1 Introduction

Osteoarthritis (OA) is a degenerative joint disease highly prevalent among middle-aged and older adults.¹ Its pathological features primarily include articular cartilage degeneration, osteophyte formation, synovial inflammation, and subchondral bone sclerosis.^2,3 Clinically, patients often present with joint pain, stiffness, and functional limitations,⁴ which severely compromise their quality of life. Recent data indicate that OA affects approximately 600 million people globally.⁵ With accelerating population aging and changes in lifestyle, the epidemiological burden of OA is projected to increase, with an estimated 1 billion individuals affected by 2050.^6,7 However, there is currently no known fully effective cure, making OA one of the leading causes of disability among middle-aged and older adults.⁷ Furthermore, economic expenditures related to OA treatment, as well as unemployment and reduced productivity due to OA, directly or indirectly exacerbate the socioeconomic burden.⁶ Therefore, prevention and timely intervention are crucial for controlling OA.

Currently, imaging techniques such as X-ray and MRI remain the primary basis for the clinical diagnosis of OA. However, these methods rely on the detection of structural changes and exhibit markedly insufficient sensitivity in the early stages of the disease.⁸ OA is not limited to the lesion of articular cartilage, but is a complex pathological process that is progressively advanced by the interaction of multiple factors, such as genetic factors, mechanical injuries, inflammatory responses, and metabolic abnormalities.⁹ Among these, chronic inflammation and metabolic dysregulation are widely regarded as core pathological drivers of OA progression.^9,10 In recent years, increasing research has focused on identifying biomarkers that reflect the pathophysiological state of OA, aiming to uncover underlying disease mechanisms, predict progression, and thereby offer new directions for identifying high-risk populations and enabling early intervention.

The Systemic Inflammation Response Index (SIRI) and Systemic Immune-Inflammation Index (SII) are systemic inflammatory markers based on peripheral blood cell counts. Multiple large-scale cohort studies have demonstrated that elevated SIRI and SII levels are significantly positively associated with OA risk (OR > 1, P < 0.05), suggesting these indices may serve as auxiliary tools for early risk assessment.^11,12 Additionally, the role of metabolic disturbances in OA pathogenesis has gained increasing attention. The Weight Adjusted Waist Index (WWI) and Triglyceride Glucose Index (TyG) are important indicators for assessing metabolic imbalances related to metabolic syndrome. Studies have shown that for every 1-unit increase in WWI, OA risk rises by 65% (95% CI: 1.48–1.84, P < 0.001) ¹³; whereas a 1-unit increase in the TyG index is associated with a substantially elevated OA risk of 634% (OR = 7.34, 95% CI: 2.25–23.93, P < 0.001),¹³ further indicating that metabolic dysregulation may be a key driver of OA. However, existing markers often focus on a single pathological dimension, failing to provide a comprehensive assessment of OA risk. Moreover, their complex composition and inconvenient calculation limit their practical utility. Hence, exploring integrated biomarkers that are easily calculable and capable of combining key mechanisms of inflammation and nutritional-metabolic imbalance holds significant importance for optimizing OA risk stratification and advancing early intervention.

The red blood cell distribution width-to-serum albumin ratio (RAR) integrates systemic inflammatory status and nutritional-metabolic levels, systematically reflecting imbalances in the body's “inflammation-metabolism” axis and showing potential value in predicting disease progression. Specifically, red cell distribution width (RDW), an indicator of erythrocyte volume heterogeneity, can signal chronic inflammation, oxidative stress, or abnormal erythropoiesis¹⁴; while albumin is not only a sensitive marker of nutritional-metabolic status but also participates in disease processes by modulating inflammatory responses.¹⁵ Multiple studies indicate that RAR holds significant value in predicting the onset and prognosis of various chronic metabolic diseases, including metabolic syndrome, diabetes and its complications, and chronic kidney disease ^16–19. Furthermore, these conditions exhibit close pathological associations with osteoarthritis. Notably, previous research by Chen et al.²⁰ has confirmed a significant nonlinear positive correlation between RAR and OA (OR > 1, P < 0.05, non-linear test P < 0.001), with a critical inflection point at RAR = 3.03. This further suggests the potential application of this indicator in the early identification of OA. As a simple and cost-effective composite indicator, RAR may facilitate large-scale screening and monitoring of high-risk populations for OA, providing a reference for timely targeted interventions.

