Application of Machine Learning for Identifying Factors Associated with Renal Function Impairment in Gouty Arthritis Patients

Abstract

Gouty arthritis (GA) and its association with kidney failure present significant challenges in healthcare, necessitating effective detection and management strategies. GA, characterized by the deposition of monosodium urate crystals in joints and other tissues, leads to inflammation and severe joint pain, often accompanied by metabolic comorbidities such as myocardial infarction and diabetes. Although GA has been widely studied in the medical field, limited research has explored the use of machine learning (ML) to identify key biomarkers affecting disease progression. This study aims to bridge this gap by leveraging ML models for predictive analysis. In this study, machine learning models such as decision trees, random forests, logistic regression, and artificial neural networks were used to classify GA using demographic, clinical, and laboratory data, and, most importantly, to identify the factors that affect GA. The analysis yielded promising results, with the decision tree model achieving the highest accuracy of 92.85%. Moreover, key factors such as urea, creatinine, and hemoglobin levels were identified during the initial attack, shedding light on the pathophysiology of GA. This study demonstrates how ML methods help identify key factors affecting GA and assist in disease management. By leveraging machine learning techniques, it is possible to refine the factors affecting GA and inform personalized interventions, ultimately improving patient care and outcomes.

1. Introduction

Gouty arthritis (GA) is a prevalent inflammatory disease that significantly impacts joint health and kidney function. GA is one of the most common inflammatory arthritides, causing severe joint pain, loss of function, and deformity [1,2]. Furthermore, GA is closely associated with metabolic comorbidities that can lead to myocardial infarction, type 2 diabetes, chronic kidney disease, and premature death [3]. High glucose levels, insulin resistance, and metabolic disorders trigger inflammation in the body, leading to the accumulation of MSU crystals and an increase in serum urate levels. In metabolic disorders, changes in insulin levels and insulin resistance can reduce the kidneys’ capacity to excrete urate, thereby contributing to the development and severity of GA [4]. If left untreated, the chronic inflammatory process against MSU crystals continues. Patients with gouty arthritis are exposed to local and systemic inflammation, especially during the attack period [5].

C-reactive protein (CRP), leukocytes, platelets, thrombocytes, and ferritin increase and albumin decreases with systemic inflammation [6]. Hematologic parameters such as albumin and hemoglobin levels and lymphocyte, neutrophil, and platelet counts are determined by laboratory tests that indicate inflammation, nutrition, and prognosis. GA and hyperuricemia are associated with serious renal complications that can affect kidney function [7]. Impaired renal function occurs in half of patients with GA [8].

One of the important criteria for the detection of Gout by medical methods is recurrent arthritis attacks caused by high serum uric acid levels (hyperuricemia) and monosodium urate (MSU) crystal deposition [9,10]. The number of attacks on arthritis is one of the important criteria for detecting the disease. Another important criterion is the most common form of inflammatory arthropathy, which is known to cause severe pain and joint changes [11,12]. In patients with gout arthritis (GA), the presence of systemic inflammation may lead to a decrease in serum albumin levels. Inflammation reduces albumin production in the liver, resulting in lower albumin levels [13]. Additionally, since albumin has the capacity to bind free urates, low albumin levels may lead to an increased amount of free urate in the bloodstream, thereby raising the risk of monosodium urate (MSU) crystal accumulation [14,15]. This condition may contribute to more severe inflammation and a worsening disease course in GA patients. Furthermore, low albumin levels are associated with impaired organ function, particularly in the kidneys, and a decline in overall health status [16].

This study applies machine learning (ML) models to analyze demographic, clinical, and laboratory data, aiming to uncover critical biomarkers associated with GA. By integrating ML-based predictive models, this research provides new insights into the disease mechanism and potential risk factors.

The important points of this study are as follows:

• Looking only at serum uric acid levels is not always sufficient to diagnose GA. Therefore, a model was developed, which is expected to provide insights into additional risk factors and the progression of GA.
• A different perspective from traditional approaches is offered by analyzing the contribution of multiple biomarkers with ML models instead of only a single biomarker (uric acid).
• Detailed analysis was performed using ML models to identify the factors affecting GA.
• This study focused on examining the biomarkers that play a role in the development of GA and ranking the importance of these factors.
• Based on the available information to date, although there are a few articles regarding the detection and classification of gout disease, there is currently no study conducted using machine learning methods for identifying the factors affecting the disease.

