Association of latent class trajectories of frailty with incident symptomatic knee osteoarthritis in Chinese postmenopausal women: a nationwide cohort study from China

Abstract

Both frailty and symptomatic knee osteoarthritis (SKOA) are prevalent among older adults, with a greater incidence noted in women. The present study aimed to classify the trajectory types pertaining to the Frailty Index (FI) in postmenopausal women, explore the longitudinal correlational link among different FI trajectory profiles as well as the likelihood of developing SKOA, and evaluate the predictive value of FI trajectories for SKOA along with the robustness of this association. A long-term prospective follow-up investigation was conducted on the basis of the CHARLS cohort of postmenopausal women. Latent Class Growth Modeling (LCGM) was applied to classify the FI trajectory patterns across five waves from 2011 to 2020. Differences in the cumulative incidence risk of SKOA among trajectory subgroups were evaluated using Kaplan-Meier survival curves paired with the log-rank statistical test. The strength of this association was assessed using Cox proportional hazards regression models. Model performance across FI trajectories, as well as baseline frailty, was assessed and contrasted using ROC curves, Net Reclassification Improvement (NRI), and Integrated Discrimination Improvement (IDI). E-value estimation, subgroup analyses, and sensitivity analyses were performed to gauge the robustness of this association. For this investigation, data pertaining to 4636 postmenopausal women within the CHARLS study cohort were analyzed, with a systematic comparison of the predictive capacity of FI trajectories and baseline frailty status in assessing SKOA. The results showed that both FI trajectories and baseline frailty status showed statistically significant correlations with SKOA. Specifically, in contrast to the low-baseline slight-increase trajectory subgroup (Class 1), the high-baseline consistently rising trajectory subgroup (Class 3) exhibited the greatest likelihood of developing SKOA (HR: 3.86, 95% CI: 3.19–4.68), followed by the moderate-baseline gentle-increase trajectory group (Class 2) (HR: 3.35, 95% CI: 2.98–3.77). In comparison with baseline frailty, FI trajectories exhibited stronger predictive capability, as evaluated by the ROC curve, NRI, and IDI. Additional analyses, such as E-value calculations, subgroup analyses, and multiple sensitivity analyses, validated the stability of these relationships. The findings from this investigation highlight the significance of frailty and FI trajectories in evaluating SKOA among postmenopausal women.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-53896-4.

Introduction

Osteoarthritis (OA) is the most prevalent chronic joint disorder globally, affecting approximately 607 million people in 2021—a 45% increase since 1990¹. In China, the age-adjusted prevalence reaches 7,030 per 100,000, with projections indicating over 64 million knee OA cases by 2050^2–5. OA disproportionately affects women, particularly those aged 50–54 years, likely reflecting hormonal transitions and differential joint loading⁴.

Frailty is a multisystem syndrome characterized by diminished physiological reserve and increased vulnerability to stressors⁶. Among community-dwelling adults aged ≥ 65 years, frailty affects 10–24% depending on the assessment tool^7,8. Population-based studies have documented consistent links between frailty and arthritis through shared pathways including chronic inflammation, physical inactivity, and sarcopenia^9–12. Specifically, symptomatic knee OA (SKOA) has been associated with 60–92% higher frailty prevalence and 66% increased risk of frailty development¹³, while a recent cross-sectional analysis of China Health and Retirement Longitudinal Study (CHARLS) data reported a strong, non-linear association between Frailty Index and knee OA prevalence (OR: 3.63)¹⁴. However, all of these investigations treated frailty as a static or single-time-point exposure, precluding understanding of how longitudinal frailty progression patterns differentially influence disease risk.

Latent Class Growth Modeling (LCGM) offers a methodologically rigorous approach to address this limitation by identifying homogeneous subgroups with distinct longitudinal trajectories^15,16. This framework captures individual heterogeneity in progression patterns that static assessments cannot detect and has been successfully applied to model trajectories of cognition, depression, and cardiovascular risk factors^17,18. Despite its potential, limited evidence exists on the application of frailty trajectory analysis to musculoskeletal outcomes.

Building on this identified gap, the present study used CHARLS data to (i) identify distinct Frailty Index (FI) trajectory patterns among postmenopausal Chinese women over nine years of follow-up; (ii) quantify the longitudinal association between these trajectories and incident SKOA risk using Cox proportional hazards regression; and (iii) evaluate whether dynamic FI trajectories provide incremental predictive value over baseline frailty status. By focusing on postmenopausal women—a high-risk population at the intersection of estrogen decline, sarcopenia, and joint vulnerability—this study aims to advance risk stratification and early intervention strategies for SKOA.