In light of the above, this study is the first to focus specifically on the middle-aged and older population, which has a high incidence of OA. It aims to thoroughly investigate the association between RAR and OA risk in this demographic and to preliminarily construct a risk prediction model based on RAR. The findings are expected to provide a theoretical foundation and methodological basis for the future development of clinical screening tools and the formulation of early risk-stratification strategies.

2 Materials and methods

The National Health and Nutrition Examination Survey (NHANES), conducted by the U.S. National Center for Health Statistics, is a nationally representative cross-sectional study that systematically collects demographic characteristics, dietary intake, physical and physiological examination data, laboratory test results, and health/nutrition questionnaire responses from approximately 5000 participants biennially. This comprehensive surveillance program provides critical data for assessing population health trends and disease patterns across different demographic groups. All participants provided informed consent approved by the Research Ethics Review Board of the National Center for Health Statistics, with complete study details available on the NHANES official website (https://www.cdc.gov/nchs/nhanes/).

In this study, we initially included 101,316 participants across 10 survey cycles (1999–2018). Subjects were rigorously screened based on the following exclusion criteria: (1) Participants with missing RAR or OA data; and (2) individuals lacking complete covariate information; (3) participants aged <40 years or pregnant individuals. Accordingly, 19,967 participants met eligibility criteria and were included in the final analysis (see Fig. 1 for detailed selection flowchart). Given NHANES's complex, stratified multistage probability sampling design, we incorporated the Mobile Examination Center (MEC) sampling weights in our analyses. Specifically, we applied a weighted approach where the 1999–2002 cycles (2 cycles) were assigned weights of (2/10) × WTMEC4YR, while the remaining eight cycles (2003–2018) were weighted as (1/10) × WTMEC2YR to ensure nationally representative estimates.

Close

Fig. 1 Participant selection flow chart for the NHANES 1999-2018.

Export to PPT

3 Results

3.4

3.4 The subgroup analysis and interaction test

The results of subgroup analyses and interaction tests across various demographic stratification variables are presented in Fig. 2. A consistently stable positive association between RAR and OA was observed across different subgroups. However, this relationship did not reach statistical significance (P > 0.05) in specific subgroups including Mexican Americans, individuals with education below high school level, those with BMI <25, borderline diabetes cases, current smokers, and participants reporting minimal or moderate physical activity levels. Furthermore, interaction analysis revealed that this positive association was not significantly modified by factors such as age, ethnicity, education level, BMI, diabetes status, smoking status, physical activity level, hypertension, or hyperlipidemia (all P for interaction>0.05). In summary, the described relationship between RAR and OA demonstrates considerable reliability and robustness across diverse population characteristics.

Close

Fig. 2 Subgroup analysis for the association between RAR and OA.

Export to PPT

3.5

3.5 Nonlinear relationships between RAR and OA

RCS analysis revealed a significant nonlinear relationship between RAR and OA in middle-aged and elderly populations in both unadjusted and fully adjusted models (all P values for non-linearity <0.05) (Fig. 3). Threshold effect analysis further identified an inflection point of 3.42 for RAR (Table 5). When RAR <3.42 (OR = 1.452, 95% CI: 1.449–1.456, P < 0.0001), each unit increase in RAR was associated with approximately a 45% higher risk of OA.

Close

Fig. 3 Four-knot restricted cubic spline relationship between RAR and OA: (A) unadjusted and (B) fully adjusted.

Export to PPT

Table 5 Threshold effect analysis of RAR on OA.

