Predictors of Inadequate Serum Urate Response to Low-Dose Febuxostat in Male Patients with Gout

Introduction

Febuxostat is an effective urate-lowering therapy (ULT) agent.^1,2 Despite titration therapies following the treat-to-target strategy were widely adopted in the management of gout,^3–5 ~20-40% of patients do not achieve a target serum urate (sUA) level of <6 mg/dL taking febuxostat even at the highest daily dose.^6,7 In previous studies, some demographic or clinical variables emerged as predictive factors for serum urate inadequate response (IR) under ULT, such as low adherence, inadequate dosage, higher baseline sUA, and longer duration of gout^8,9 in clinical practice. However, the factors and underlying mechanisms specifically associated with the therapeutic response to febuxostat have not been well elucidated.

It was held that a “start low go slow” target strategy is recommended after initiating ULT,² as acute gout flares occur more often in the setting of a rapid and intensive reduction of the serum urate level.^10,11 However, such an approach embedded in the guidelines requires frequent monitoring of sUA and continuous supply of ULT.¹² To date, data on the optimal titration regimens are lacking. Exploratory clinical trials have recently described the efficacy of low-dose febuxostat over a mid-term period and reported that the rate of achieving the target (sUA < 6 mg/dL) was 32–45.7%.^13–15 A proportion of patients required a lower dose of febuxostat during early ULT. In contrast, continuous administration of low-dose febuxostat (10/20 mg/day) has been widely adopted after ULT initiation in clinical situations in Asia (eg China, Japan)^16,17 considering patient preferences and adherence. Dose adjustments of ULT medication occurred infrequently, with a mean time to any dose change of 211 days in the febuxostat group.¹⁸ Taken together, identifying who will benefit from low-dose febuxostat has the potential to facilitate dose choice, while reducing frequent blood draws and patient visits. Patients treated with ULT need more optimal and tailored titration ULT regimens according to patient preferences, the timing of ambulatory encounters, and other factors in real-world management of gout.² Based on individual participant data from available trials, this study aimed to investigate the predictors associated with inadequate serum urate response to low-dose febuxostat during early ULT in pooled patients with gout.

Methods

Patients

We included participants from a prospective observational cohort in which gout patients were treated with 20 mg febuxostat and visited every 4 weeks to determine the predictive ability of CA-724 on gout flares (#ChiCTR2100043573)¹⁹ and two RCT trials: a trial evaluating the efficacy and safety of chitosan oligosaccharide treatment (#ChiCTR2100042424), and a trial comparing the efficacy and safety of low-dose febuxostat and low-dose benzbromarone in gout patients (#ChiCTR1800019352). All participants took 20 mg of febuxostat without any dosage adjustments during 12 weeks. Overall, the inclusion criteria were: male patients with gout, age ≥ 18 years old, received at least 12 weeks of 20 mg febuxostat treatment, had available demographic, clinical, and follow-up data. Key exclusion criteria were estimated glomerular filtration<45 mL/min/1.73m², transaminases >2-fold of the upper normal limit, baseline sUA <7 mg/dL, and suffering from rheumatoid arthritis, or other serious diseases. Participants were defined as inadequate responders if they did not reach the target sUA (<6 mg/dL) at any time point during the short-term (4 weeks) or mid-term (12 weeks) follow-up. The study design is illustrated in Figure 1.

Figure 1 Study flow diagram. CA-724 cohort, cohort evaluating the predictive ability of CA-724 on gout flares (#ChiCTR2100043573); chitosan cohort, cohort evaluating the efficacy and safety of chitosan oligosaccharide treatment (#ChiCTR2100042424); febuxostat cohort, cohort comparing the efficacy and safety of low-dose febuxostat and low-dose benzbromarone in patients with gout (#ChiCTR1800019352). Abbreviation: LR, logistic regression; RF, random forest; SGD, stochastic gradient descent classifier; SVC, support vector classifier; XGBoost, extreme gradient boosting.

Variables

Demographic, clinical, biological, and radiological data (Table 1) were collected for each cohort at baseline. Tophi were assessed by clinical assessment. If the laboratories drawn varied between the discovery cohorts, only the common variables were reserved. Finally, 218 samples were selected as the discovery dataset with relatively complete 29 variables. To deal with missing data, k-nearest-neighbors-based imputation (KNN) was used on the discovery set when the missing value was <20%. In total, the 218 discovery samples were randomly divided into training and internal test sets (5:1 ratio). A variable selection process, least absolute shrinkage and selection operator (LASSO) regression according to the minimum criteria (lambda.min), was applied to all 29 available variables to select the most predictive features for the models, augmented with 10-fold cross-validation of the training dataset for internal validation. The R package ‘glmnet’ and ‘impute’ were used to implement the LASSO regression and the KNN imputation methods.

