Ultra-Processed Food Intake, Circulating NMR Metabolomic Signatures and Risk of Osteoarthritis in Women

Abstract

Purpose

Prospective evidence reporting statistical associations between ultra-processed food (UPF) intake and osteoarthritis (OA) risk in women is sparse, women suffer a disproportionately heavy OA burden with distinct metabolic and inflammatory characteristics, and the circulating biological correlates of UPF intake have not been integrated with OA risk in a unified analytical framework.

Patients and Methods

In a prospective multi-omics analysis, we studied women without baseline OA, with NOVA-based UPF intake reconstructed from repeated Oxford WebQ 24-hour dietary records. Cox proportional hazards models (age as time scale) estimated hazard ratios (HRs) per 1-SD UPF increase and across quartiles, adjusting for ethnicity, education, income, smoking, alcohol, physical activity and total energy intake (Model 2); stabilised inverse probability weighting (IPW) addressed UPF availability. Nightingale NMR metabolomic and Olink proteomic features were screened sequentially (UPF-to-omics, omics-to-OA; FDR<0.05); candidate correlates were summarised by principal component analysis into a UPF-related omics signature. We performed two-stage correlational screening for omics markers rather than formal causal mediation analysis. Notably, 61.8% of the female cohort lacked valid UPF dietary records; stabilised inverse probability weighting (IPW) with dual trimming thresholds was applied to mitigate potential selection bias from missing dietary data.

Results

Among 13,122 women with UPF data (base female OA cohort 34,322), 528 knee OA, 473 hip OA, 1,908 any OA, 963 knee/hip OA and 57 hand OA events accrued. Model 2 demonstrated per-1-SD UPF increments were associated with higher any OA (HR 1.070, 95% CI 1.023–1.120, P=0.003) and knee/hip OA risk (HR 1.072, 95% CI 1.005–1.143, P=0.035), with the top UPF quartile linked to 28.8% higher knee OA risk; energy-adjusted and IPW-weighted analyses produced consistent estimates, while associations fully attenuated after adiposity adjustment in Model 3. UPF intake was related to 178 NMR features (FDR<0.05), including GlycA, omega-6/linoleic acid percentages and HDL-related lipoprotein indices. The combined PCA NMR signature independently predicted increased knee OA (HR 1.187, 95% CI 1.124–1.253, P<0.001) and any OA risk (HR 1.104, 95% CI 1.070–1.139, P<0.001). In exploratory Olink proteomic analyses, five proteins were associated with UPF intake at FDR<0.05, but no protein passed the pre-specified two-stage FDR screening for both UPF association and OA risk; nominal Olink signals are reported in the Supplementary Materials and interpreted as hypothesis-generating only.

Conclusion

Higher UPF intake was associated with modestly increased OA risk in women; NMR signatures of inflammation and lipid metabolism provided exploratory biological context overlapping adiposity pathways.

Introduction

Osteoarthritis (OA) is a leading cause of chronic pain, mobility limitation and loss of independence in older adults, with the knee and hip joints disproportionately affected and the global burden continuing to rise alongside population ageing and obesity.^1,2 Women carry a substantially larger share of OA-related disability than men: prevalence rises sharply around midlife and after menopause, women report more severe symptoms and are more likely to progress to total joint replacement than age-matched men.^3,4 Female-specific factors—including hormonal changes across the reproductive life course, body composition and joint biomechanics, and the higher prevalence of obesity and metabolic syndrome in post-menopausal women—are widely invoked to explain this excess burden.^5,6 In the absence of disease-modifying therapies, the identification of modifiable risk factors, including diet, is a public-health priority.^1,7 Yet female-specific evidence on dietary determinants of OA risk remains limited and has been dominated by joint-replacement endpoints and selected micronutrients rather than overall diet quality or dietary processing. Women bear a far heavier burden of osteoarthritis-related disability than men: their OA prevalence rises sharply after menopause, accompanied by more severe joint pain and higher rates of joint replacement. Unique female risk factors including postmenopausal hormonal fluctuation, distinct body composition and joint biomechanical characteristics also drive gender disparities in OA risk, while prospective evidence exploring UPF-OA associations exclusively in women remains scarce, justifying our female-only study population.

Over the past two decades, the food supply in high-income countries has shifted markedly toward ultra-processed foods (UPF), defined within the NOVA classification as industrial formulations made from food substances, additives and cosmetic ingredients, engineered for convenience, palatability and shelf-life.⁸ Typical NOVA Group 4 ultra-processed foods include packaged snacks, sugary beverages, instant ready meals, refined bakery products and processed meat substitutes, which are manufactured with large amounts of added sugar, salt, emulsifiers, preservatives and artificial flavorings without intact whole food ingredients. UPF now contribute more than half of total energy intake in some European and North American populations, raising broad concerns about their cardiometabolic and inflammatory consequences. A tightly controlled inpatient feeding trial demonstrated that ultra-processed diets produce greater ad libitum energy intake and weight gain than nutrient-matched unprocessed diets, even when macronutrient profiles are matched,⁹ establishing a direct causal link between food processing and short-term metabolic risk. Several biological pathways have been proposed as plausible bridges from UPF to joint pathology, including positive energy balance and adiposity,^9,10 systemic low-grade inflammation indexed by markers such as glycoprotein acetyls (GlycA),^11,12 altered circulating fatty-acid composition and lipoprotein subclasses,^13,14 and disturbances of the gut microbiota with downstream effects on the gut-joint axis.¹⁵ Heat-processed and reformulated foods are also a major dietary source of advanced glycation end-products, which may accumulate within cartilage matrix and contribute to its stiffening and to chondrocyte dysfunction over time.⁷ Consistent mechanistic evidence supporting these correlational pathways has been systematically synthesised in recent dedicated review articles. A 2025 scoping review specifically summarised epidemiological and preclinical data linking higher UPF intake to elevated knee OA risk in postmenopausal women, highlighting overlapping inflammatory, lipid and adiposity-related pathways that connect industrial food processing to degenerative joint damage. A complementary mechanistic systematic review further elaborated that the destruction of natural food matrix in UPF disrupts satiety signalling, triggers persistent metabolic dysregulation and low-grade systemic inflammation independent of macronutrient composition, providing a unifying biological framework to interpret the short-term feeding trial observations and long-term population chronic disease associations. Such consistent mechanistic foundations lay a biological basis for interpreting the large-scale population-based prospective epidemiological associations between UPF intake and chronic diseases outlined below.