RAR	OR (95%Cl)	P-value
Fitting by standard linear model	1.150 (1.149-1.152)	<0.0001
Fitting by the two-piecewise linear model
Inflection point	3.42
RAR<3.42	1.452 (1.449-1.456)	<0.0001
RAR>3.42	0.907 (0.904-0.909)	<0.0001
Log likelihood ratio		<0.0001

3.6

3.6 Identification of key predictive variables for OA

Fig. 4 illustrates the selection of key variables for the OA prediction model using LASSO regression. Through 10-fold cross-validation and iterative analysis, the optimal penalty parameter (λmin = 0.0006) was determined. Building upon preliminary findings from univariate and multivariate logistic regression analyses, we identified significant predictors of OA. We included 11 variables in the predictive model: age, education level, race, sex, BMI, RAR, diabetes mellitus, physical activity, hypertension, hyperlipidemia, and smoking status.

Close

Fig. 4 Variable Selection in the OA Prediction Model Using LASSO Regression. (A) Trajectories of coefficient estimates across λ values. (B) Cross-validation error profile (optimal λ indicated by the red dot). Vertical axis: coefficient magnitude; top labels: number of retained variables; horizontal axis: log(λ).

Export to PPT

3.7

3.7 Assessment of the predictive performance of the OA prediction model

Fig. 5A displays the ROC curves of the RAR and the final predictive model, with their performance evaluated by the AUC. The predictive model achieved an AUC of 0.723 (95% CI: 0.714–0.733), with a sensitivity of 68.74% and specificity of 64.94% (Supplementary Table 4). In contrast, the RAR exhibited a significantly lower AUC of 0.545 (95% CI: 0.533–0.557). A formal comparison test confirmed a statistically significant divergence between the two models (P < 0.0001), indicating the superior discriminative power of the final model. Further assessment using DCA demonstrated that the model provided substantial net clinical benefit across a clinically relevant probability threshold range (Fig. 5B), reinforcing its practical utility.

Close

Fig. 5 Performance Evaluation of the Predictive Model. (A).ROC Curves of the Predictive Model vs. RAR: Combined predictive model in Red curve and RAR model in Blue curve. The predictive model achieved an AUC of 0.723 (95% CI: 0.714–0.733), with a sensitivity of 68.7% and specificity of 64.9% at the optimal cutoff. (B). DCA: The red curve denotes the net benefit of the prediction model, whereas the black line represents the baseline scenario without model application.

Export to PPT

3.8

3.8 Development of a clinical nomogram for OA risk prediction

To enhance the clinical utility of our predictive model, we constructed a visual nomogram incorporating all final predictor variables, providing physicians with an intuitive and efficient tool for quantifying individual OA risk (Fig. 6). This nomogram allows clinicians to simply sum the standardized point values assigned to each predictor in the model, where higher scores reflect greater OA likelihood. The total score was then mapped directly to an individualized probability, enabling instant risk stratification. By transforming regression coefficients into clinically interpretable point contributions, the nomogram preserves statistical accuracy while offering immediate, actionable risk assessment during patient evaluations.

Close

Fig. 6 Nomogram for predicting the risk of osteoarthritis. The total points obtained from the top score axis correspond to the predicted probability of osteoarthritis. Variable codings are as follows: Sex (0 = Male, 1 = Female); Race (0 = Mexican American, 1 = Non-Hispanic White, 2 = Non-Hispanic Black, 3 = Other races); Education Level (0 = Below high school, 1 = High school or equivalent, 2 = College graduate or above); BMI (0 = <25, 1 = 25-30, 2 = >30); Diabetes (0 = No, 1 = Borderline, 2 = Yes); Physical Activity (0 = Hardly, 1 = Mild, 2 = Moderate, 3 = Vigorous); Smoking Status (0 = Never, 1 = Former, 2 = Current); Hypertension (0 = No, 1 = Yes); Hyperlipidemia (0 = No, 2 = Yes).

Export to PPT

3.9

3.9 Identification of key predictive variables

To identify the most significant predictors, we applied the XGBoost algorithm to evaluate variable importance across all fully adjusted covariates in Model 3. As illustrated in Fig. 7A, the key variables were ranked in descending order: age, sex, race, RAR, education level, BMI, hypertension, smoking status, hyperlipidemia, physical activity, and diabetes mellitus.