Table 1 Characteristics of Participants in Each Cohort

Models

Five machine-learning models were evaluated: logistic regression, random forest, stochastic gradient descent (SGD), extreme gradient boosting (XGBoost), and linear support vector classifier (SVC). The Python library Scikit-Learn was used to implement the machine learning models.

First, cross-validation was performed during the modeling process to avoid overfitting and to allow for an unbiased model. An external test dataset from another trial (ChiCTR2000034138) was then used to evaluate the best-trained model output. The area under the receiver operating curve (AUC) was used to evaluate model performance. The positive and negative predictive values, sensitivity, and specificity were also computed, allowing model evaluation. Calibration of the prediction model was evaluated using calibration plots of observed and predicted risks and goodness-of-fit tests, with the number of boxes used to calculate quantiles set to 10. The larger the P value, the better the calibration of the prediction model. The R package ‘rms’ was used to plot calibration curves and calculate P-value statistics. To estimate the clinical utility of the prediction model, a decision curve analysis was employed using the R package “rmda” with the aim of calculating the net benefit of the threshold probability range in the training set and the test set. The threshold probability of decision curve analysis is that the expected return of the prediction is equal to the expected return to avoid the prediction. Decision curve analysis, a trade-off between false positives and false negatives, is primarily used to measure medical intervention strategies, screen beneficiaries, and evaluate the practical value of the entire model. To show the correlation between the model features, the correlation analysis between the features is visualized using the R package “circlize”.

We performed SHAP (Shapley Additive exPlanations) on the training dataset to describe the weight of the given variables responsible for the model output. The SHAP value was used to quantify the proportion of positive and negative influences of the features on the prediction outcome for each patient, displaying an explanation diagram. The Python library SHAP and matplotlib were used for this analysis.

Statistical Analysis

For continuous variables, the t-test or Kruskal–Wallis test was performed as appropriate. Binary variables were compared between the two groups using the chi-squared test. A two-tailed p-value <0.05 was considered to indicate statistical significance. For all cohorts (340 patients), multilevel mixed-effect models were used to compute odds ratios (OR), with the study level as a random effect. Statistical analyses were performed using Stata Version 18.0.

Ethics Approval

The protocol of this study was approved by the Ethics Committee of the Affiliated Hospital of Qingdao University and conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent. The scientific committees of the four cohorts approved the use of the data for this study.

Results

Study Cohorts

A total of 340 participants were pooled to investigate febuxostat responses. At week 4, 68.9% of the participants were non-responders, reporting sUA>6 mg/dL, and 51.8% were non-responders during the mid-term (12 week) follow-up. Of the 237 non-responders at short-term follow-up, 54 became responders during the mid-term follow-up period. Baseline characteristics of the patients with gout are summarized in Table 1.

Potential Factors Associated with IR

Age, BMI, sUA, TG, and Ccr during the 4-week follow-up period were significantly associated with febuxostat IR in the pooled population. With multivariate analysis, sUA was the only factor (OR 1.012, 95% CI 1.006–1.017). sUA and TG levels were both significant predictors during the mid-term follow-up (OR 1.017, 95% CI 1.012–1.023 and OR 1.439, 95% CI 1.061–1.951) (Table 2).

Table 2 Factors Associated with Not Achieving the Serum Urate Target During Short- and Mid-Term Follow-Up

There was no difference in the febuxostat response during the short- or mid-term follow-up according to disease duration, combined with hypertension or nephrolithiasis, radiographic severity, or signs of inflammation (CRP and ESR) (Table 2).

Variable Selection and Model Performance

In the discovery dataset, 218 patients receiving 20 mg of febuxostat daily for 12 weeks were included, of which 17% were selected as the internal validation set. A total of 138 patients out of 218 (63.3%) were identified as having febuxostat IR. Low-dose febuxostat inadequate responders were younger (43 vs 50 years, p<0.001) and had higher body mass index (BMI) (27.4 vs 26.4 kg/m², p=0.012), higher baseline sUA (585.4 vs 513.0 μmol/L, p<0.001), higher triglyceride (TG) (2.4 vs 1.6 mmol/L, p<0.001), lower high-density lipoprotein cholesterol (1.2 vs 1.3 mmol/L, p < 0.001), higher C-reactive protein (CRP) level (3.4 vs 1.8 mg/L, p=0.008), and a larger count of lymphocytes and white blood cells compared with those responders at baseline (Supplementary Table S1). Eighty patients were included in the external test dataset, with 37 (46.3%) identified as having febuxostat IR.