2. Literature Survey

The use of the decision tree (DT) model for the classification of glomerular filtration rate (GFR) values in the determination of gout disease holds significant promise. Decision trees have been widely utilized in various domains for classification tasks, and their effectiveness in providing interpretable and accurate predictions makes them suitable for medical diagnostics. Decision tree models have been successfully applied in medical research for the determination of disease risk factors and support for clinical diagnosis [17]. Additionally, decision tree models have been used for the classification of cardiac disorders based on electrocardiogram data, demonstrating their potential in medical classification tasks [18]. Moreover, decision tree models have been employed in the analysis of infectious disease distribution status, indicating their capability to describe the level of knowledge distribution based on the decision tree [19].

The use of random forest (RF) models for the classification of glomerular filtration rate (GFR) values in the context of gout disease determination holds significant promise. Random forest models have demonstrated high accuracy and robustness in various domains, including ecology, neuroimaging data analysis, and cardiovascular disease prediction [20,21,22].

The use of the logistic regression (LR) model for the classification of glomerular filtration rate (GFR) values in the determination of gout disease presents a significant avenue for research and clinical application. Logistic regression has been widely employed in medical research for disease prediction and risk factor identification. For instance, a study [23] utilized logistic regression to examine the decadal trends of gout and hyperuricemia prevalence in the United States, demonstrating the model’s applicability in epidemiological investigations. Furthermore, logistic regression has been used in the identification of gout-related metabolic indices, highlighting its role in understanding the genetic variation associated with gout risk [24]. In addition, logistic regression has been applied in the context of gout-related comorbidities. For example, a study by [25] investigated the association of gout with an increased risk of hypertension and diabetes mellitus among stroke survivors, showcasing the model’s utility in exploring disease interrelationships. Moreover, logistic regression has been utilized in the identification of patients with gout, demonstrating its effectiveness in developing questionnaires for epidemiological studies [26]. Furthermore, logistic regression has been employed in the context of gout risk assessment and classification. The authors of [27] demonstrated the independent relationship between higher Healthy Eating Index-2015 scores and a decreased risk of gout, highlighting the model’s role in assessing lifestyle factors associated with the disease. Additionally, logistic regression has been used in the context of gout and hyperuricemia risk assessment, showcasing its potential in understanding the relationship between dietary habits and disease susceptibility [27].

In a different study, the aim was to develop a model that predicts future estimated glomerular filtration rate (eGFR) values in individuals with type 2 diabetes and chronic kidney disease (CKD) and validates them externally. The study used data from three large cohorts in Europe to predict how kidney function might change over a 5-year period, based on 13 variables (e.g., hemoglobin A1c, hemoglobin, serum cholesterol, mean arterial pressure, urine albumin-to-creatinine ratio) collected on routine clinic visits. The multivariable linear mixed-effects model was used for estimation in the article. The R² score of the model was determined as 0.74 [28]. In a study by Haishuai Wang et al., a new deep learning model (dlGFR) was developed and tested in comparison with the MDRD and CKD-EPI equations based on traditional statistical methods. They used serum creatinine, age, gender, and race variables to predict GFR. The dataset they used includes 22,991 observations of 5186 patients. They used the dlGFR, CKD-EPI, and MDRD equations and LR model in their study. The dlGFR model they designed had the highest score of 88.3% accuracy [29]. Another study compared various machine learning models with the traditional EKFC (European Kidney Function Consortium) equation for glomerular filtration rate (GFR) estimation. The study used 13 different cohorts of 19,629 patients and compared the EKFC (European Kidney Function Consortium) equation with machine learning models (random forest, XGBoost, linear regression, etc.). When the results of the models were compared, the lowest error rate (MSE) belonged to the RF model with 294.24 [30].

This study demonstrates the effectiveness of DT, RF, and LR models in classifying GFR values to assess gout disease risk, offering a data-driven approach for identifying key contributing factors.