Close

Fig. 7 Identification of Key Feature Variables in the Machine Learning Model. (A) Relative importance of each predictor in determining OA incidence was quantified using the XGBoost algorithm, with corresponding importance scores assigned. The Y-axis displays the selected variables, while the X-axis represents their importance scores. (B) Feature selection was performed via the Boruta algorithm, ranking predictors by importance. A variable was deemed statistically significant if its mean importance Z-score exceeded the maximum Z-score of shadow variables (green shaded area); otherwise, it was excluded from the final model.

Export to PPT

Further validation was conducted using the Boruta algorithm, performing 500 iterations to compute stability-based Z-scores, as shown in Fig. 7B. The ranking slightly differed: age, sex, race, education level, hypertension, RAR, BMI, diabetes mellitus, hyperlipidemia, smoking status, and physical activity. Despite minor discrepancies in feature importance between methods, we retained all variables in subsequent machine learning models to ensure comprehensive predictive performance.

3.10

3.10 Comparative analysis of five machine learning models for predictive performance assessment

The included data were divided into a validation set (30%) and a training set (70%). The basic information of subjects in both sets is summarized in Supplementary Table 5. ROC curves and DCA curves were plotted for the test sets of five machine learning models, and the AUC values of ROC curves were calculated to evaluate predictive capability and clinical applicability (Fig. 8). The AUC values for RF, Logistic, SVM, Decision Tree, and NB in the training set were 0.839, 0.815, 0.806, 0.796, and 0.712, respectively. In the validation set, their AUC values were 0.836, 0.810, 0.800, 0.795, and 0.714. To comprehensively evaluate predictive performance, this study further calculated Accuracy, Prevalence, Recall, F1-Score, MCC, Precision, Specificity, and FNR for each model on both training and validation sets (Supplementary Table 6). Results indicate that the RF model consistently demonstrates superior performance across most metrics. Additionally, DCA curves reveal that the RF-based model achieves higher net benefit than either the “all interventions” or “no intervention” strategies across the entire threshold range in both training and validation sets, suggesting broad clinical applicability. Five-fold cross-validation results further confirmed the robustness of the Random Forest model (Fig. 8), with fold-wise AUC values ranging from 0.828 to 0.838 and averaging approximately 0.832 (Supplementary Table 7). Furthermore, the model demonstrated relatively balanced performance across metrics including accuracy (0.762), recall (0.784), and F1 score (0.796), indicating a favorable equilibrium between accuracy and sensitivity in positive predictions, thereby possessing certain clinical reference value.

Close

Fig. 8 ROC and DCA Plots of Five Machine Learning Models. (A) ROC in the training set. (B) ROC in the validation set. (C)DCA in the training set. (D)DCA in the validation set. (E) ROC curve for five-fold cross-validation of Random Forest models. (F) DCA curves for five-fold cross-validation of Random Forest models.

Export to PPT

4 Discussion

Based on NHANES data from 1999 to 2018, this study investigated the association between the RAR and OA in middle-aged and older adults over 40 years of age. Following stringent screening criteria, a total of 19,967 participants were included, among whom 2997 were diagnosed with OA, yielding a prevalence rate of 15.01%. Multivariable logistic regression analysis demonstrated a significant positive association between RAR levels and the risk of OA, which remained consistent across unadjusted, partially adjusted, and fully adjusted models (all P < 0.0001). RCS analysis further revealed a significant nonlinear dose–response relationship between RAR and OA (P for trend <0.001, P for nonlinear <0.05), with an inflection point at RAR = 3.42. Subgroup analyses did not reveal significant interaction effects (P > 0.05), suggesting the robustness of this association across different population characteristics. Building upon these findings, a risk prediction model for OA based on RAR was developed and visualized as a clinically practical nomogram. Additionally, the RF model demonstrated consistently excellent predictive performance and clinical applicability in both the testing and validation sets, and its robustness was confirmed through five-fold cross validation. In summary, elevated RAR is an independent risk factor for OA and shows potential value for risk monitoring and early prediction of OA in middle-aged and older adults, which may aid in identifying high-risk individuals and guiding early interventions.