Prospective epidemiological work has consistently associated UPF intake with a range of chronic outcomes. In the NutriNet-Santé cohort, a 10-percentage-point increment in UPF intake was associated with a 12% higher risk of overall cardiovascular disease and a 13% higher risk of coronary disease,¹⁶ and a 15% higher risk of incident type 2 diabetes per 10-percentage-point increment.¹⁷ A systematic review and meta-analysis of 43 observational studies linked higher UPF consumption with greater risk of all-cause mortality (pooled risk ratio 1.25), cardiovascular disease (1.29) and depression (1.20),¹⁸ and a 2024 umbrella review of epidemiological meta-analyses reported convincing or highly suggestive evidence for adverse associations with cardiometabolic, mental and mortality outcomes.¹⁹ In a multinational EPIC cohort study, higher UPF consumption was associated with multimorbidity of cancer and cardiometabolic diseases.²⁰ Within this study, plant-sourced UPF have been associated with cardiovascular morbidity and mortality—in contrast to plant-sourced non-UPF—underscoring that food processing, not nutrient content alone, contributes to risk,²¹ and higher UPF intake has been linked to incident dementia.²²

By contrast, prospective evidence linking UPF intake to musculoskeletal outcomes is comparatively sparse. Dietary research in OA has so far focused on individual nutrients—for example, whole grains and dietary fibre have been inversely associated with radiographic knee OA in the Osteoarthritis Initiative^23,24—or on broad dietary patterns such as the Mediterranean diet, which has been associated with slower OA progression and lower OA-related symptom burden,²⁵ rather than on overall food processing. Crucially, no large prospective study has integrated multi-omic profiling to characterise candidate biological intermediates between UPF and OA risk, and even fewer have focused on women, who shoulder the larger share of OA-related disability.

We therefore conducted a prospective multi-omics analysis in women,^26,27 leveraging NOVA-based UPF exposure reconstructed from the validated Oxford WebQ 24-hour dietary instrument,²⁸ ICD-10-defined incident OA, large-scale NMR metabolomic profiling on the Nightingale Health platform^12,29 (the primary mechanistic layer) and Olink proteomic profiling from the UKB-PPP^30,31 (an exploratory layer). Our aims were threefold: (i) to quantify the association between UPF intake and risk of incident knee, hip, knee/hip and any OA in women, accounting for sociodemographic, lifestyle and energy-intake confounders, and addressing selection bias through inverse probability weighting; (ii) to identify circulating correlates jointly associated with UPF intake and OA risk; and (iii) to summarise these features into a UPF-related omics signature and evaluate its association with OA. We framed the analysis as exploratory and hypothesis-generating, viewing the omics features as providing biological context rather than as strict causal mediators or as diagnostic markers.

Materials and Methods

Study Design and Data Source

This analysis was based on^26,27 a population-based study of approximately 500,000 adults aged 40–69 years recruited from 22 assessment centres across England, Scotland and Wales between 2006 and 2010. We constructed an analytical framework, restricted to women, that linked NOVA-based ultra-processed food (UPF) intake to incident osteoarthritis (OA), with circulating nuclear magnetic resonance (NMR) metabolomics as the primary mechanistic layer and Olink proteomics as a complementary, exploratory layer. Dietary information was derived from previously curated Oxford WebQ dietary data²⁸ and NOVA-based reconstruction of UPF exposure.⁸ Outcomes were derived from medical-record linkage to hospital admissions and ICD-10 diagnosis dates. The overall study design and analytical workflow are summarised in Figure 1.

Figure 1.

Study design and analytic workflow. The analysis focused on women without baseline OA, reconstructed UPF exposure, incident OA follow-up, and exploratory NMR/Olink omics analyses. Dashed arrows indicate exploratory biological context rather than causal mediation.

The analysis was pre-specified as exploratory and hypothesis-generating. Throughout the manuscript, the omics components are described as providing exploratory biological context, candidate correlational phenotypes or hypothesis-generating evidence; they are not interpreted as strict causal mediators, and the omics signatures are not presented as diagnostic models. This study has approval from the North West Multi-Centre Research Ethics Committee, and all participants provided written informed consent. The current application number is 78619.

To streamline understanding of the full analytical pipeline, we briefly summarise the integrated workflow across subsequent subsections: we constructed multiple complementary UPF exposure metrics; applied two pre-specified age-scale Cox models (primary Model 2 and adiposity sensitivity Model 3) to quantify UPF-OA associations; used dual-trimmed stabilised IPW and layered sensitivity analyses (energy-adjusted UPF substitution, early-event exclusion, restricted cubic splines) to address selection bias and test result robustness; and conducted two-stage FDR-controlled correlational screening on NMR metabolomic and exploratory Olink proteomic data, before summarising shared omics correlates into a unified UPF-related signature via PCA for OA risk assessment.

Study Population

Analyses were restricted to female participants. The base female OA cohort included women with a valid baseline assessment date and age, valid follow-up time, and no record of OA at any anatomical site (knee, hip, hand, or other) before baseline; women with such pre-baseline diagnoses were excluded. For the primary UPF analyses, we further restricted the sample to women with a valid UPF grams/day exposure value. A comparison of UPF-available versus UPF-unavailable participants, together with inverse probability weighting (IPW), was used as a key bias assessment to evaluate whether differential dietary completion plausibly affected the primary estimates. Baseline characteristics, displayed by quartile of UPF grams/day, are reported in Table 1. The participant selection and database cleaning process, including exclusion criteria, sample sizes at each stage and missing-data screening rules, is shown in Supplementary Figure S7.

Table 1.

Baseline Characteristics of Women by Quartiles of UPF Grams/Day

Characteristic	Overall	Q1	Q2	Q3	Q4
N	13,122	3281	3280	3283	3278
Age, years	54.5 (7.9)	55.2 (7.6)	54.8 (7.9)	54.4 (7.9)	53.5 (8.1)
BMI, kg/m2	26.0 (4.8)	25.3 (4.4)	25.8 (4.6)	26.2 (4.9)	26.6 (5.2)
Waist circumference, cm	82.1 (11.9)	80.4 (11.2)	81.8 (11.5)	82.6 (11.9)	83.7 (12.6)
White ethnicity, n (%)	12,519 (95.4%)	3131 (95.4%)	3161 (96.4%)	3154 (96.1%)	3073 (93.7%)
College/university education, n (%)	6364 (48.5%)	1806 (55.0%)	1618 (49.3%)	1533 (46.7%)	1407 (42.9%)
Current smoker, n (%)	763 (5.8%)	192 (5.9%)	188 (5.7%)	172 (5.2%)	211 (6.4%)
Current alcohol drinker, n (%)	12,297 (93.7%)	3114 (94.9%)	3097 (94.4%)	3087 (94.0%)	2999 (91.5%)
Physical activity, log MET-min/week	7.4 (1.2)	7.4 (1.2)	7.4 (1.1)	7.3 (1.3)	7.3 (1.3)
Diabetes, n (%)	329 (2.5%)	82 (2.5%)	81 (2.5%)	81 (2.5%)	85 (2.6%)
Energy intake, kcal/day	1933.1 (589.5)	1772.3 (537.7)	1890.1 (542.6)	1969.4 (576.8)	2100.9 (645.7)
UPF intake, g/day	167.1 (158.7)	32.7 (17.9)	89.5 (17.3)	165.5 (28.4)	381.0 (171.7)
UPF weight percentage, %	5.4 (5.2)	1.1 (0.7)	3.0 (0.9)	5.5 (1.6)	12.1 (6.0)
Knee OA events, n	528	116	118	141	153
Any OA events, n	1908	456	454	497	501