3. Materials and Methods

3.1. Subject and Limitations

In this study, we analyzed patients over the age of 18 years who presented to the rheumatology outpatient clinic between 1 January 2018 and 1 July 2023 and were diagnosed with acute gouty arthritis. The reason for selecting the dataset within a specific time period is to ensure both the consistency of the data and the reliability of the models to be used. This period was chosen because an experienced rheumatology specialist accurately and comprehensively recorded patients’ anamnesis, physical examination findings, laboratory measurements, imaging results, and treatment responses in the hospital data system in an up-to-date manner. Patients with data during acute gouty arthritis were included in this study. These data belong to patients admitted to the Department of Internal Medicine, Rheumatology Department of Rize Recep Tayyip Erdoğan University Rize Training and Research Hospital. The Ethics Committee report was obtained for use within the scope of this study. In our study, in addition to identifying the risk factors for gout, parameters that may be associated with impaired kidney function, which could affect the course and prognosis of the disease, were carefully selected. Patients with blood diseases affecting hematologic parameters, liver disease, renal disease such as glomerulonephritis, autoimmune disease, acute infection, antiplatelet and immunosuppressive drug use, or missing complete blood count and albumin data were excluded. The liver is the organ that plays a crucial role in the production of inflammatory markers such as albumin and C-reactive protein (CRP). Therefore, patients with liver diseases such as liver cirrhosis, which could affect liver function, were excluded from this study [31,32].

3.2. Dataset

Preprocessing studies were carried out on the data first. Patients with a null value were excluded from the evaluation. The features in the dataset were normalized and the standard scalar distribution was applied so that the weights would not lose their effectiveness during the classification process. After this stage, the data were first divided into training and test data for training the models. Since the dataset used in this study was limited to training machine learning models, 90% of the data were allocated for training and 10% for testing. Demographic and clinical characteristics including age, gender, body mass index, smoking status, alcohol consumption, comorbidities (diabetes, hypertension, coronary artery disease, and hyperlipidemia), physical examination findings, and medications were recorded. Investigations were performed during the attack, and patients with at least two attacks were included in this study. The criteria for gouty arthritis are based on the work of [33]. Laboratory values included complete blood count (hemoglobin, neutrophil count, lymphocyte count, and platelet count), biochemical parameters (urea, creatinine, glucose, alanine aminotransferase (ALT), aspartate aminotransferase (AST), triglyceride (TG), total cholesterol, high-density lipoprotein (HDL), low-density lipoprotein (LDL), lactic dehydrogenase (LDH), uric acid, total protein, and albumin), erythrocyte sedimentation rate (ESR), and C-reactive protein (CRP). Patients were divided into two groups: GFR < 60 mL/min/1.73 m² (CKD) and GFR ≥ 60 mL/min/1.73 m². This study included 276 patients. Of these patients, 97 (35.1%) were female and 179 (64.9%) were male. The mean age of the patients was 66.85 ± 12.17 years, and the mean age was found to be significantly higher in female than in male patients with GFR ≥ 60 mL/min/1.73 m² compared to GFR < 60 mL/min/1.73 m². The mean GFR of the patients was 65.31 ± 20.97 mL/min/1.73 m², 162 (58.7%) had a GFR ≥ 60 mL/min/1.73 m², and 114 (41.3%) had a GFR below 60 mL/min/1.73 m². The mean GFR and the proportion with GFR ≥ 60 mL/min/1.73 m² were higher in males than in females.

Table 1. Demographic characteristics and GFR levels of patients.

Variables	Male	Female	Total
Number of patients	179 (64.90%)	97 (35.10%)	276 (100.00%)
Age, mean + standard deviation (SD)	64.49 ± 12.40	71.22 ± 10.47	66.85 ± 12.17
Age at first attack + SD	58.87 ± 12.78	65.62 ± 10.94	61.24 ± 12.56
GFR + SD	68.97 ± 21.09	58.56 ± 19.07	65.31 ± 20.97

shows the summary of the information presented.

3.3. Hyperparameter Tuning of Machine Learning Model

Randomized search hyperparameter optimization (RHO) is a systematic method used to identify the optimal combination that provides the best model performance. The basic principle of RHO is to sample a certain number of configurations of values in the hyperparameter space. This technique allows for better exploration of the hyperparameter space when the number of combinations can be very large [34]. One of the advantages of RHO is that it reduces the computational burden associated with combinations of values in the hyperparameter space [35]. RHO reaches the optimum values by random sampling. This random sampling helps to avoid local optima that trap deterministic methods such as grid search [36]. In this study, since the combination of values in the hyperparameter space is large, the optimal hyperparameters were determined by random search.

3.4. Decision Tree Classification

Decision trees (DTs) are a fundamental component of machine learning and were first introduced in 1986 [37]. They have a flowchart-like structure where each internal node represents a test on an attribute, each branch denotes the outcome of the test, and each leaf node holds a class label [38]. The decision tree model is widely used for classification tasks and is known for its interpretability and ease of implementation [39]. This model is particularly valuable in providing explainable machine learning, as it explicitly demonstrates how different features contribute to the prediction [39].