Methods

Study population

Our investigation employs a longitudinal analysis that draws a dataset from CHARLS. CHARLS functions as a nationwide representative survey developed to examine the social, economic, and health conditions of middle-aged and older populations (aged ≥ 45 years) in China¹⁹. Its nationwide baseline assessment (wave 1) took place in 2011, enrolling 17,708 study subjects, with follow-ups completed in waves 2–5 in 2013, 2015, 2018, and 2020, respectively. This study was conducted in line with the STROCSS 2025 guidelines²⁰.

Eligible study participants were required to satisfy the following inclusion criteria: aged ≥ 45 years at baseline, postmenopausal, free of SKOA diagnosis at baseline, with complete baseline FI data, and holding valid FI evaluations in no fewer than two follow-up survey waves. The exclusion criteria included unknown menopausal status, missing information on SKOA diagnosis, missing baseline FI data, or FI data from fewer than two follow-up waves. Ultimately, 4,636 participants were included in this analysis. The comprehensive process for selecting the study subjects is illustrated in Fig. 1.

Fig. 1.

Flowchart.

Definition of symptomatic knee osteoarthritis

The SKOA diagnosis in this study followed the criteria established in prior research^21,22, specifically characterized as arthritis diagnosed by a medical practitioner coupled with knee pain. In detail, this classification was determined based on responses to the following three questions: (1) “Have you received a diagnosis of arthritis or rheumatism from a medical professional?”; and (2) “Do you frequently experience physical discomfort?”; and (3) If the response to query (2) was “yes,” the participant was further asked, “In which parts of your body is the pain located?“. Study subjects who responded affirmatively to both of the initial queries and indicated that the pain site included the “knee(s)” were identified as having SKOA.

Definition of frailty

Frailty status was evaluated using FI derived from the cumulative deficit model, a framework proposed by Searle et al.²³. The FI employed in the present investigation consisted of 36 items encompassing questionnaire-derived indicators across several dimensions: basic activities of daily living, instrumental activities for daily functioning, chronic disease conditions, depressive symptomology, and self-perceived health status²³. Comprehensive descriptions of the items are presented in Table S1. FI functions as a continuous metric spanning 0 to 1, where elevated scores signify increased frailty severity. In line with prior studies, the present investigation categorized frailty as an FI ≥ 0.25²⁴.

Assessment of covariates

The covariates collected in this study spanned multiple dimensions, including sociodemographic characteristics, health-related behaviors, medication use, anthropometric measures, and laboratory indicators. The social and demographic attributes included age, age at menopause, ethnicity, residential type, educational attainment, marital status, employment status, disability, health insurance, and life satisfaction. Health-related behaviors included smoking and drinking history, and daily sleep duration. Medication use included a history of use of antihypertensive agents, lipid-lowering medications, glucose-lowering treatments, and arthritis-associated pharmaceuticals. The anthropometric measurements included systolic blood pressure (SBP) and diastolic blood pressure (DBP). Laboratory tests included high-density lipoprotein cholesterol (HDL-C), serum creatinine (Scr), blood urea nitrogen (BUN), low-density lipoprotein cholesterol (LDL-C), triglycerides (TG), total cholesterol (TC), uric acid (UA), and C-reactive protein (CRP).

Latent class growth modeling

The LCGM was employed to classify individual subgroups with comparable FI change trajectories over five waves (2011, 2013, 2015, 2018, 2020)¹⁵. This model assumes that individuals belong to a fixed number of latent classes, with members within each class following a similar developmental trajectory. To identify the optimal class count, we fitted models with 1 to 5 latent classes using the lcmm package in R (version 4.5.1) and evaluated them against the following criteria: (1) lower scores for the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and sample size-adjusted BIC (SABIC) signal a superior model fit; (2) an entropy score exceeding 0.8 signals strong class separation; (3) each latent class should account for at least 5% of the total sample and reach statistical significance (P < 0.05); (4) successful model convergence; and (5) clear clinical interpretability.