Notes: Values are mean (SD) or n (%). OA, osteoarthritis. Q1 denotes the lowest UPF intake and Q4 denotes the highest UPF intake.

Abbreviation: UPF, ultra-processed food.

Ultra-Processed Food Exposure

The primary exposure was UPF intake expressed in grams per day, derived from best-available NOVA-based reconstruction using repeated Oxford WebQ dietary variables.^8,28 This metric represents the average daily intake of foods classifiable as NOVA group 4 (ultra-processed) across the available WebQ rounds. In the main model, UPF grams/day was winsorised at the 0.5th and 99.5th percentiles and standardised; hazard ratios are reported per 1-SD increase to facilitate comparison across UPF metrics.

To test whether the UPF grams/day signal merely captured total food mass or total energy intake, we additionally constructed three alternative UPF metrics: (i) UPF weight percentage (UPF grams as a percentage of total food weight); (ii) UPF grams per 1,000 kcal (UPF grams/day divided by total energy in kcal, multiplied by 1,000); and (iii) a UPF servings/day proxy using 100 g as an approximate serving size, for interpretive purposes only. Item-level UPF energy data were not available in this release, so a UPF energy percentage metric could not be reliably constructed. UPF grams/day was also modelled as quartiles and quintiles, with the lowest category (Q1) as the reference.

To improve reproducibility, we provided the item-level Oxford WebQ-to-NOVA mapping table used for UPF exposure reconstruction in Supplementary Table S1, detailed IPW model specification, stabilised weight calculation and trimming rules in Supplementary Methods S1.

Osteoarthritis Outcome Definitions

The primary outcome was incident knee OA (ICD-10 M17). Secondary outcomes were hip OA (M16), any OA (M15-M19), knee/hip OA (M17 or M16) and hand OA (M18). Knee or hip joint replacement was pre-specified as a clinically meaningful secondary endpoint; however, reliable surgical event data were not available in this data release and joint replacement was therefore not reported as a primary result. Follow-up began at the baseline assessment date and ended at the first occurrence of the target OA diagnosis, death, or administrative censoring, whichever occurred earliest. Cox proportional hazards models used age as the underlying time scale, implemented as Surv (age_start, age_end, event), which is recommended for age-related chronic disease outcomes in cohorts with a broad recruitment age range.

Covariates

The primary model (Model 2) was pre-specified to adjust for sociodemographic and lifestyle factors that are established or plausible confounders of the UPF-OA association: ethnicity, educational attainment, household pre-tax annual income, smoking status, alcohol drinking status, physical activity, and total energy intake. Ethnicity was categorised as White, Non-white or Unknown; education as College/University, A/AS levels, School/Vocational, Other, None or Unknown; income as <£31k, £31–52k, £52–100k, >£100k or Unknown; smoking status as Never, Previous, Current or Unknown; alcohol drinking status as Current, Previous, Never or Unknown. Physical activity was modelled as the log1p-transformed total metabolic equivalent (MET-min/week) and total energy intake as kcal/day. Missing covariate categories were modelled explicitly to preserve sample size.

Model 3 extended Model 2 with additional adjustment for body mass index (BMI), waist circumference and physician-diagnosed diabetes. Because BMI, waist circumference and diabetes may simultaneously act as confounders and lie on the metabolic/adiposity pathway between UPF intake and OA risk, Model 3 was specified as an adiposity/metabolic sensitivity model rather than as the principal model.

NMR Metabolomic and Olink Proteomic Analyses

Circulating NMR metabolomics, profiled on the Nightingale Health platform,^12,29 served as the primary mechanistic layer. Pre-specified candidate features included glycoprotein acetyls (GlycA), glucose, lactate, citrate, albumin, branched-chain and aromatic amino acids, HDL/LDL/VLDL particle subclasses, triglycerides, total cholesterol, ApoA1/ApoB and fatty-acid composition (omega-3, omega-6, monounsaturated, polyunsaturated and saturated fatty acid percentages of total fatty acids, and the corresponding ratios). NMR features were filtered for missingness, winsorised for extreme values and standardised. UPF-to-NMR associations were estimated using linear regression adjusted for Model 2 covariates; NMR-to-OA associations were estimated using Cox models additionally adjusted for UPF grams/day and the same covariates.

Olink proteomic profiling from the UKB-PPP was treated as a complementary,^30,31 exploratory mechanistic layer. The UKB-PPP uses the proximity extension assay (PEA) to quantify circulating proteins in plasma.³¹ Pre-specified protein families of interest spanned inflammation/immune signalling, extracellular matrix, cartilage and tissue remodelling, MMP/ADAMTS, IGF/IGFBP, and fibrosis/tissue repair. Because Olink-overlap sample sizes and OA event counts were considerably smaller than for NMR, Olink results are reported only as exploratory supplements and did not contribute to the principal conclusions.

Statistical Analysis

We clearly distinguish pre-specified primary analyses and post-hoc exploratory analyses here. Cox proportional hazards models, stabilised inverse probability weighting, restricted cubic spline non-linear dose-response tests, and three alternative UPF exposure metrics (UPF weight percentage, UPF g per 1000 kcal, UPF servings proxy) were fully pre-planned and documented before accessing the study dataset. By contrast, Olink proteomic correlation and outcome analyses were post-hoc exploratory assessments carried out after completion of primary NMR metabolomic analyses, and these proteomic analyses were not part of the original pre-specified analytical framework. Baseline characteristics were summarised by quartile of UPF grams/day. Continuous variables were expressed as mean (standard deviation) or median (interquartile range) and compared using analysis of variance or the Kruskal–Wallis test as appropriate; categorical variables were expressed as frequencies (percentages) and compared using the chi-squared test.