In the decision trees, the concept of entropy is used to find the best discriminative features. Entropy is used to measure the uncertainty, inequality, and randomness of the data in a dataset [40]. Entropy measures the available information and, as shown in Equation (1), assumes a dataset represented by T consisting of n numbers, assigning values between 0 and 1.

$$T=\left\{C_1,\right.\left.C_2,C_3,\dots ,C_n\right\}$$

(1)

The inference obtained using entropy is shown in Equation (2). The entropy formula was developed by [41]. After the nodes are weighed using entropy, the tree structure emerges.

$$T\left(C_1,C_2,C_3,\dots ,C_n\right)={\displaystyle \sum }_{i=1}^n\left(P_i\right)*lo{g}_2\left({\displaystyle \frac{L}{P_i}}\right)$$

(2)

Table 2. DT model hyperparameters.

Hyperparameter Name	Value
Criterion	Entropy
Max depth	5
CCP alpha	0.018918996053595438
Max depth	5
Max leaf nodes	15
Max features	None
Min samples leaf	10
Min samples split	11
Splitter	Best

shows the hyperparameter values for the DT model used in this study.

3.5. Random Forest Classification

Random forest (RF) is an ensemble machine learning method that was introduced by [42]. It is a combination of tree predictors, where each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest [42]. This approach aims to improve the predictive performance and robustness of individual decision trees by aggregating their outputs. Random forest is particularly suitable for large datasets and has shown rapid development due to its effectiveness in handling complex data [43].

The working principle of random forest involves the construction of multiple decision trees during the training phase. Each tree is built using a random subset of the training data and a random subset of the features. This randomness introduces diversity among the trees, which helps in reducing overfitting and capturing different aspects of the data. During the prediction phase, the individual trees “vote” on the outcome, and the most popular class label is chosen as the final prediction [44].

Although the RF model uses entropy and Gini Index formulas to determine the best discriminative feature selection, it predominantly uses the Gini Index formula. The Gini Index was developed by [45], and its formula is given in Equation (3).

$$Gini Index=1-{\displaystyle \sum }_{i=1}^n{\left(P_i\right)}^2$$

(3)

Table 3. RF model hyperparameters.

Hyperparameter Name	Value
N Estimators	242
Class weight	Balanced
Criterion	Gini
Max depth	6
CCP alpha	0.009390071578941839
Bootstrap	True
Max samples	0.9670121495383591
Max features	Sqrt
Min samples leaf	15
Min samples split	3

shows the hyperparameter values of the RF model used in this study.

3.6. Logistic Regression Classification

Logistic regression (LR), developed by David Cox in 1958, is a statistical method used for modeling the probability of a binary outcome based on one or more predictor variables [46]. The working principle of logistic regression involves fitting the data to a logistic function, which transforms the output of a linear regression model to a range between 0 and 1, representing the probability of the occurrence of a particular event [46]. This method is particularly useful for analyzing observational data when adjustment is needed to reduce potential bias resulting from differences in the groups being compared [46].

Now, let us delve into the LR model, considering the count of explanatory variables as p [47]. If we regard the collection of measurement outcomes comprising the response and explanatory variables $\left(y_i; x_{i1}, x_{i2 },\dots ,x_{ip}\right)$ as a singular data point, then the LR model, utilizing a total of n data points, can be formulated in Equation (4).

$$\begin{gathered} \left[\begin{array}{c}{y}_1\\ y_2\\ y_3\\ ⋮\\ ⋮\\ ⋮\\ y_n\end{array}\right]=\left[\begin{array}{ccccc}1& x_{11}& x_{12}& \cdots & x_{1p}\\ 1& x_{21}& x_{22}& \cdots & x_{1p}\\ 1& x_{31}& x_{32}& \cdots & p\\ ⋮& ⋮& ⋮& \cdots & ⋮\\ ⋮& ⋮& ⋮& \cdots & ⋮\\ ⋮& ⋮& ⋮& \cdots & ⋮\\ 1& x_{n1}& x_{n2}& \cdots & x_{np}\end{array}\right]\left[\begin{array}{c}{\beta }_0\\ {\beta }_1\\ {\beta }_2\\ ⋮\\ ⋮\\ ⋮\\ {\beta }_p\end{array}\right]+\left[\begin{array}{c}{\epsilon }_1\\ {\epsilon }_2\\ {\epsilon }_3\\ ⋮\\ ⋮\\ ⋮\\ {\epsilon }_p\end{array}\right] \\ y=X·\beta +\epsilon \end{gathered}$$