Five of the 29 variables available were selected as highly predictive combinations by LASSO: age, BMI, TG, sUA, and CRP levels (Figure 2A, B). Five machine learning models were then employed to predict the febuxostat therapeutic IR: random forest classifier, logistic regression, linear SVC, SGD Classifier and XGBoost. The performance of each model and their comparison with an external test dataset are shown in Figure 2C. The SGD Classifier performed best on the internal test dataset and achieved an AUC of 0.873 [95% CI: 0.763–0.942)]. The results replicated well in the external test set, with an AUC of 0.706 (95% CI [0.636, 0.727]) (Figure 3A).

Figure 2 Variable selection and model performance. The area under the receiver operating characteristic curve (AUC) score of the 29 variables was determined using the LASSO model in the training set (A). Red dots represent the AUC score, gray lines represent the standard error, and vertical dotted lines represent the optimal values based on the minimum criteria. The upper abscissa is the number of nonzero coefficients in the model at this time, and the lower abscissa is log lambda, which is the tuning parameter used for the 10-fold cross-validation in the LASSO model. LASSO coefficient profiles of the 29 variables (B). The performance of each model and comparisons of the external test dataset (C). AUC: the area under the receiver operating characteristic curve. Abbreviation: SGD, stochastic gradient descent classifier; SVC, support vector classifier; XGB classifier, extreme gradient boosting classifier.

Figure 3 Model performance and evaluation. The area under the receiver-operator characteristic curve on the internal validation and external test sets of stochastic gradient descent (SGD) classifier (A), the calibration plot of the SGD model in the test set (B), the decision curve (C), and the correlation analysis between the features (D). Abbreviation: AUC, the area under the receiver operating characteristic curve; sUA, serum urate; TG, triglyceride; BMI, body mass index; CRP, C-reactive protein.

The accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of the external test dataset are shown in Supplementary Table S2. Given that the SGD model had a higher AUC than the other models, it was further evaluated as the best model. The SGD model identified 70.3% (sensitivity, 95% CI: 65.9%, 88.2%) of inadequate responders in the test cohorts. The calibration curve of the SGD model was plotted in the testing set using five features (Figure 3B), with Hosmer-Lemeshow P=0.996. In addition, the decision curve results (Figure 3C) showed that the selected “age+sUA+BMI+CRP+TG” model exhibited a higher net benefit than the other models, as well as “All” (gray line) and “None” (black line). Patients would benefit more from the prediction of our model than from the no-prediction schemes for most ranges. The correlation analysis between the features is shown in Figure 3D. Serum urate was positively correlated with TG (r=0.280) and negatively correlated with age (r=−0.217). These five features are not redundant to each other.

Explanation of Model

SHAP values were computed for the five variables selected by LASSO to investigate their impact on the IR prediction model. The magnitude of the importance was similar for the training/test and external test sets. When displaying the importance of the variables in the training dataset, the most influential variable was age, followed by sUA, BMI, CRP, and TG levels (Figure 4A).

Figure 4 Febuxostat inadequate response interpreted by the trained stochastic gradient descent model with Shapley additive explanations (SHAP). The global importance of five features based on the average SHAP value magnitude (A). A set of bee-swarm plots corresponding to feature summary (B). The vertical axis shows the sorted five features while the horizontal axis displays the impact on the model output. Each data point represents a predicted output and the color indicates the feature values. Abbreviation: sUA, serum urate; TG, triglyceride; BMI, body mass index; CRP, C-reactive protein; SHAP, Shapley Additive exPlanations.

Age exerted the most negative effect on IR prediction (Figure 4B). When plotted individually (Supplementary Figure S1), younger age was associated with inadequate response and older age was associated with response. Patients with a higher BMI and higher sUA, CRP, or TG levels were less likely to reach the serum urate goal, particularly for values above normal. Furthermore, the interaction analysis among the five variables showed that they contributed to the predictive model in a linear fashion (Supplementary Figure S2). For example, older age corresponded to a lower value of sUA (red dots), which interacted most with sUA.

Discussion

In the present study, we demonstrated that younger age, higher BMI, and higher baseline sUA, triglyceride, CRP levels were associated with failure to achieve the sUA target in patients with gout receiving low-dose febuxostat during the early ULT period. Based on machine learning algorithms using these variables, we could robustly predict inadequate responders to low-dose febuxostat. These findings may have clinical applicability, allowing for the identification of patients who are likely to require low-dose febuxostat treatment to achieve a serum urate target.

The promoted sUA target appeared realistic for almost half of the patients, followed by a low-dose strategy in both the short- and mid-term follow-up timeframes. Although a higher dose was recommended by the guidelines and the target rate was increased, the low dose was prone to be maintained in usual daily care. Moreover, excessive medication burden is linked to poor ULT adherence in patients with gout,²⁰ with consequent undertreatment.^21,22 It is likely that participants who respond to low-dose febuxostat will benefit from a personalized febuxostat dose regimen, allowing patients to receive the right dose without underdosing or using excessively high doses of medication.