Both linear and quadratic functional forms for the FI trajectories were compared. While the quadratic specification yielded substantially lower AIC and BIC values (ΔAIC ≈ 1,980; ΔBIC ≈ 1,961), the class membership proportions were essentially identical to those of the linear model. Given the comparable class structure, greater parsimony, and easier clinical communication of linear trajectories, the linear specification was retained for the primary analyses. The robustness of this choice was confirmed by sensitivity analyses using quadratic-derived trajectories.

The final 3-class linear model was selected as the optimal solution: it achieved the lowest BIC and SABIC among models satisfying all criteria, with an entropy of 0.854 (exceeding the 0.80 threshold), all class proportions greater than 5% (59.4%, 33.5%, and 7.1%), and clear clinical interpretability corresponding to low-baseline slight-increase, moderate-baseline gradual-increase, and high-baseline steady-increase trajectories. The average posterior probabilities were 0.899, 0.930, and 0.952 for Classes 1, 2, and 3, respectively, which all exceeded the 0.70 threshold recommended for high classification accuracy. All models converged successfully.

Statistical analysis

Continuous variables are reported as mean ± standard deviation (SD), while categorical variables are summarized as frequencies (percentages). Standardized mean differences (SMD) were calculated as Cohen’s d comparing Class 3 (highest-risk) vs. Class 1 (reference). Absolute values are reported to facilitate interpretation of effect magnitude regardless of direction. The LCGM was employed to identify potential FI trajectories. Missing covariate data were interpolated using the random forest method in R. Five imputed datasets were generated (m = 5). Convergence was assessed by comparing the mean and SD of the interpolated values from the last two iterations; the relative variation for all variables was less than 0.5%, indicating good convergence. The missing rates for each covariate are shown in Table S2.

Cox proportional hazards regression models were used to examine the relationship between distinct FI trajectories and the risk of SKOA occurrence. The proportional hazards assumption was tested using Schoenfeld residual test. All covariates satisfied this assumption; the FI trajectory class variable showed significant non-proportionality (P < 0.001), consistent with the expected pattern that high-frailty individuals develop events disproportionately early in follow-up. Kaplan-Meier curves were generated to visually illustrate the cumulative incidence of SKOA, while the log-rank test was employed to compare differences across groups.

Model performance was assessed using Receiver Operating Characteristic (ROC) curves, Net Reclassification Improvement (NRI), and Integrated Discrimination Improvement (IDI) to compare the predictive capacity of FI trajectories with baseline frailty status. To assess multicollinearity, the variance inflation factor (VIF) was calculated for all the covariates. Furthermore, E-value analysis was performed to assess the reliability of the results, calculating the lowest magnitude of unmeasured confounders needed to potentially reverse the detected relationship^25,26.

To explore heterogeneity in effects, subgroup analyses were conducted based on age, age at menopause, employment status, residential location, marital status, life satisfaction, disability status, smoking history, history of arthritis medication use, and history of antihypertensive medication use. Given the presence of multiple comparisons, we used the Benjamini–Hochberg false discovery rate method for correction.

Several sensitivity analyses were implemented to confirm the robustness of the findings: (1) complete-case analysis to validate the imputation strategy; (2) exclusion of participants with asymptomatic knee OA at baseline to ensure specificity for symptomatic disease; (3) exclusion of those reporting only knee pain without arthritis diagnosis to address potential reverse causation; (4) additional adjustment for BMI to assess residual confounding by adiposity; (5) use of quadratic-derived trajectories to evaluate robustness to the assumed linear functional form; and (6) to better capture individual heterogeneity, trajectory analysis and association testing were also conducted using the Latent Class Growth Mixture Model (LCGMM), which allows for within-class trajectory variation by assuming individuals belong to multiple latent classes with different developmental trajectories²⁷. The LCGMM relaxes the assumption that all individuals within a class follow an identical trajectory, permitting individual-level deviation from the class mean. To compare the statistical findings, we estimated the association between LCGMM-derived trajectories and SKOA risk using the same Cox regression framework as the primary LCGM analysis. The hazard ratios were compared directly to assess consistency. All analyses were conducted using the R software (version 4.5.1).