(4)

In this context, y represents the response variable, x stands for the explanatory variable, β denotes the regression co-efficient, and ε signifies the residual error unaccounted for by the explanatory variables [47]. It should provide a concise and precise description of the experimental results, their interpretation, as well as the experimental conclusions that can be drawn. The hyperparameter values of the LR model used in this study are presented in

Table 4. LR model hyperparameters.

Hyperparameter Name	Value
C	1.8183496720710062
L1 ratio	0.18340450985343382
Max Iter	500
Penalty	L1
Solver	Saga

.

3.7. Artificial Neural Networks

Artificial neural networks (ANNs) are computational models inspired by the organization and functioning of biological neurons [48,49,50]. These networks consist of interconnected nodes, or artificial neurons, that process information and are capable of learning complex patterns and relationships [51]. The concept of ANNs is rooted in the idea of mimicking the human brain’s neural structure to perform tasks that require pattern recognition, classification, and prediction [52].

The first and simplest type of ANN is the feedforward neural network, which was among the earliest designs in artificial neural network development [52]. ANNs work by receiving input data, processing it through multiple layers of interconnected neurons, and producing an output based on learned patterns and weights assigned to connections between neurons [53]. Through a process known as training, ANNs adjust these weights iteratively to minimize errors and improve the network’s ability to make accurate predictions [54]. The hyperparameter values of the ANN model used in this study are presented in

Table 5. DT model hyperparameters.

Hyperparameter Name	Value
Activation function	ReLU
Optimizer	SGD
Learning rate	0.001
Loss	Sparse categorical entropy
Validation split	0.1
Epoch	100
Batch size	4

.

$$y={\displaystyle \frac{1}{\left(1+e^{-\beta x}\right)}}$$

(5)

4. Results

In this section, the confusion matrix, accuracy, precision, recall, and F1 score values of the models are analyzed to examine the performance of the models used. In Section 4.1, these scores are explained, and in Section 4.2, detailed results are presented.

4.1. Evaluation Metrics

A confusion matrix is a table utilized for delineating the efficacy of a model [55], as shown in Figure 1.

Accuracy (A) represents the proportion of correctly classified instances out of the total instances and is a fundamental measure of overall model performance, as shown in Equation (6).

$$A={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(6)

Precision (P), on the other hand, quantifies the proportion of true positive predictions out of all positive predictions made by the model. It is particularly relevant when the cost of false positives is high, and is denoted as Equation (7).

$$P={\displaystyle \frac{TP}{TP+FP}}$$

(7)

Sensitivity (S) measures the proportion of true positive predictions out of all actual positive instances in the dataset. It is crucial for scenarios where the cost of false negatives is significant, and is shown as Equation (8).

$$P={\displaystyle \frac{TP}{TP+FN}}$$

(8)

F1 score (F), which is the harmonic means of precision and recall, provides a balanced assessment of a model’s performance, especially when dealing with imbalanced datasets, as seen in Equation (9).

$$F=2*{\displaystyle \frac{P*S}{P+S}}$$

(9)

The Specificity (SP) score is a measure used in diagnostic tests to determine the proportion of true negative results among all individuals without the condition of interest, shown in Equation (10).

$$S={\displaystyle \frac{TN}{TN+FP}}$$

(10)

4.2. Performance of Models

This section contains information about the performance results of the models used in this study. The accuracy, precision, recall, F1 score, and ROC accuracy results of all models were extracted, and these results are given in detail in

Table 6. Results of ML models.

Model	A	P	S	F	SP	ROC AUC
Decision tree	0.928571	0.866667	1.000000	0.928571	0.866667	0.933333
Random forest	0.857143	0.846154	0.846154	0.846154	0.866667	0.856410
Logistic regression	0.821429	0.833333	0.769231	0.800000	0.933333	0.817949
ANN	0.857100	0.800000	0.923100	0.857100	0.800000	0.861500

. Looking at the results in Table 6, the DT model gave the best result, with an accuracy rate of 92.85%. The worst result was given by the LR model, with an accuracy rate of 82.14%.

Explanations of TP, TN, FP, and FN values shown in the confusion matrix results in

Table 7. Detailed CM results of the models.