For the main exposure-outcome analyses, Cox proportional hazards models with age as the time scale estimated hazard ratios (HRs) and 95% confidence intervals (CIs) for the association between UPF intake and OA risk. The proportional-hazards assumption was assessed using Schoenfeld residual tests. UPF was modelled both as a continuous variable (per 1-SD increase) and as quartiles (reference: lowest quartile); a linear trend test (P for trend) was performed by assigning the median UPF value of each quartile. Potential non-linear dose-response was assessed using Cox-based restricted cubic splines (RCS) with knots at the 5th, 35th, 65th and 95th percentiles of the exposure.

To address potential selection bias arising from incomplete UPF reporting, we constructed a logistic model for UPF availability conditional on Model 2 covariates and derived stabilised inverse probability weights (IPW). Weights were trimmed at the 1st/99th and 5th/95th percentiles, and weighted Cox models re-estimated the UPF-OA associations under each trimming rule.

For reproducibility, the implemented IPW model was specified as logit{Pr(A_i=1 | X_i)} = alpha_0 + alpha_1 age_at_recruitment_i + alpha_2 BMI_i + alpha_3 waist_circumference_i + alpha_4 log(total MET)_i + alpha_5 energy_kcal_i + alpha_6 ethnicity_i + alpha_7 qualifications_i + alpha_8 household_income_i + alpha_9 smoking_i + alpha_10 alcohol_i + alpha_11 diabetes_i, where A_i=1 indicated availability of a non-missing UPF grams/day value. Stabilised weights for UPF-available participants were calculated as SW_i=Pr(A=1)/Pr_hat(A_i=1 | X_i), then truncated at the 1st/99th and 5th/95th percentiles before re-estimating the Cox models. The formula, trimming rules and implementation details are provided in Supplementary Methods S1.

Multi-omics screening was conducted in two stages. First, UPF-to-omics associations were estimated using multivariable linear regression on filtered, winsorised and standardised features. Second, omics-to-OA associations were estimated using Cox models. Multiple testing was controlled using the Benjamini-Hochberg false-discovery rate (FDR<0.05) within each layer. Candidate correlates were defined as those passing FDR thresholds in both directions. The two-stage correlational screening procedure adopted in this study only screens biomarkers that are statistically associated with both ultra-processed food intake and osteoarthritis risk. Although this analytical workflow is structurally similar to mediation analysis, formal causal mediation decomposition cannot be conducted in our research. Dietary and multi-omics indicators were only measured once at baseline cross-sectionally, lacking longitudinal repeated data to confirm the chronological order of exposure, intermediate markers and disease onset, which fails to meet the core prerequisite for reliable causal mediation analysis.

The two-stage FDR screening identified candidate bridge correlates linking UPF and OA; all NMR features input to PCA were preprocessed by missing value filtering, winsorization and standardization, principal component analysis was run without orthogonal or varimax rotation, and only the first principal component (PC1) was extracted as the unified UPF-related metabolomic signature score. The markedly smaller analytical sample for Olink proteomics stemmed from study design constraints: plasma proteomic quantification via PEA was exclusively performed within the independent UKB-PPP substudy, rather than for all participants who received NMR metabolomic profiling, resulting in a limited overlapping population with complete UPF dietary records, proteomic data and incident OA follow-up outcomes.

A UPF-related omics signature was constructed by principal component analysis (PCA) on the candidate features, retaining the first principal component as the summary score, and the resulting signature was modelled in Cox regression for each OA outcome.

Sensitivity analyses tested the robustness of the main results, including: (i) Model 3 with additional adjustment for BMI, waist circumference and diabetes; (ii) exclusion of OA events occurring within the first 2 or 4 years of follow-up to mitigate reverse-causation; and (iii) use of energy-adjusted UPF metrics in place of UPF grams/day. All analyses were performed in R (version 4.3.1) using the survival, rms, mice and ipw packages. Beyond multiple-testing adjustments, two-sided P<0.05 was considered statistically significant.

Results

Study Population and UPF Exposure Distribution

We stratified all incident OA outcome findings by the strength of statistical evidence for clear interpretation. First, robust, statistically significant associations were observed for any OA and combined knee/hip OA across primary, energy-adjusted and IPW-weighted models. Second, the continuous exposure model for knee OA yielded a borderline significant association (P=0.05), which represents a marginal signal with weak statistical certainty and should not be interpreted as a definitive positive association independently. Third, hip OA showed no consistent directional or statistically meaningful association with UPF intake. Fourth, analyses of Olink proteomic signatures were limited by small sample size and low event counts and are only reported as exploratory hypothesis-generating data.

The base female OA cohort comprised 34,322 women. UPF grams/day was available for 13,122 women (missing rate 61.8%). During follow-up in this UPF-available sample, we observed 528 incident knee OA events, 473 hip OA events, 1,908 any OA events, 963 knee/hip OA events and 57 hand OA events; reliable knee/hip joint replacement event data were not available. In the NMR overlap, the UPF-available subsample was identical to the main analytical sample. The Olink overlap included 1,343 women, with 50 knee OA and 177 any OA events (Figure 1).

Across quartiles of UPF grams/day, both UPF grams/day and UPF weight percentage rose monotonically by design. Baseline age, BMI, waist circumference, smoking, alcohol drinking, physical activity and total energy intake also varied across quartiles (Table 1), indicating that UPF intake is correlated with the broader sociodemographic and metabolic background. Accordingly, all subsequent models adjusted for sociodemographic, lifestyle and energy covariates, while adiposity/metabolic factors were reserved for sensitivity adjustment in Model 3.

UPF Grams/Day and OA Risk

In Model 2, each 1-SD increment in UPF grams/day was associated with higher risk of any OA (HR 1.070, 95% CI 1.023–1.120, P=0.003) and of knee/hip OA (HR 1.072, 95% CI 1.005–1.143, P=0.035). For the primary outcome of knee OA, the direction was consistent but the association was borderline (HR 1.087, 95% CI 1.000–1.181, P=0.050). Hip OA showed no clear association (HR 1.036, 95% CI 0.941–1.142, P=0.468). The main Cox results are shown in Table 2 and Figure 2.

Table 2.