OA is a heterogeneous disease influenced by multiple factors, and its development is closely associated with age, sex, obesity, inflammation, and metabolic dysregulation.^7,9,31 The baseline analysis in this study further supports this perspective. Results showed that individuals aged 60 and above accounted for 71.30% of OA patients, significantly higher than those under 60 (28.70%). The proportion of females with OA was notably higher than that of males (66.33% vs. 33.67%), and the prevalence among obese individuals was significantly greater than among those with normal weight (45.61% vs. 32.90%). Notably, the proportion of OA patients with comorbid metabolic disorders such as diabetes, hypertension, and dyslipidemia was significantly higher than in those without these conditions (P < 0.001). Concurrently, the prevalence of OA was markedly higher among individuals with such metabolic disorders. Existing research indicates that metabolic imbalance diseases, including diabetes and metabolic syndrome, are closely linked to OA progression.^10,32 Chronic hyperglycemia and insulin resistance can trigger oxidative stress and systemic inflammatory responses, which in turn disrupt chondrocyte metabolism, impair cartilage matrix homeostasis, exacerbate synovial inflammation, and promote pathological changes in subchondral bone and microvasculature.³² These findings further suggest that both nutritional metabolic imbalance and inflammatory responses play significant roles in the pathogenesis of OA and may interact to contribute to disease initiation and progression. Additionally, univariate regression analysis revealed associations between OA and factors such as race, education, smoking, and physical activity levels, underscoring the multifactorial complexity of the disease etiology.

RAR is a novel biomarker that integrates systemic inflammatory status and nutritional metabolic state. Currently, this index has been applied in the risk assessment and prognosis prediction of various diseases, including diabetes and its complications, malignancies, and cardiovascular diseases, all of which share common pathological features of concurrent inflammation activation and metabolic dysregulation ^{17,18,33–35}. Cross-sectional studies suggest that higher RAR levels are associated with increased prevalence and poorer prognosis of diabetes.¹⁷ Furthermore, other research indicates that for each 1-unit increase in RAR, the risk of metabolic syndrome rises by 67%.¹⁶ This evidence suggests that RAR may help evaluate the pathophysiological state related to OA and serve as a potential predictive marker. The release of inflammatory mediators and oxidative stress are critical factors influencing OA progression.³⁶ RDW, an indicator of red blood cell volume heterogeneity, is modulated by both factors.¹⁶ On one hand, inflammatory mediators can interfere with erythrocyte homeostasis by suppressing erythropoietin activity or inducing iron metabolism disorders, leading to increased RDW.^37,38 On the other hand, oxidative stress can increase erythrocyte membrane fragility, reduce deformability, damage erythrocyte structure and function, and cause oxidative damage, which similarly elevates RDW.³⁹ Therefore, elevated RDW levels can partly reflect the activity of inflammation and oxidative stress in OA patients. Albumin, synthesized by the liver, is the most abundant circulating protein in the body, and its level reflects protein metabolism status. Hypoalbuminemia often indicates abnormal protein metabolism.^40,41 In OA patients, dyslipidemia and persistent inflammatory responses may affect liver synthetic function via related inflammatory cytokines, leading to protein metabolism disorders and decreased albumin levels ^16,42–45. Moreover, albumin not only maintains plasma colloid osmotic pressure but also possesses important physiological functions such as anti-inflammatory and antioxidant activities.⁴⁶ Hypoalbuminemia weakens these protective effects. Concurrently, inflammatory states increase capillary permeability and promote albumin extravasation, creating a pathological cycle.^47,48 Thus, albumin levels can reflect both the nutritional metabolic status and the systemic inflammatory level in individuals with OA. Unlike indices such as SII, TyG, and WWI, which focus on a single pathological dimension, RAR incorporates both a systemic inflammation marker (RDW) and a nutritional metabolic indicator (albumin) into a unified assessment framework. This enables a more comprehensive reflection of the systemic imbalance in the “inflammation-metabolism axis” during the OA pathophysiological process, potentially enhancing the identification and stratification of high-risk populations. Furthermore, RAR relies solely on routine clinical laboratory parameters, offering advantages such as easy accessibility, simple calculation, and low cost, which support its practical application in primary care and large-scale population screening.

