Data for this study were obtained from the NHANES dataset. As described by Zhao M et al, NHANES (2013–2016) is a cross-sectional survey program using complex random probability samples with a multi-stage stratified design conducted by the National Center for Health Statistics of the Centers for Disease Control and Prevention (15). The survey covers demographics, lifestyle, anthropometric, laboratory analysis, questionnaire interviews, and dietary data. The NHANES study plan was approved by the National Center for Health Statistics (NCHS) Ethics Review Board, Approval No. #2011-17 (16). Written informed consent was obtained from all participants for this study including the Home Interview Consent Form, Consent for Specimen Storage and Continuing Studies, and Consent for Examination at the Mobile Examination Center.

Because NHANES 2013 to 2016 has complete data on male and female sex steroid hormones, we chose these 2 cycles as the source of data for our study. We collected a total of 20,146 male and female samples for the initial survey. Inclusion criteria were (1) age ≥20 years (2); participation in dietary interviews and sex steroid hormone testing; and (3) full participant gout information. Exclusion criteria were (1) age < 20 years (2); missing self-report of gout (3); pregnancy in female participants; and (4) missing information on SII, DA, sex steroid hormones, and other important variables. Finally, 8550 participants were included in the study, including 4160 males and 4390 females.

The 2013-2016 NHANES data that we initially included in the study was an all-age population, and the age limit for those who conducted the Interview Gout Questionnaire was 20 years of age and older; therefore, we needed to exclude those under 20 years of age. We used self-reported personal interview data from the MCQ160N questionnaire ( Supplementary Material 1 ) to categorize whether study participants had gout. Participants were asked, “Has a doctor ever told you that you had gout?” Choice code 1 indicated gout, code 2 indicated no gout, code 7 indicated refusal to answer, and code 9 indicated not sure.

Serum specimens were separated after venous blood collection, stored at -30°C, and shipped to the Centers for Disease Control and Prevention for uniform testing by the Laboratory Sciences Division of the National Center for Environmental Health. Serum total testosterone (TT) levels and estradiol (E2) concentrations were determined by isotope dilution liquid chromatography-tandem mass spectrometry (ID-LC-MS/MS). The sex hormone-binding globulin (SHBG) assay is based on chemiluminescence measurements, and readings are compared to instrument- and lot-specific calibration curves. Detailed analytical methods are described in the NHANES Laboratory Procedures Manual. The lower detection limits for TT, E2, and SHBG were 0.75 ng/dL, 2.994 pg/mL, and 0.800 nmol/L, respectively.

The demographic profiles we included in the study included gender, age, race, educational attainment, and family poverty income ratios (PIR). Tobacco use, alcohol consumption, hypertension, and diabetes were determined by an interview questionnaire, with at least 100 cigarettes smoked in a lifetime as definitive tobacco use. Total dietary calories and total lycopene intake data for 2 days were obtained through nutrition interviews. They got BMI and waist measurement data from body measurements. Other laboratory tests include triglycerides (TG), hemoglobin (Hb), UA, high-density lipoprotein (HDL), low-density lipoprotein (LDL), apolipoprotein B (ApoB), glycosylated hemoglobin (GHB) and 1, 25-hydroxyvitamin D3 (VitD).

According to the method of Wei C et al. (10), the calculation formula of SII is platelet count multiplied by neutrophil count divided by lymphocyte count. Dietary data were obtained through two dietary recall interviews. The NHANES-assisted dietary interview system calculated the energy and nutrient intakes obtained from the number and type of food, beverage, and water within 24 hours before the interview. The six antioxidants we extracted included vitamin A (VitA), vitamin C (VitC), vitamin E (VitE), zinc, selenium (Se), and lutein + zeaxanthin (LZ). DA in this study mainly refers to six antioxidants and lycopene. The CDAI was calculated by subtracting the mean value from each of the six antioxidants and dividing it by the standard deviation to standardize and add them together (17).

To avoid overfitting, the participant data used for ML development were subjected to principal component analysis (PCA), which ultimately consisted of 29 features and 1 label, of which 4 features were categorical variables and the remaining 25 features were continuous variables. Variables with missing values greater than 20% were directly deleted, and variables with missing values less than 20% were interpolated for missing values using the ML-based random forest (RF) algorithm (18). You J et al. found that RF interpolation significantly outperformed other interpolation methods in terms of both average prediction error and consistency correlation coefficient after comparing multiple missing value interpolation methods (19). We randomly split the data into the 80% training set (male group n=3328, female group n=3512) and the 20% test set (male group n=832, female group n=878), subjected the training set to Synthetic Minority Over-Sampling Technique (SMOTE) to deal with data imbalance, normalized the feature data using a MinMaxScaler, and represented the categorical variables using one-hot coding (20, 21).

We used six ML models, support vector machine (SVM), Light Gradient Boosting Machine (LGBM), RF, Gradient Boosting Decision Trees (GBDT), Extreme Gradient Boosting (XGBoost), and Category Boosting (CatBoost), to identify gout associated with SII, sex steroid hormones, and DA. To optimize the performance of each ML algorithmic model, we tuned the hyperparameters using a grid search (22). We adopt the method of Cao Y et al. using 5-fold cross-validation, where the training set data is divided into 5 subsets, and each model is trained on 4 of the partitioned data and validated on the remaining partitioned data to calculate the average performance metrics (23). Different ML algorithms have their adaptations to the data and analysis, so we used the training set to develop six ML models and to verify the reliability, accuracy, and robustness of the models we used the test set (24). Important reference metrics used to evaluate the performance and generalization ability of ML models include the area under the ROC curve (AUC), accuracy, precision, recall, F1 score, brier score, and the area under the P-R curve (AP), as described by Goodswen SJ et al. (25). After summarizing seven metrics representing model performance and generalization ability for each ML model, the overall best algorithmic model for identifying gout was selected and the model was interpreted using SHAP. SHAP is one of the most commonly used post hoc interpretability tools. It enhances the user’s trust in the model by calculating the marginal contribution of each feature to the model’s output, interpreting the model both globally and locally, and giving the process of prediction and decision-making provided by the model (26).