Results

FI Trajectories

Based on the analysis using the LCGM and the following model selection criteria including the AIC, BIC, entropy, and a minimum class proportion exceeding 5% (Table S3), participants were grouped into three separate FI trajectories (Fig. 2): low-baseline slight-increase (Class 1, 59.4%), moderate-baseline gradual-increase (Class 2, 33.5%), and high-baseline steady-increase (Class 3, 7.1%). Specifically, participants in trajectory Class 1 had a comparatively low FI level at baseline and showed a slight increase during the follow-up period. Those in Class 2 trajectory had a moderate FI level at baseline, displaying a gradual, mild upward trend during follow-up. Participants in Class 3 started from a high baseline FI level and demonstrated a sustained, gentle increase throughout the observation period.

Fig. 2.

FI trajectories derived from the LCGM results.

Baseline characteristics

Table 1 presents the baseline characteristics of the 4,636 participants stratified by FI trajectory class. Overall, individuals in the high-baseline steady-increase trajectory (Class 3, n = 328) differed markedly from those in the low-baseline slight-increase trajectory (Class 1, n = 2,755). Notable imbalances (|SMD| > 0.10) were observed for age (|SMD| = 0.93), employment status (0.66), antihypertensive medication use (0.67), arthritis medication use (0.57), lipid-lowering medication use (0.50), disability (0.48), SBP (0.44), and CRP (0.25). Class 3 participants were generally older, more likely to be unemployed, had lower educational attainment, higher prevalence of disability and medication use, and exhibited elevated inflammatory markers. Conversely, minimal imbalance was observed for age at menopause, drinking history, TC, and LDL-C (all |SMD| < 0.10). Detailed laboratory biochemical profiles are presented in Supplementary Table 4. The distribution of FI from 2011 to 2020 is shown in Figure S1.

Table 1.

Baseline characteristics stratified by latent class.

Characteristic	Overall	Class_1	Class_2	Class_3	|SMD|(Class 3 vs. 1)	P
N	4,636	2,755	1,553	328
Age, years	61.13 (8.50)	59.09 (7.56)	63.57 (8.76)	66.81 (9.08)	0.93	< 0.001
Age of menopause, years	49.32 (4.10)	49.26 (3.90)	49.35 (4.36)	49.65 (4.40)	0.09	0.254
Nation, n (%)					0.12	0.075
Other	312 (6.73)	178 (6.46)	102 (6.57)	32 (9.76)
Han ethnicity	4324 (93.27)	2577 (93.54)	1451 (93.43)	296 (90.24)
Education, n (%)					0.28	< 0.001
Below high school	4312 (93.01)	2513 (91.22)	1479 (95.24)	320 (97.56)
High school	223 (4.81)	169 (6.13)	49 (3.16)	5 (1.52)
Above high school	101 (2.18)	73 (2.65)	25 (1.61)	3 (0.91)
Work, n (%)					0.66	< 0.001
No	2108 (45.47)	1059 (38.44)	821 (52.87)	228 (69.51)
Yes	2528 (54.53)	1696 (61.56)	732 (47.13)	100 (30.49)
Residence, n (%)					0.16	< 0.001
City	1845 (39.80)	1159 (42.07)	574 (36.96)	112 (34.15)
Rural	2791 (60.20)	1596 (57.93)	979 (63.04)	216 (65.85)
Marital status, n (%)					0.34	< 0.001
Married	3616 (78.00)	2256 (81.89)	1139 (73.34)	221 (67.38)
Not married or no partner	1020 (22.00)	499 (18.11)	414 (26.66)	107 (32.62)
Health insurance, n (%)					0.15	0.003
No	307 (6.62)	157 (5.70)	118 (7.60)	32 (9.76)
Yes	4329 (93.38)	2598 (94.30)	1435 (92.40)	296 (90.24)
Smoke, n (%)					0.11	0.009
No	4261 (91.91)	2560 (92.92)	1406 (90.53)	295 (89.94)
Yes	375 (8.09)	195 (7.08)	147 (9.47)	33 (10.06)
Drinking, n (%)					0.02	0.721
No	3974 (85.72)	2371 (86.06)	1323 (85.19)	280 (85.37)
Yes	662 (14.28)	384 (13.94)	230 (14.81)	48 (14.63)
Life satisfaction, n (%)					0.27	< 0.001
Unsatisfied	627 (13.52)	309 (11.22)	249 (16.03)	69 (21.04)
Satisfied	4009 (86.48)	2446 (88.78)	1304 (83.97)	259 (78.96)
Disability, n (%)					0.48	< 0.001
No	3907 (84.28)	2441 (88.60)	1238 (79.72)	228 (69.51)
Yes	729 (15.72)	314 (11.40)	315 (20.28)	100 (30.49)
Arthritis medication, n (%)					0.57	< 0.001
No	3908 (84.30)	2490 (90.38)	1194 (76.88)	224 (68.29)
Yes	728 (15.70)	265 (9.62)	359 (23.12)	104 (31.71)
Antidiabetic medication, n (%)					0.37	< 0.001
No	4444 (95.86)	2693 (97.75)	1461 (94.08)	290 (88.41)
Yes	192 (4.14)	62 (2.25)	92 (5.92)	38 (11.59)
Lipid-lowering medication, n (%)					0.5	< 0.001
No	4383 (94.54)	2677 (97.17)	1435 (92.40)	271 (82.62)
Yes	253 (5.46)	78 (2.83)	118 (7.60)	57 (17.38)
Antihypertensive medication, n (%)					0.67	< 0.001
No	3600 (77.65)	2345 (85.12)	1071 (68.96)	184 (56.10)
Yes	1036 (22.35)	410 (14.88)	482 (31.04)	144 (43.90)
Sleep duration, h	6.24 (1.91)	6.49 (1.77)	5.97 (1.99)	5.44 (2.26)	0.52	< 0.001
DBP, mmHg	75.12 (11.81)	74.51 (11.53)	75.97 (11.88)	76.25 (13.37)	0.14	< 0.001
SBP, mmHg	131.25 (22.29)	128.24 (20.71)	135.04 (23.29)	138.60 (25.65)	0.44	< 0.001