The study by Chen et al.²⁰ found that RAR levels were significantly higher in OA patients compared to non-OA individuals (P < 0.001), a result consistent with the findings of the present study. Furthermore, Chen et al.²⁰ also observed a significant nonlinear relationship between RAR and OA (P < 0.001) and noted that OA incidence increased with rising RAR levels, aligning with the trend identified in our research. It is worth noting that the inflection point identified by Chen et al.²⁰ was at RAR = 3.03, slightly lower than the value observed in this study (RAR = 3.42). This difference may be attributable to variations in the characteristics of the populations included in the two studies. Interestingly, our analysis further revealed that below the inflection point (RAR <3.42), each 1-unit increase in RAR was associated with a 45% increase in OA risk. Above the inflection point (RAR >3.42), OA risk slightly decreased and then plateaued with further increases in RAR, a pattern similar to that reported by Chen et al. This exploratory finding suggests that the complex association between RAR and OA risk may reflect two potential factors. On one hand, extremely high RAR levels may be accompanied by as-yet-undefined compensatory or protective physiological mechanisms. Prior studies have indicated that compensatory anti-inflammatory responses are often activated concurrently in systemic inflammatory states ^49–51. On the other hand, the observed attenuation in risk at higher RAR levels could be influenced by confounding effects, such as more aggressive therapeutic interventions in individuals with severe comorbidities or survivor bias within this subgroup. Given the inherent limitations of cross-sectional study designs, future prospective cohort studies and basic experiments incorporating additional biomarkers are needed to validate this nonlinear relationship and elucidate its underlying mechanisms. Based on current evidence, maintaining RAR below 3.42 may provide a useful reference for identifying OA risk, stratifying individuals, and guiding early interventions in middle-aged and older adults.

Building on these findings, we developed and validated an OA risk-prediction model incorporating RAR. The nomogram derived from the model indicated that, among all included predictors, RAR carried the second-highest contribution weight after age, exceeding that of conventional risk factors such as sex and BMI. This underscores the value of RAR as an integrative marker of systemic inflammation and metabolic dysregulation, offering pathophysiological insights that complement traditional demographic and anthropometric assessments and thereby improving overall risk stratification.

In terms of clinical translation, the model and the RAR index hold potential utility across multiple stages of care. In prevention, RAR could serve as a primary screening tool in community or metabolic clinics to identify high-risk individuals who may benefit from joint health education and early behavioral interventions. In diagnosis, for patients with typical symptoms but only early-stage radiographic changes (e.g., Kellgren-Lawrence grade I-II), an elevated RAR could provide auxiliary diagnostic support and suggest that disease activity may be closely linked to systemic metabolic-inflammatory processes, rather than mechanical wear alone. In management, RAR may also function as a dynamic monitoring indicator for treatment response and disease activity, with changes potentially reflecting the modulation of systemic inflammatory-metabolic status by interventions, thus offering an objective parameter beyond symptoms and imaging for the long-term management of OA.

However, it is important to note that the prediction model developed in this study (AUC = 0.723) is primarily intended as an auxiliary screening and risk-communication tool, not a replacement for clinical diagnosis. Its discriminative ability is comparable to, or even superior to, many established risk-prediction models for chronic diseases, making it suitable for preliminary risk stratification at the population level.^26,52,53 The outstanding value of this model lies in its cost-effective integration of traditional risk factors with a novel blood-based marker (RAR), offering a more comprehensive and mechanistically grounded practical framework for OA risk assessment.

However, the model constructed based on LASSO regression focuses primarily on clinical interpretability and ease of practical application. To further evaluate the predictive potential of RAR and validate its robustness as a key variable, we additionally employed five machine learning algorithms for modeling. Such methods are capable of capturing potential nonlinear relationships and complex interactions among variables, thereby helping to explore the upper limit of predictive performance.⁵⁴ Concurrently, we applied SHapley Additive exPlanations (SHAP) to identify the key predictive features in each model. Feature importance analysis demonstrated that RAR was consistently recognized as an important predictor across multiple machine learning models, further methodologically supporting its robust value in OA risk prediction. Results from receiver operating characteristic (ROC) curves and decision curve analysis (DCA) showed that the machine learning models exhibited good and stable discriminative ability and clinical applicability in both the training and test sets. Among them, the RF model achieved AUCs of 0.839 and 0.836 in the training and validation sets, respectively, and consistently outperformed the other four machine learning models across multiple performance metrics. Five-fold cross-validation also indicated that the RF model possesses strong robustness and generalization ability.