Ultra-Processed Food Intake and Risk of Osteoarthritis Outcomes

Outcome	Exposure	Model	N	Events	HR (95% CI)	P Value	Notes
Knee OA	UPF grams/day, per 1 SD	Model2	13,122	528	1.087 (1.000, 1.181)	0.050
Hip OA	UPF grams/day, per 1 SD	Model2	13,122	473	1.036 (0.941, 1.142)	0.468
Any OA	UPF grams/day, per 1 SD	Model2	13,122	1908	1.070 (1.023, 1.120)	0.003
Knee/hip OA	UPF grams/day, per 1 SD	Model2	13,122	963	1.072 (1.005, 1.143)	0.035
Any OA	UPF weight %, per 1 SD	Model2	13,122	1908	1.043 (0.996, 1.092)	0.073
Any OA	UPF g/1000 kcal, per 1 SD	Model2	13,122	1908	1.067 (1.020, 1.115)	0.004
Knee OA	UPF grams/day, per 1 SD	Model2, ipw_1_99	13,117	527	1.078 (0.990, 1.173)	0.083
Knee OA	UPF grams/day, per 1 SD	Model2, ipw_5_95	13,117	527	1.081 (0.994, 1.175)	0.068
Any OA	UPF grams/day, per 1 SD	Model2, ipw_1_99	13,117	1907	1.074 (1.025, 1.125)	0.003
Any OA	UPF grams/day, per 1 SD	Model2, ipw_5_95	13,117	1907	1.077 (1.028, 1.128)	0.002
Knee OA	UPF grams/day quartile 4	Model2	13,122	528	1.288 (1.003, 1.654)	0.047
Knee OA	UPF grams/day quartile trend	Model2	13,122	528	1.001 (1.000, 1.002)	0.018	Trend test
Any OA	UPF grams/day quartile 4	Model2	13,122	1908	1.139 (0.999, 1.299)	0.051
Any OA	UPF grams/day quartile trend	Model2	13,122	1908	1.001 (1.000, 1.001)	0.019	Trend test

Notes: Model 2 adjusted for ethnicity, education, household income, smoking, alcohol drinking, physical activity, and total energy intake. IPW models used stabilized inverse probability weights for UPF availability. The IPW model formula, stabilized weight definition, and trimming rules are provided in Supplementary Methods S1.

Figure 2.

UPF intake and osteoarthritis risk. Forest plot showing hazard ratios for UPF grams/day and OA outcomes. Continuous estimates are per 1-SD increase; quartile estimates compare Q4 with Q1. The x-axis is shown on a log scale.

In quartile analyses, the highest quartile of UPF grams/day was associated with higher knee OA risk relative to the lowest (Q4 vs Q1: HR 1.288, 95% CI 1.003–1.654, P=0.047; P-trend=0.018). For any OA, the corresponding Q4 vs Q1 HR was 1.139 (95% CI 0.999–1.299, P=0.051; P-trend=0.019). These results indicate a weak but directionally consistent dose gradient rather than a single strong linear effect (Table 2 and Figure 2).

Alternative UPF Metrics, IPW and Adiposity/Metabolic Adjustment

Using energy-adjusted UPF (grams per 1,000 kcal), any OA remained associated with UPF intake (HR 1.067, 95% CI 1.020–1.115, P=0.004), indicating that the any-OA result was not fully explained by total energy intake. UPF weight percentage showed a directionally consistent but non-significant association with any OA (HR 1.043, 95% CI 0.996–1.092, P=0.073).

IPW analyses addressed potential selection bias driven by UPF availability. For any OA, the IPW-weighted HR was 1.074 (95% CI 1.025–1.125, P=0.003) with 1/99 trimming and 1.077 (95% CI 1.028–1.128, P=0.002) with 5/95 trimming; direction and statistical significance were stable across trimming rules. For knee OA, IPW-weighted HRs were approximately 1.08 with P-values between 0.068 and 0.083, indicating a directionally consistent but weaker signal. Under Model 3, additionally adjusting for BMI, waist circumference and diabetes, the HR for knee OA attenuated to 1.025 (P=0.565), and the HR for any OA to 1.040 (P=0.090), consistent with substantial overlap between UPF, the adiposity/metabolic background and OA risk—though this attenuation may also reflect over-adjustment if these factors lie on the causal pathway. Complete IPW-weighted models are reported in Supplementary Table S3 and Supplementary Figure S3; additional sensitivity models are summarised in Supplementary Table S4.

Non-Linear Analyses

Cox-based restricted cubic spline (RCS) models showed a statistically significant overall association between UPF grams/day and any OA (overall P=0.010), but no evidence of non-linearity (non-linear P=0.341). For knee OA, the overall association did not reach conventional significance (overall P=0.115; non-linear P=0.411). Knee/hip OA and hip OA also showed no consistent non-linear signal. Overall, the RCS results support a weak monotonic dose gradient rather than a clear threshold effect. Full RCS results are reported in Supplementary Table S2, and the RCS curves for knee OA and any OA are shown in Supplementary Figures S1 and S2, respectively.

UPF-Related NMR Metabolomic Features

UPF intake was associated with widespread changes in the circulating NMR metabolome. Across UPF metrics, 522 UPF-NMR associations passed FDR<0.05; using the primary exposure UPF grams/day, 178 NMR features were significant at FDR<0.05. The most prominent UPF-related features involved fatty-acid composition, lipoprotein and lipid subclasses, amino acids, and inflammation/metabolism-related indices. Figure 3 displays representative NMR candidate correlates that also passed the omics-to-OA filter; the global UPF-to-NMR distribution is shown in Supplementary Figure S4. These features are interpreted as providing biological context, not as strict causal mediators. Consistent alterations across multiple lipid subclasses, fatty acid fractions and inflammatory markers jointly form a coherent metabolic signature linking UPF intake to joint disease risk. GlycA, a validated systemic inflammatory marker, and disturbed circulating omega-6, linoleic acid and HDL lipoprotein profiles collectively reflect the adverse metabolic milieu induced by high ultra-processed food consumption. As our primary mechanistic readout, these reproducible NMR biomarkers lay the central biological foundation for interpreting the UPF-OA association observed in the Cox regression models.

Figure 3.

Exploratory UPF-related omics candidate correlates. Top candidate NMR candidate correlates ranked by the exploratory product index of the UPF-to-omics beta and the omics-to-OA log hazard ratio. These features provide biological context and should not be interpreted as causal mediators.

Among the candidate correlates (320 in total, including 264 primary candidates predominantly from NMR), representative NMR features included the omega-6 fatty acids to total fatty acids percentage, the linoleic acid to total fatty acids percentage, the polyunsaturated to monounsaturated fatty acid ratio, the monounsaturated fatty acids to total fatty acids percentage, several HDL-related lipid composition indices, and GlycA (Figure 3 and Supplementary Table S5). Together, these results indicate that higher UPF intake corresponds to a circulating phenotype enriched for altered fatty-acid composition, lipoprotein composition and metabolism-related inflammation.

