, and index PRO indicators), and high BMI.

In China, the number of studies on the assessment of risk factors for CKD and the investigation of methods for risk prediction is increasing and LR analysis is commonly being performed. Feng et al [24] used an adjusted LR model to investigate CKD prevalence and related risk factors in 38 megacities across China. Liu et al [25] and Yang et al [26] performed cross-sectional studies to analyze risk factors for diabetic nephropathy in Shanghai, whereas a community-based, 7-year-long cohort study from Tianjin used LR to examine the association between the high triglyceride waist phenotype and risk of CKD development [27]. Yan et al [28] performed LR analysis to assess the correlation between residual cholesterol levels and CKD, and identify other significant risk factors affecting middle-aged and older individuals residing within a city. Gradual advancements in machine learning models have prompted further scrutiny of the divergent performance and inherent limitations of the conventional LR approach. To distinguish this study from previous studies that followed the LR approach for exploratory purposes, we used the RF algorithm to rank risk factors that were subjected to single-factor analysis according to their relevance and consequently evaluated comparative predictive precision by performing LR analysis using training samples. Our results reveal that both the RF and LR models achieved an overall accuracy rate exceeding 90% in the prediction test set. Conversely, the dichotomous LR model exhibited a marginally superior predictive performance than the RF model. Nevertheless, one should pay attention to the tendency of LR to result in excessive ORs when imbalanced data are used. Although LR exhibits excellent predictive abilities and desirable attributes such as high accuracy and stability, and ease of operation with a minimal possibility of overlearning during classification prediction, RF has the ability to assess the importance of variables when classifying data into suitable categories while compensating for errors in imbalanced sets of categorical data.

Our results indicate that age was the primary significant factor in the RF model, and LR analysis confirmed that higher age was significantly associated with CKD. Compared to participants aged ≤64 years, those aged 65-75 years and older were at a significantly higher risk of CKD, which is in line with previous results [29,30]. The risk of CKD increases with age; thus, early screening and risk prediction for CKD are crucial for middle-aged and older people.

A cross-sectional study published in The Lancet [31], using a nationally representative sample of Chinese adults also identified independent factors associated with kidney damage, which included age and gender. Age and gender are independent CKD risk factors [32]. Many studies worldwide have shown that women are at a higher risk of CKD [33,34], and similar observations have been reported in China [24,30]. This correlation may be attributed to differences in the prevalence of primary diseases and the availability of medical resources across genders [35]. However, our results show that females in the survey population were at a lower risk of CKD than were males, which is inconsistent with the majority of previous results. Our data include information regarding the registered population in a district of Shanghai. The exclusion of samples with incomplete information and regional differences, as well as the presence of unregistered patients, may have led to bias, ultimately yielding inconsistent results.

Next, this study shows that people with a higher-than-normal BMI were at a higher risk of CKD, similar to a time-series study that investigated risk factors regarding CKD burden in China from 1991 to 2011 and identified the correlation between high BMI and CKD [36]. Obesity is an important risk factor for CKD worldwide [24,25,37-39]. Potential obesity-associated factors that may lead to or aggravate CKD include hemodynamic disorder and renal tissue hypoxia [40,41]. However, weight loss through diet and regular exercise can reverse kidney damage; hence, maintaining a healthy lifestyle and controlling body weight could prevent or decelerate CKD progression to a certain extent [42]. Additionally, this study shows that CKD risk was higher in people who had urban employee medical insurance. These people were employed and had relatively better economic conditions; however, health risk factors such as work stress and unhealthy lifestyles probably contribute to an increased CKD risk [43].

Moreover, people with abnormal urine test results (Alb, UACR, and PRO indicators) were at a higher CKD risk, which is consistent with previous results reported worldwide [36,44,45]. Similarly, a Chinese study using 4 machine learning models, comprising 19,270 adult samples, showed that UACR, Alb, age, and gender were important CKD risk factors [44]. Urine tests can serve as an early warning system for CKD detection. Similarly, our risk prediction model could guide decision-making regarding early CKD screening.

Herein, we effectively assessed the risk of CKD by combining internal data for model construction and testing. However, this study has some limitations. First, the generalization ability of the model remains unknown because the study did not include external data for external validation. Second, owing to the bias in data collection, our results were inconsistent with those of the previous studies. Finally, more prospective studies are required to verify the predictive power and practical utility of our model. Thus, health care professionals should routinely evaluate the level of agreement within and between models before reaching any clinical decision on the basis of the present limitations and previous findings [46].

In conclusion, the RF model has significant predictive value for assessing risk factors associated with CKD and is capable of correcting errors in imbalanced categorical data sets. It can be used to screen individuals with risk factors, which is of great significance for early intervention and prevention of CKD.

For the prevention and treatment of CKD, early intervention can involve a low-protein diet, regular physical examination, actively promoting urine examination, and screening of high-risk groups to achieve early detection, early treatment, early diagnosis, and early intervention of CKD, and to reduce the social and personal losses caused by diseases and improve people’s quality of life.