Models	TP	TN	FP	FN	Samples
Decision tree	13	13	2	0	28
Random forest	11	13	2	2	28
Logistic regression	10	13	2	3	28
ANN	12	12	3	1	28

are as follows:

• TP (true positive) number indicates the classification of patients with non-attack gout patients (Non-Gout) according to the test data.
• TN (true negative) number indicates the classification of gout attacks patients (Gout) according to the test data.
• FP (false positive) count indicates the number of patients in whom gout attack patients without non-attack gout patients according to the test data.
• FN (false negative) count indicates the number of patients in whom non-attack gout patients without gout attack patients according to the test data.

The classification performance of DT, RF, LR, and ANN models used in this study in the classification of gout attack patients and non-attack gout patientes is given in

Table 8. Disease classification performance of the models.

Models
Class	DT	RF	LR	ANN
Gout	13	11	10	12
Non-gout	13	13	13	12

.

Along with the results of the classification performance of the models, the features affecting the classification performance of the model were also determined. The purpose of determining these features is to determine the features that affect the performance of the models and to reveal the conditions that are effective in gout attack with ML models. Figure 2 shows the ROC curve of the RF model.

Table 9. The most important factors identified with RF.

Feature Name	RF Classification (%)
CREATIN	0.234179
UREA	0.162304
HAEMOGLOBIN	0.093477
ALBUMIN	0.043521
URIC ACID	0.042693
LDH	0.031623
GLUCOSE	0.027593
MCV	0.026070
ALT	0.025176
LYMPHOCYTE	0.024879
EOSINOPHIL	0.023152
PLT	0.022635
HDL	0.022486
TOTAL PROTEIN	0.021319
SEDIMENTATION	0.020957
MONOCYTE	0.020675
WBC	0.020099
CRP	0.019134
LDL	0.018854
MPV	0.017552

shows the most important factors affecting classification performance with the random forest (RF) model and determining gout disease status. Accordingly, the most important factor determining the gout disease status is whether the patient has Creatin or not with 23.41%.

When

Table 10. The most important factors identified with DT.

Feature Name	DT Classification (%)
CREATIN	0.588792
HAEMOGLOBIN	0.108985
URIC ACID	0.065655
HDL	0.058617
UREA	0.046587
SEDIMENTATION	0.023664
LDH	0.016008
PLT	0.015976
TOTAL PROTEIN	0.014848
TRIGLYCERIDE	0.013363
EOSINOPHIL	0.011136

is examined, the decision tree (DT) model shows the most important factors that determine gout disease status. According to this, the most important factor determining gout disease status is whether the patient has Creatin with 58.87%, while the least important factor is Eosinophil with 1.11%. Figure 3 shows the ROC curve of the DT model.

In tree-based models, the factors affecting the performance of the models are calculated with percentage values, while in the LR model, the co-efficient value is considered. In ANN model, Permutation Importance Score, SHAP (SHapley Additive exPlanations) Values, or LIME (Local Interpretable Model-agnostic Explanations) Explainable values are examined to determine the factors affecting the performance of the model. In this study, the co-efficient value of the factors affecting the performance of the LR model is examined, while the Permutation Importance Score values are examined for the ANN model.

Table 11. The most important factors identified with LR.

Feature Name	LR Classification Co-Efficient
CREATIN	3.418836
UREA	2.827242
URIC ACID	1.579555
HYPERTENSION	1.329566
LDH	0.963876
HYPERLIPIDEMIA	0.660478
URICOLYSIS	0.651395
EOSINOPHIL	0.469698
NOTROPHILE	0.456201
HDL	0.386265
MPV	0.350879
GLUCOSE	0.329433
MONOCYTE	0.057605

shows the most important factors determining gout disease status in the logistic regression (LR) model. Accordingly, Creatin was the most important factor determining gout status, while Monocyte was the factor with the lowest effect. Figure 4 shows the ROC curve of the LR model.

Finally, the factors in

Table 12. Most important factors identified with ANN.