UPF-Related Omics Signature and OA

A principal-component summary of the UPF-related NMR features (the UPF-related NMR signature) was positively associated with OA risk. In Model 2, the UPF-related NMR signature was associated with knee OA (HR 1.187, 95% CI 1.124–1.253, P<0.001) and any OA (HR 1.104, 95% CI 1.070–1.139, P<0.001). In the UPF-available subsample with additional adjustment for UPF grams/day, the signature remained associated with knee OA (HR 1.266) and any OA (HR 1.106). However, in Model 3, with additional adjustment for BMI, waist circumference and diabetes, the signature was no longer associated with OA, suggesting substantial overlap of the omics phenotype with the adiposity/metabolic background. Full signature results are reported in Supplementary Table S6, and a supplementary forest plot of signature-OA associations is provided in Supplementary Figure S5.

Olink Exploratory Analyses and Sensitivity Analyses

Olink proteomic analyses were exploratory only due to limited sample size and scarce OA endpoints. Only five markers reached FDR significance for UPF exposure, and no stable protein mediator linking UPF and OA was identified (Supplementary Figure S6).

Sensitivity analyses confirmed that the association between UPF intake and any OA was the most robust finding, while the knee OA association was borderline and sensitive to model adjustment. Combined consistent Cox models, energy-adjusted UPF metrics, IPW weighting results and core NMR metabolomic signatures collectively support a hypothesized association between UPF consumption, circulating metabolic profiles and female OA risk.

Discussion

In this prospective multi-omics analysis of women without baseline OA, higher ultra-processed food (UPF) intake was associated with modestly increased risks of incident any OA (HR per 1-SD 1.070, 95% CI 1.023–1.120) and knee/hip OA (HR 1.072, 95% CI 1.005–1.143), with a directionally consistent but borderline signal for knee OA in continuous models and a clearer signal in the highest quartile (Q4 vs Q1 HR 1.288, 95% CI 1.003–1.654; P-trend=0.018). Associations persisted when UPF was expressed per 1,000 kcal and after inverse probability weighting for UPF availability, and attenuated only after additional adjustment for BMI, waist circumference and diabetes, suggesting that part of the association operates through, rather than independently of, the adiposity/metabolic axis. To our knowledge, this is the first large prospective female cohort study investigating UPF-OA associations with a multi-omics framework centered on NMR metabolomics as the primary mechanistic layer, with Olink proteomics added solely as an exploratory supplementary analysis. Consistent with this pre-specified analytical hierarchy, UPF intake was associated with broad changes in the circulating NMR metabolome—including GlycA, fatty-acid composition and HDL-related lipoprotein indices—and a PCA-derived UPF-related NMR signature was independently associated with OA risk in models unadjusted for adiposity. By contrast, Olink proteomic findings were merely exploratory without stable mechanistic signals.

These findings extend the broader UPF-chronic disease literature to a musculoskeletal endpoint that has rarely been examined in this context. In the NutriNet-Santé cohort, a 10-percentage-point increment in UPF intake was associated with a 12% higher risk of cardiovascular disease and a 13% higher coronary risk,¹⁶ and a 15% higher risk of incident type 2 diabetes.¹⁷ A systematic review and meta-analysis of cohort studies linked higher UPF consumption with greater risks of all-cause mortality (pooled risk ratio 1.25), cardiovascular disease (1.29) and depression (1.20),¹⁸ and a 2024 umbrella review of 45 unique pooled analyses found convincing or highly suggestive evidence for adverse associations between UPF intake and multiple cardiometabolic, mental and mortality outcomes.¹⁹ In the multinational EPIC cohort, higher UPF consumption was associated with multimorbidity of cancer and cardiometabolic diseases.²⁰ Within this study, plant-sourced UPF were associated with higher cardiovascular morbidity and mortality, in contrast to plant-sourced non-UPF, underscoring that processing—not only nutrient content—contributes to risk,²¹ and higher UPF intake has been linked to incident dementia in the same cohort.²² The Monteiro et al NOVA framework provides the conceptual basis for these analyses, emphasising that UPFs are formulations engineered for palatability and shelf-life and may exert health effects beyond their nutrient composition,⁸ a hypothesis supported by the inpatient feeding trial of Hall et al, in which ultra-processed diets produced greater energy intake and weight gain than nutrient-matched unprocessed diets over only two weeks.⁹ Our results add a women-specific, OA-focused prospective signal to this literature: the effect sizes per 1-SD UPF and per quartile are modest but consistent with the magnitudes reported for other chronic outcomes when expressed on comparable scales, and they are concentrated in the OA outcomes most directly tied to weight-bearing joints (knee and knee/hip). Although the per-SD HR increment of approximately 7% and 28.8% elevated knee OA risk in the highest UPF quartile represent modest individual-level effects, the widespread high consumption of UPF globally means population-wide dietary restrictions on ultra-processed foods could substantially reduce the overall OA disease burden among women. Nevertheless, the observational nature of this study cannot support immediate clinical dietary guidelines, and further intervention trials are required to validate whether limiting UPF intake improves long-term joint health.