It has been shown that age, BMI, sUA, TG, and CRP levels might be early predictors of which patients could benefit from low-dose febuxostat therapy. Inadequate responders were younger than responders, which is in line with the results of a real-world study.²³ One possible interpretation is that the included older patients may have received persistent ULT treatment for a period prior to enrollment. Additionally, younger age was a predictor of poor adherence to ULT. Low-dose febuxostat is more suitable for older patients with a high prevalence of comorbidities.²⁴ Baseline BMI and sUA levels have been consistently associated with ULT responses in clinical trials and real-world studies,^14,23,25 which is consistent with our results. A higher serum urate level reflects a higher total urate burden, thereby requiring a higher dose to achieve the target serum urate.²⁶

Multivariate analysis and the present model highlight the potential role of TG as a predictor of febuxostat IR. The association between TG and gout has been well established, however, the effect of TG on patient responses to febuxostat has not been determined. It was recently reported that febuxostat could regulate free fatty acid and lipid metabolism,^27,28 indicating that hypertriglyceridemia may reflect the state of xanthine oxidase activity in the early phase of ULT. Notably, both BMI and TG levels were included in the model. It is possible that a high TG level could attenuate the urate-lowering effect of low-dose febuxostat or that a high TG level reflects more severe metabolic syndrome. Additionally, inflammatory status, as assessed by CRP levels, is a predictor of IR to low-dose febuxostat. The other imaging features and blood parameters did not provide any additional information. Further studies are required to identify the underlying mechanisms, as well as to explore how these predictors perform in clinical settings.

The ML model in the present study demonstrated reliable prediction of IR to low-dose febuxostat in the development set, with satisfactory replication in an independent test cohort. This finding is consistent with the predictors selected in the univariate and multivariate analyses. This is the first assessment using the ML approach to predict IR to febuxostat, with an AUC comparable to that of other studies such as methotrexate.²⁹ Our model supports that patients were appropriate for febuxostat dose escalation with a confidence of 61.9% (PPV), whereas others tended to benefit from maintaining a low-dose regimen. This model has several potential benefits and can be aligned with the individualization of ULT titration recommended in the 2020 American College of Rheumatology gout management guidelines.² At present, guidelines recommend febuxostat dose up-titration at intervals of 2–4 weeks.³⁰ Such a prediction model allows the identification of patients who require more frequent blood test monitoring. Recently, Qi et al assessed the proportion of achieved the serum urate target among hyperuricemia subtypes in another cohort receiving dose-escalation febuxostat, concluding that combined subtype has a lower response to febuxostat, compared to those with either overload or underexcretion subtype.³¹ Overall, combined subtype hyperuricemia (OR = 0.64, 95% CI 0.41–0.99, P = 0.048) and baseline serum urate (OR = 0.74, 95% CI 0.62–0.89, P = 0.001) were independently associated with lower rates of achieving SU target among six variables including age, disease duration, BMI, baseline serum urate, TG/HDL-C and subtypes. In Qi et al, the important finding was prescribing febuxostat according to hyperuricemia subtype, while it is available for clinicians to tailor individualized febuxostat treatment according to our findings.

The limitations of this study include patients with severe chronic kidney disease and low baseline serum urate levels being excluded from rigorous clinical trials, which limits the generalizability of the potential predictors. Additionally, a general limitation of this study is that the basic mechanism(s) of IR to febuxostat are not known. Although febuxostat inhibits the urate excretory transporter ABCG2 at clinically relevant concentrations,³² previous studies have not shown that the ABCG2 genotype (rs2231142 allele) alters the febuxostat response.³³ Febuxostat can impede the access channel to the xanthine oxidase/dehydrogenase active site, and febuxostat disposition is primarily mediated by the activity of hepatic UDP-glucuronosyltransferase and Cytochrome P450 (CYP) enzymes.³⁴ Importantly, it is not yet known whether individualizing febuxostat doses according to selected predictors will improve clinical outcomes. Although the quality of evidence for the efficacy of low-dose febuxostat is relatively limited, the benefits of low-dose febuxostat may include fewer gout flares in the first few months after treatment initiation.⁶

In conclusion, participants who were administered low-dose febuxostat may respond according to baseline sUA but also TG, CRP, age, and BMI at the mid-term follow-up. These findings suggested a stratified treatment regimen for patients with gout. The robust ML prediction model for inadequate serum urate response may allow the identification of patients who are likely to require dose escalation of febuxostat to achieve the serum urate target.