SMD, standardized mean difference. Absolute SMD > 0.10 suggests potentially meaningful baseline imbalance.

Association between frailty and symptomatic knee osteoarthritis

Analysis using Kaplan-Meier curves showed that, compared to the Class 1 group, the survival curves for the Class 2 and Class 3 groups declined more rapidly, indicating that higher FI trajectories were associated with an increased cumulative incidence of SKOA. The log-rank test demonstrated a notable difference in the cumulative incidence across the three subgroups (Fig. 3).

Fig. 3.

Kaplan-Meier curves.

Table 2 presents the link between FI trajectories and SKOA risk across the multiple adjusted models. Using Class 1 as the reference group in the fully adjusted Model 3, the risk of SKOA was notably higher in the Class 2 subgroup (HR = 3.35, 95% CI: 2.98–3.77) and rose even more in the Class 3 subgroup (HR = 3.86, 95% CI: 3.19–4.68). Furthermore, individuals frail at baseline faced a 56% higher risk of SKOA occurrence than their non-frail peers (HR = 1.56, 95% CI: 1.37–1.77). Collinearity diagnostics indicated that the VIF for all covariates was below 5 (Table S5), suggesting no significant multicollinearity issue in the models.

Table 2.

Regression analysis on the link between frailty and SKOA risk.

	Event (n%)	Crude Model HR (95%CI)	P value	Model 1 HR (95%CI)	P value	Model 2 HR (95%CI)	P value	Model 3 HR (95%CI)	P value
Frailty_baseline	1544 (33.3)	2.04 (1.82, 2.29)	< 0.001	2.09 (1.86, 2.35)	< 0.001	1.93 (1.70, 2.18)	< 0.001	1.56 (1.37, 1.77)	< 0.001
FI trajectory
Class_1	555 (20.1)	Ref.	-	Ref.	-	Ref.	-	Ref.
Class_2	810 (52.2)	3.36 (3.02, 3.75)	< 0.001	3.77 (3.37, 4.22)	< 0.001	3.69 (3.29, 4.14)	< 0.001	3.35 (2.98, 3.77)	< 0.001
Class_3	179 (54.6)	3.84 (3.24, 4.54)	< 0.001	4.61 (3.87, 5.49)	< 0.001	4.45 (3.71, 5.34)	< 0.001	3.86 (3.19, 4.68)	< 0.001

Model 1: adjusted for age and age at menopause.

Model 2: adjusted for Model 1 + education level, marital status, ethnicity, residence, employment status, disability, health insurance, life satisfaction, sleep duration, smoking history, and drinking history.

Model 3 was adjusted for Model 2 + SBP, DBP, HDL-C, LDL-C, TC, TG, UA, CRP, BUN, Scr, and medication use history.