Compared with the existing diet-OA literature, our findings complement and extend several earlier lines of evidence. Higher whole-grain and dietary-fibre intakes have been associated with reduced incidence of radiographic knee OA in the Osteoarthritis Initiative and Framingham cohorts,^23,24 in directional opposition to higher UPF intake, which is typically energy-dense, low in dietary fibre and rich in refined starches and added sugars. Adherence to a Mediterranean diet—itself characterised by minimally processed foods, whole grains, fruit, vegetables and unsaturated fats—has been prospectively associated with slower OA progression and lower OA-related symptom burden in older adults.²⁵ Taken together with our results, this evidence base is increasingly compatible with the view that overall dietary patterns and food processing, rather than single nutrients in isolation, shape long-term OA risk. Our NMR findings—particularly the inverse association between UPF intake and the omega-6 to total fatty acids percentage, the linoleic acid percentage and the polyunsaturated to monounsaturated fatty acid ratio, and the positive associations with monounsaturated fatty-acid composition and GlycA—are consistent with prior observations that circulating linoleic acid and other omega-6 biomarkers are inversely associated with cardiovascular disease and incident type 2 diabetes,^13,14 suggesting that the metabolic milieu indexed by these features is broadly cardiometabolically adverse and may also be relevant to joint health. These consistent NMR-derived metabolic alterations provide clear biological pathways explaining the UPF-OA association, with circulating metabolomic signatures serving as critical intermediate readouts bridging UPF consumption and joint pathology. GlycA reflects chronic systemic inflammation, while disturbed fatty acid and HDL lipoprotein metabolism directly modulate synovial and chondrocyte function; collectively this NMR panel constitutes our primary circulating mechanistic evidence. In contrast, Olink proteomic markers were included only as an exploratory supplementary layer with insufficient statistical power, and no stable circulating protein mediator was validated in this cohort. We rely on these measurable circulating molecular signatures to dissect how dietary ultra-processing drives systemic metabolic dysfunction that further elevates OA risk. Elevated GlycA signals chronic low-grade systemic inflammation, which penetrates joint tissues, activates NF-κB inflammatory signalling, accelerates chondrocyte apoptosis and cartilage matrix degradation. Disturbed fatty acid and HDL lipoprotein profiles further disrupt lipid homeostasis within synovial fluid, exacerbating synovitis and joint degenerative damage. Importantly, the PCA-integrated NMR signature aggregated all these inflammatory and lipid metabolic perturbations into a unified biomarker score that independently predicted elevated OA risk without adiposity adjustment, and these large, reproducible NMR data form the primary mechanistic evidence of our analysis. Accumulated systematic evidence further confirms that disordered circulating lipoprotein and fatty acid profiles, alongside elevated systemic inflammatory markers including GlycA, directly drive synovial inflammation, chondrocyte dysfunction and cartilage matrix breakdown via NF-κB and lipid metabolic signalling pathways. GlycA itself integrates signals from multiple acute-phase glycoproteins (such as alpha-1 acid glycoprotein, haptoglobin and alpha-1-antitrypsin), persists within individuals over years and predicts long-term risk of severe infection and cardiovascular disease.^12,29 Our results extend this line of work by showing that an UPF-anchored NMR phenotype, summarising these features, also tracks with OA incidence in women.

Collectively, the present study delivers three core novel observations. First, higher UPF consumption independently predicts elevated risk of incident knee/hip and all-cause OA among women after comprehensive confounder adjustment and IPW sensitivity correction. Second, large-scale NMR metabolomics identifies a consistent inflammatory and lipid metabolic signature bridging UPF intake and OA, which serves as the primary mechanistic evidence supporting our epidemiological associations. Third, exploratory Olink proteomics fails to detect stable protein mediators linking UPF to joint disease, with limited statistical power restricting reliable proteomic inference. Taken together, these epidemiologic and multi-omic results expand existing diet-OA research by highlighting food ultra-processing as a modifiable joint health risk factor and revealing circulating metabolic biomarkers as potential intermediate pathways.

Olink proteomics served merely as exploratory supplementary analysis with limited sample size and OA events, and no consistent protein mediator linking UPF intake to OA risk was identified. No further mechanistic interpretation of these preliminary proteomic results is provided, and replication in larger proteomic cohorts is required for reliable inference.

Beyond proteomic limitations, another general constraint of existing mechanistic evidence should be noted: metabolic responses documented in short-term inpatient feeding trials only capture acute energy balance changes and cannot be used to infer long-term causal relationships between UPF consumption and chronic OA risk over decades of follow-up. Despite these analytical limitations of multi-omic profiling, several biologically plausible pathways may explain the observed UPF-OA association in women.

Notably, all NMR and proteomic biomarkers identified herein were obtained via cross-sectional correlational screening rather than validated causal mediation analysis. We did not perform BMI-stratified analyses for the NMR signature-OA association due to limited event numbers in each BMI subgroup, which prevents us from fully separating mediation and confounding explanations for the complete attenuation after adiposity adjustment. Larger cohorts with adequate OA events across BMI strata are needed to resolve this question in future work. Several biological mechanisms may plausibly link UPF intake to OA pathogenesis in women. First, UPF intake promotes positive energy balance, adiposity and metabolic syndrome,^9,18 and obesity is the strongest established modifiable risk factor for knee OA, acting via both biomechanical loading and systemic, low-grade inflammation.^10,32 Notably, although we observed that adiposity-related metabolic biomarkers partially mediated the statistical correlation between UPF intake and OA risk, cross-sectional single measurement data prevent us from quantifying the magnitude of indirect effects mediated by adiposity. Full attenuation of the NMR signature-OA association after Model 3 adjustment for adiposity and glycemic markers supports an adiposity-mediated biological pathway. However, we lack sufficient OA events to conduct stable stratified analyses across low, medium and BMI-high subgroups, so we cannot definitively disentangle whether the attenuation arises from confounding or true mediating effects. Over-adjustment bias may also exist since adiposity phenotypes sit on the plausible causal pathway between UPF intake and OA risk. We can only infer that adiposity may act as a potential intermediate correlate rather than a confirmed causal mediator linking UPF to OA. The substantial attenuation of UPF-OA and signature-OA associations after adjustment for BMI, waist circumference and diabetes in our Model 3 is consistent with this adiposity/metabolic-mediation interpretation, although such adjustment is also susceptible to over-adjustment bias when the adjusted covariates lie on the causal pathway. Second, OA is increasingly recognised as an inflammatory and metabolic disease in which low-grade systemic and synovial inflammation, captured in part by GlycA and related markers, contributes to cartilage breakdown and synovitis.^6,7,11,32 Third, alterations in fatty-acid composition—lower omega-6 and polyunsaturated proportions and higher monounsaturated proportions—and HDL-related lipoprotein indices may reflect dysregulated lipid metabolism that contributes to chondrocyte dysfunction and synovial inflammation.^13,14 Fourth, UPF intake has been linked to gut dysbiosis, and the gut-joint axis has been proposed as a contributor to OA pathology, with intestinal microbiome composition being associated with circulating inflammatory markers and with joint pain in population-based studies.¹⁵ Heat-processed and reformulated foods are also a major dietary source of advanced glycation end-products, which may accumulate within cartilage matrix and contribute to its stiffening and to chondrocyte dysfunction with ageing.⁷ Together, these mechanisms make the UPF-OA association biologically plausible, but the present cross-sectional omics design cannot disentangle their relative contributions.

From a clinical and public-health perspective, the observed effect sizes are modest at the individual level—roughly 7–9% higher risk per 1-SD increase in UPF intake, and approximately 29% higher knee OA risk in the highest versus the lowest quartile—but are potentially meaningful at the population level given the high and rising prevalence of UPF in modern diets and the very large global burden of OA in women.^1,2,4 Importantly, UPF is a modifiable exposure: dietary patterns that limit UPF and emphasise minimally processed foods, whole grains, fruit and vegetables already feature in cardiovascular and metabolic prevention guidance. Our results suggest that reductions in UPF may yield collateral benefits for women’s joint health, alongside the better-established cardiovascular, metabolic and cognitive benefits documented elsewhere.^16–22 We emphasise, however, that the present findings are hypothesis-generating rather than directly actionable: they do not establish causality, and they should not be interpreted as a basis for targeting specific foods or food groups in clinical practice in the absence of independent replication and intervention evidence. Future trials and Mendelian randomisation analyses will be required to determine whether reducing UPF intake yields measurable changes in joint health outcomes.