In this study, the demographic characteristics of study participants with and without gout were described and grouped by gender. We used the Wilcoxon rank sum test for complex survey samples to compare differences between two groups for continuous variables and the Rao & Scott second-order corrected chi-square test to compare differences between groups for categorical variables. The Kruskal-Wallis test was used for comparison of data between quartile groups. Depending on the distribution of the data, continuous variables are expressed as mean ± standard deviation or median and quartiles (25%, 75%), while categorical variables are expressed as frequencies (percentages). ML model performance was discriminated using the AUC, accuracy, precision, recall, F1 score, brier score, and the AP for each model. All data were analyzed using R software version 4.3.1 and Python software version 3.11.5, respectively. The R packages used are gtsummary, survey, haven, tableone, plyr, dplyr, tidyverse, arsenal, missForest. ML was primarily analyzed using the scikit-learn 1.2.2 library. P-values less than 0.05 were considered statistically different.

Figure 1 illustrates the baseline process of participant selection and study design. A total of 8550 participants were included in the study, and the demographic characteristics of participants diagnosed with and without gout are summarized in Table 1 . Participants were 48.65% (n=4160) male and 51.35% (n=4390) female, with an overall gout prevalence of 4.32% (n=369), with 70.19% male and 29.81% female, which is a significant gender difference (P<0.05). In the overall population gout patients were more likely to be male, older, have higher TT levels, lower E2 levels, lower SHBG levels, higher TG levels, higher UA levels, lower HDL levels, higher GHB levels, diabetes, Non-Hispanic White, alcohol consumption, tobacco use, obesity, and hypertension (P<0.05).

We categorized males and females into Q1-Q4 groups based on quartiles of UA levels, respectively. Table 2 shows that among the male participants, the differences between the Q2, Q3, and Q4 groups compared to the Q1 group were statistically significant for all variables except SII, Lycopene, zinc, Se, and LZ (P<0.05). Table 3 shows that among the female participants, the differences between the Q2, Q3, and Q4 groups compared to the Q1 group were statistically significant for all variables except SII, lycopene, VitA, Se, and LZ (P<0.05).

Table 4 shows a summary of the model performance metrics of the six ML models for identifying gout in different gender groups based on the test set. In the male group, XGBoost showed optimal AUC performance with an AUC (95%CI) of 0.795 (0.746-0.843) based on the test set data ( Figure 2 ). The XGBoost model demonstrated superior discrimination ability, achieving an accuracy of 98.7%, precision of 100%, recall of 82.8%, F1 score of 0.906, brier score of 0.067, and AP of 0.238 based on the test set data ( Figure 3 ). The XGBoost algorithm demonstrated superior performance in terms of AUC (95%CI) with a value of 0.822 (0.754-0.883) based on the test set data, specifically within the female group ( Figure 4 ). The XGBoost model also demonstrated superior discrimination ability, achieving an accuracy of 99.2%, precision of 100%, recall of 76.7%, F1 score of 0.868, brier score of 0.035, and AP of 0.207 based on the test set data ( Figure 5 ).

The SHAP plot illustrates the impact of all features in the XGBoost model on identifying gout within the test dataset. The importance feature by SHAP showed that UA was the most important feature for the identification of gout by the XGBoost model in both male and female groups. Using the SHAP summary plot, we found that in the male group, the lower the feature values of LZ, VitC, lycopene, zinc, TT, VitE, and VitA, the greater the positive impact on the model output, while the lower the feature values of UA, Se, SHBG, SII, and E2, the greater the negative impact on the model output ( Figure 6 ). In the female group, SHAP values showed that lower feature values of E2, Zinc, Lycopene, LZ, TT, and Se had a greater positive effect on model output. In comparison, lower feature values of VitA, SHBG, and SII had a greater negative effect on model output ( Figure 7 ).

We plotted decision plots revealing the process by which individuals identify gout from a complex XGBoost model. The gray vertical line passing through zero in the decision plot is the baseline of the XGBoost model. The red fold in the decision plot indicates that the individual identification result in the XGBoost model is gout, and the output value for each feature is labeled, along with an indication of whether or not that output value is higher than the average output value. Figure 8A illustrates the individual decision-making process of the XGBoost model for the male group. Figure 8B shows the individual decision-making process of the XGBoost model for the female group.

From the decision plot we can observe that starting from the bottom of the plot, the decision curve shows the final prediction results of the SHAP values in the XGBoost model accumulating from the base value at the bottom to the top, with the final identification boiled down to 0 for no gout prediction and 1 for gout prediction. The decision plot shows the results of the top 30 individual identifications from the XGBoost model, with the red folds being gout predictions and the purple folds being no gout predictions. Figure 9A illustrates the individual decision-making process for the top 30 of the XGBoost model for the male group. Figure 9B shows the individual decision-making process for the top 30 of the XGBoost model for the female group.