Assessment of FI trajectories’ predictive performance

Based on the ROC curve analysis (Fig. 4) and comparison of the associated Area Under the Curve (AUC) values (Table S6), the FI trajectory model exhibited good discriminative capacity, with an AUC of 0.712 (95% CI: 0.694–0.731). This performance outperformed the model that relied solely on the baseline frailty status (AUC = 0.590).

Fig. 4.

ROC curves demonstrating the stronger predictive ability of FI trajectories in predicting SKOA risk.

To evaluate the incremental predictive value of FI trajectories over traditional risk factors, the NRI and IDI were computed, and the results are detailed in Table 3. With Model 3 serving as the baseline prediction model, adding FI trajectory information resulted in a notable enhancement in risk reclassification performance (NRI = 0.287, P < 0.001), and this improvement was greater than that achieved by adding baseline frailty status. Regarding discrimination ability, the addition of FI trajectories resulted in significant enhancement (IDI = 0.025, P < 0.001), whereas the contribution of baseline frailty status was relatively limited (IDI = 0.004, P = 0.013). This indicates that the dynamic trajectory pattern of frailty has additional predictive value compared to a single baseline frailty assessment.

Table 3.

Incremental predictive value of FI trajectories versus baseline frailty.

	C-statistic (95%CI)	P value	Continuous NRI (95%CI)	P value	IDI (95%CI)	P value
Basic model	0.685 (0.663, 0.707)	< 0.001	Ref		Ref
Basic model + FI trajectory	0.772 (0.753, 0.791)	< 0.001	0.287 (0.210, 0.346)	< 0.001	0.025 (0.016, 0.038)	< 0.001
Basic model + Frailty_baseline	0.702 (0.681, 0.724)	< 0.001	0.072 (0.009, 0.149)	0.020	0.004 (0.001, 0.010)	0.013

The basic model was adjusted for age, age at menopause, education, marital status, ethnicity, residence, employment status, disability, health insurance, life satisfaction, sleep duration, smoking history, drinking history, SBP, DBP, HDL-C, LDL-C, TC, TG, UA, CRP, BUN, Scr, and medication use history. NRI, net reclassification improvement; IDI, integrated discrimination improvement. NRI > 0.20 and IDI > 0.01 are generally considered indicative of clinically meaningful incremental predictive value.

E-value analysis

To assess the potential impact of unmeasured confounding factors on the identified associations, E-value analysis was conducted. For the different FI trajectories, the E-value for the comparison between the high-baseline steady-increase group (Class 3) and low-baseline slight-increase group (Class 1) was 4.44. For baseline frailty status, the E-value for the comparison of frail and non-frail subgroups was 2.05. These E-values are close to or surpass the widely referenced robustness benchmark of 1.82²⁶, indicating that the links between frailty and SKOA identified in this study are fairly reliable.

Subgroup analysis

To investigate whether demographic characteristics, lifestyle, and health status might alter the link between FI trajectories and SKOA risk, subgroup analyses and interaction tests were performed (Fig. 5). Among all prespecified subgroups (such as age, age at menopause, employment status, and residence), using Class 1 as the reference, both Class 2 and Class 3 were consistently linked to a notably elevated risk of SKOA.

Fig. 5.

Subgroup analysis.

Interaction analysis (Fig. 5) showed that employment status and history of arthritis medication use significantly altered these associations (P for interaction < 0.05). However, the interaction P-values for other factors (including age, age at menopause, and marital status) were all above 0.05, indicating that these variables did not significantly alter the association between FI trajectories and SKOA risk. After FDR adjustment, no statistically significant interactions were found in any subgroup (Adjusted P > 0.05). This result indicates that the main effect of FI trajectories on SKOA risk remained strong and consistent across all subgroups, supporting the robustness of the primary association.

Sensitivity analysis

To confirm the reliability of the longitudinal link between FI trajectories and SKOA, several sensitivity analyses were performed (Tables S7–S10). The results consistently showed that the core association remained stable across different analytical strategies. Specifically, a significant positive correlation between higher FI trajectories and SKOA risk persisted after: (1) excluding samples with missing covariates (n = 2,110); (2) removing samples with (asymptomatic) arthritis at baseline (n = 3,159); (3) excluding samples reporting only knee pain at baseline (n = 4,329); and (4) adjusting for BMI. Sensitivity analysis of quadratic-derived trajectories yielded hazard ratios that were largely consistent with the linear results (Table S11); the goodness-of-fit indices are shown in Table S12.