Strengths and Limitations

This analysis has several strengths. It is, to our knowledge, among the first prospective multi-omics analyses of UPF and OA, with a substantial number of incident OA events in women in a population-based cohort. We pre-specified women as the analytical population, in recognition of the disproportionate OA burden in women,^3–5 rather than relying on sex-stratified subgroup analyses post hoc. The exposure was constructed via NOVA-based reconstruction from repeated Oxford WebQ records,²⁸ a dietary instrument with established biomarker-based validation, and we evaluated multiple alternative UPF metrics (weight percentage, grams per 1,000 kcal, servings/day proxy), IPW for UPF availability under two trimming rules, and energy-adjusted versions, all of which consistently supported the primary findings. We integrated the Nightingale NMR metabolomics platform^12,29 as the primary mechanistic layer, with Olink proteomics^30,31 as an exploratory complement, and we adopted a pre-specified two-stage screening (UPF-omics followed by omics-OA) with Benjamini-Hochberg FDR control. The use of age as the underlying Cox time scale, with sensitivity analyses excluding events occurring within the first 2 or 4 years of follow-up, mitigates reverse-causation concerns. The omics signature was explicitly framed as exploratory and was not presented as a diagnostic or causal-mediation model.

Several limitations should be acknowledged. UPF exposure was reconstructed from Oxford WebQ data rather than from dedicated UPF dietary instruments, which may yield exposure misclassification and limit direct comparability with cohorts using purpose-built UPF questionnaires; the missing-data rate for UPF was substantial (61.8%), although IPW analyses suggested a limited impact on the primary inference. ICD-10-based ascertainment captures clinically diagnosed OA and underestimates radiographic or symptomatic disease that does not reach secondary care; agreement between radiographic, clinical and self-reported OA definitions is modest,³³ so OA case ascertainment is incomplete and may bias estimates toward the null. Reliable joint-replacement event data were not available in this release. Despite extensive covariate adjustment, residual confounding by unmeasured factors—including occupational and recreational joint loading, prior joint injury, reproductive history and menopausal hormone use—cannot be excluded. The multi-omics analyses were strictly observational and exploratory; we did not perform formal causal mediation analyses, and the marked attenuation of signature-OA associations after adiposity adjustment is compatible with both mediation and over-adjustment. This study is predominantly of White European ancestry, with healthy-volunteer characteristics, limiting generalisability to other populations and to women in lower-income settings, where UPF consumption is rising particularly rapidly. We did not utilise official sampling stratification or cluster weights across Cox regression, IPW adjustment and NMR omics PCA analyses. To mitigate potential population-level inferential bias from unweighted sampling, we comprehensively adjusted sociodemographic covariates capturing cohort stratification features, implemented dual-trimmed IPW to correct dietary missingness selection bias, and conducted multiple pre-specified sensitivity analyses. We acknowledge that residual minor representativeness bias cannot be fully eliminated, and replication with weighted population models is warranted to confirm generalisable population associations. The Olink overlap was small, multiple-testing control used FDR rather than family-wise error correction, and hand-OA event counts were limited, so the corresponding analyses should be interpreted as descriptive. The Olink overlap was small, multiple-testing control used FDR rather than family-wise error correction, and hand-OA event counts were limited, so the corresponding analyses should be interpreted as descriptive. In addition, metabolic responses documented in short-term inpatient feeding trials only capture acute energy balance changes and cannot be used to infer long-term causal relationships between UPF consumption and chronic OA risk over decades of follow-up. Specifically, subsequent research should recruit larger multi-omics cohorts with adequate Olink proteomic samples, adopt longitudinal repeated exposure and biomarker measurements to enable formal causal mediation analysis and multi-omics joint integration modeling, and carry out dietary intervention trials to verify whether reducing UPF intake alleviates joint metabolic dysfunction and OA progression.

Conclusion

In this prospective multi-omics analysis of women, higher ultra-processed food intake was associated with modestly increased risks of incident any osteoarthritis and knee/hip osteoarthritis, with a borderline but directionally consistent signal for knee osteoarthritis. Higher UPF intake was associated with a broad NMR metabolomic phenotype covering systemic inflammation, disturbed fatty acid metabolism and abnormal HDL lipoprotein profiles; the PCA-derived UPF-NMR signature independently predicted elevated OA risk without adiposity adjustment, providing tentative correlational evidence pointing to overlapping adiposity-related metabolic pathways underlying the UPF-OA association. Olink proteomic results were merely exploratory supplementary data with no stable mechanistic signals. The findings are hypothesis-generating and add joint health to the spectrum of female-relevant chronic outcomes for which limiting UPF may be relevant, while identifying candidate circulating biological features—most prominently GlycA and fatty-acid composition—for follow-up. Notably, all circulating metabolic markers were measured only once at baseline alongside dietary assessment in a cross-sectional fashion; without longitudinal repeated biomarker data, we cannot establish causal temporal sequences, and the adiposity-related metabolic pathways suggested by NMR signatures remain exploratory correlational hypotheses rather than validated causal mechanisms. Future research includes four targeted directions. First, replicate UPF-OA associations in large independent female cohorts with repeated refined UPF records and joint replacement endpoints, and conduct menopause-stratified analyses to interpret the borderline knee OA association. Second, longitudinal datasets with serial diet, adiposity, biomarker and OA measurements are required for temporal causal mediation or Mendelian randomisation to quantify adiposity and inflammation mediation effects. Third, expanded Olink proteomic cohorts with adequate OA events and combined lipid-proteomic profiling are needed to identify stable protein mediators of joint injury. Fourth, preclinical cartilage/synovium models can validate whether altered fatty acids and GlycA induce chondrocyte dysfunction; short-term UPF-restricted dietary trials are also needed to track dynamic metabolic and inflammatory biomarker shifts.

Funding Statement

This work was supported by Ningbo Top Medical and Health Research Program [grant numbers 2022020506].

Data Sharing Statement

The data is publicly available on reasonable request.

Ethics Approval and Informed Consent

This study has approval from the North West Multi-Centre Research Ethics Committee, and all participants provided written informed consent.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors have declared that no competing interests exist.