In summary, although machine learning models represented by random forest showed slightly better predictive performance than the LASSO-based model, the LASSO-derived nomogram, due to its intuitive and user-friendly nature, better meets the practical needs for interpretability and operational convenience in current clinical practice. Therefore, this study still recommends the LASSO-nomogram as the primary clinical auxiliary tool, while the machine learning analysis provides important methodological support for the robustness and generalizability of the key biomarker RAR and lays a preliminary foundation for the future development of models integrating more indicators and possessing dynamic prediction capabilities.

This study has several main limitations, which should be addressed in future research. First, as NHANES is a large-scale cross-sectional survey aimed at collecting broad population health data, standard joint imaging assessments were not performed for all participants. The diagnosis of OA relied on a validated self-reported questionnaire. Although the use of self-reported, physician-diagnosed data is a common and methodologically reasonable strategy in large observational studies—especially in the preliminary stages of exploring novel biomarkers—and multiple authoritative studies have confirmed its high concordance with clinical diagnoses ^22–25, this approach may still introduce recall bias and non-differential misclassification. Such misclassification is typically independent of the exposure (RAR) and, in epidemiological analysis, often biases effect estimates toward the null. This implies that the association observed in this study may be somewhat underestimated, which indirectly suggests the potential robustness of the current findings. Second, self-reported data cannot provide key clinical information such as the specific joints affected, radiographic grading (e.g., Kellgren-Lawrence grade), symptom severity, or structural changes in OA, limiting our ability to conduct an in-depth analysis of the association between disease phenotypes and RAR.

Third, constrained by the cross-sectional study design, this research cannot infer a causal relationship between RAR and OA. Finally, regarding predictive modeling, although we conducted a preliminary evaluation of the internal robustness of the model through five-fold cross-validation, the model still requires further refinement in the following aspects: First, the current model lacks external validation in independent population cohorts, and its true generalizability has not been fully tested. Second, feature engineering could be further optimized; integrating multimodal data in the future may enhance predictive performance.

Future research directions should include: validating the association between RAR and OA of different sites and severity levels in prospective cohorts with imaging-confirmed OA and detailed clinical phenotyping, and further exploring its biological mechanisms; conducting external validation of the model in independent large-scale populations to systematically assess its cross-population robustness; and promoting clinical translation research of the prediction model, ultimately building a dynamically updatable risk assessment tool through multicenter collaboration to provide more clinically applicable decision support for early screening and stratified intervention of OA.

5 Conclusion

In summary, this study demonstrates a robust and significant positive association between elevated RAR levels and the risk of osteoarthritis in middle-aged and older adults. The RAR-based prediction model shows potential clinical utility for early identification of high-risk individuals, formulation of targeted prevention strategies, implementation of early interventions, and dynamic monitoring of disease progression. However, due to the limitations of the present study design and data sources, these conclusions require further validation through more rigorous research designs and independent cohorts in future investigations.

Guardian/patient's consent

This study is based on publicly available data from the NHANES database. All human subjects research has been reviewed by the Ethics Review Committee of the National Center for Health Statistics and conducted in accordance with the Declaration of Helsinki. All participants signed informed consent forms. Therefore, this study waives additional ethics review.

Consent for publication

All authors have reviewed and approved the final version of the manuscript for submission to this journal.

Ethics approval and consent to participate

This study is based on publicly available data from the NHANES database. All human subjects research has been reviewed by the Ethics Review Committee of the National Center for Health Statistics and conducted in accordance with the Declaration of Helsinki. All participants signed informed consent forms. Therefore, this study waives additional ethics review.

Credit author statement

GQ contributed to data analysis, manuscript drafting, and literature search, data collection, curation, and manuscript writing. ZX D supervised the research and edited the manuscript. All authors participated in the study design and manuscript revision, critically reviewed the final version, and approved the submission of this manuscript.

Funding

This study was funded by the Postgraduate Research & Practice Innovation Program of Jiangsu Province (Grant No. SJCX25_0910).