Feature Name	ANN Feature Importance (%)
CREATIN	0.125542
UREA	0.114828
HAEMOGLOBIN	0.083080
DIABETES	0.067443
URIC ACID	0.063294
CORONARY ARTERY DISEASE	0.057614
HYPERTENSION	0.044978
HYPERLIPIDEMIA	0.041536
PLT	0.040198
ALBUMIN	0.038311
CRP	0.038147
SEDIMENTATION	0.032884
ALT	0.028782
LDH	0.025137
LYMPHOCYTE	0.020704
EOSINOPHIL	0.017293
MCV	0.016175
NOTROPHILE	0.014993
WBC	0.012409
MONOCYTE	0.011735
MPV	0.011646
LDL	0.011591

show the most important factors determining gout disease status in the artificial neural network (ANN) model. Figure 5 shows the ROC curve of the ANN model. Accordingly, the most important factor determining gout disease status is whether the patient has Creatin, while the least important factor is LDL. Figure 6 and Figure 7 show the training and loss graphs of the ANN model, respectively.

Gouty arthritis (GA) is an inflammatory disease characterized by the accumulation of monosodium urate (MSU) crystals in the joints. This condition is particularly associated with kidney function and metabolic diseases. Urea is a waste product related to protein metabolism that is excreted from the bloodstream by the kidneys. In GA, kidney function declines due to tubulointerstitial nephritis and stone formation. Kidney failure or dysfunction leads to elevated urea levels. Moreover, high urea levels, resulting from impaired kidney function, can cause changes in serum uric acid levels. Therefore, elevated urea levels in GA may influence the severity of the disease and the frequency of attacks. High urea levels are associated with kidney function and hyperuricemia. Low albumin levels are linked to inflammation and urate accumulation. Urate accumulation reduces nitric oxide, leading to renal ischemia; increases pro-inflammatory cytokines, causing kidney damage; and impairs the kidney’s ability to excrete nitrogenous waste products. Glucose levels, on the other hand, contribute to kidney dysfunction by triggering metabolic disorders and inflammatory processes. Examining the results obtained from identifying factors influencing GA using ML models, it has been observed that variables such as urea and uric acid are among the factors affecting model performance. These findings indicate that the models align with real-world findings in determining the factors influencing GA.

Summary information on similar studies in the literature is given in

Table 13. Comparison of study findings with the literature.

References	Year	Dataset	Model	Metrics	Results
[28]	2023	2010–2019 collected dataset of 4.637 patients with type 2 diabetes and chronic kidney disease	Multivariable linear mixed-effects model	R2	0.74
[29]	2022	22991 observation data for 5186 patients	dlGFR	Accuracy	0.8830
[30]	2024	13 different cohorts of 19.629 patients	RF	MSE	294.24
Ours	2025	Data on 276 patients between 2018 and 2023	DT	Accuracy	0.9285

.

5. Conclusions

In this study, we investigated the application of machine learning (ML) techniques to identify key factors influencing the progression of gouty arthritis (GA) and its impact on renal function. By leveraging demographic, clinical, and laboratory data, we utilized decision tree (DT), random forest (RF), logistic regression (LR), and artificial neural network (ANN) models to classify GA and determine the most significant biomarkers associated with disease severity.

The evaluated ML models showed that they were able to effectively classify GA cases based on glomerular filtration rate (GFR) values, and showed effective performances in identifying important biomarkers affecting GA. Among the models, the decision tree (DT) model outperformed the others, achieving the highest accuracy (92.85%) and recall (100%), highlighting its effectiveness in detecting GA. In contrast, the random forest (RF) model showed a more balanced performance with an accuracy of 85.71%, indicating that it can generalize better to larger datasets. Logistic regression stood out with the highest specificity (93.33%), effectively reducing false positive classifications. On the other hand, the artificial neural network (ANN) model exhibited strong recall (92.31%) but the lowest precision (80.00%), indicating a tendency to misclassify cases without gouty arthritis as positive. Among the most influential factors, creatinine, urea, hemoglobin, and albumin levels were identified as key indicators of GA progression. These findings highlight the potential of ML-based predictive models in improving early diagnosis, risk assessment, and personalized treatment strategies for GA patients.

Despite the promising results, this study has certain limitations. The dataset size was relatively small, which may impact on the generalizability of the findings. Future research should focus on expanding the dataset to include a larger and more diverse patient population. Additionally, integrating more advanced deep learning techniques and longitudinal patient data could enhance predictive accuracy and provide deeper insights into disease progression.

In conclusion, this study underscores the potential of ML-driven approaches in medical research, particularly in the assessment of GA and renal function impairment. By incorporating data-driven methodologies, clinicians can refine diagnostic processes, improve patient outcomes, and develop more effective treatment strategies for individuals at risk of GA-related complications.