Furthermore, the FI trajectories were refitted using an LCGMM to capture within-individual variation. The dynamic characteristics of these trajectories are shown in Figure S2. Analyses of the LCGMM-derived trajectories also showed that higher FI trajectories were linked to a greater risk of SKOA (Table S13). The LCGMM-derived trajectories yielded hazard ratios nearly identical to the primary LCGM (Class 3 vs. 1: 3.79 vs. 3.86; Class 2 vs. 1: 3.31 vs. 3.35), with overlapping confidence intervals, confirming that allowing within-class trajectory variation did not alter the substantive conclusions (Table S14). The goodness-of-fit indices for the LCGMM (including AIC, BIC, SABIC, and entropy) are listed in Table S15.

Discussion

Using a nationally representative sample, this study carried out a 9-year longitudinal follow-up of 4,636 postmenopausal Chinese women. We examined the link between frailty level and SKOA risk from both static and dynamic viewpoints. Analyses of baseline frailty status showed that baseline-frail participants (FI ≥ 0.25) faced a notably greater risk of SKOA when matched against non-frail peers (FI < 0.25). Using a dynamic trajectory model, we identified three distinct FI evolution trajectories: low-baseline slight-increase, moderate-baseline gradual-increase, and high-baseline steady-increase. The results showed that relative to the low-baseline slight-increase trajectory, the high-baseline steady-increase trajectory was notably linked to a greater risk of SKOA onset. In terms of predictive performance, the FI trajectory model outperformed the single baseline frailty assessment in ROC, NRI, and IDI.

FI trajectories have been linked to the risk of several age-related conditions such as stroke, chronic kidney disease, cognitive decline, and cancer^28–31. Epidemiological and mechanistic evidence converges to support an association between frailty and KOA. Zhu et al.¹⁴ recently documented a strong nonlinear association between the FI and the prevalence of SKOA in the CHARLS cohort (odds ratio 3.63, 95% CI: 3.48–3.78). This finding is supported by longitudinal evidence: Waghorn et al.³² demonstrated that frailty increases the risk of osteoarthritis onset, progression, and adverse outcomes; Tian et al.³³ demonstrated a bidirectional association between frailty and KOA in Chinese adults, particularly women; The Multicenter Osteoarthritis Study reported that SKOA was associated with a 92% increase in frailty prevalence and a 66% increase in frailty risk¹³; advanced KOA was associated with worsening frailty status in KNHANES participants³⁴. In addition to prevalence, frailty can predict the 9-year trajectory of knee pain, with greater frailty increasing the likelihood of progression to severe pain³⁵. These associations have biological significance through shared mechanisms such as chronic low-grade inflammation (CRP, IL-6, TNF-α)^12,36, joint instability caused by sarcopenia³⁷, and estrogen-deficiency-mediated cartilage degeneration and bone loss^38–40, collectively forming a vicious cycle of musculoskeletal self-deterioration^11,41.

However, all existing studies have treated frailty as a static or single-point-in-time exposure. As Zhu et al.¹⁴ acknowledge, future research should model frailty as “a time-varying exposure rather than a static endpoint.” Our study directly addresses this call by identifying three distinct FI trajectories over a nine-year period. We demonstrate that frailty progression patterns provide independent prognostic information for the onset of SKOA beyond baseline status, with dynamic trajectories significantly outperforming static assessments (AUC 0.712 vs. 0.590; continuous NRI 0.287 vs. 0.072). The AUC increase from 0.590 to 0.712 represents a 20.7% relative improvement in discriminative accuracy, moving from poor-fair to moderate-good discrimination. The continuous NRI of 0.287 indicates that nearly 29% more participants were correctly reassigned to appropriate risk categories, exceeding the 0.10–0.20 threshold typically considered clinically actionable. For every 100 postmenopausal women screened, approximately 29 would receive more accurate risk stratification. Incorporating dynamic FI trajectory assessment into routine health evaluations could improve identification of high-risk individuals for early, targeted interventions—such as resistance training, nutritional optimization, or anti-inflammatory management—while women in stable low-risk trajectories could be spared unnecessary overtreatment, supporting more efficient allocation of limited healthcare resources.

In constructing the FI, this study employed an index comprising 36 items, which provides more comprehensive coverage than the 32-item indices commonly used in previous research³³. The newly included indicators include dyslipidemia, liver conditions, kidney disorders, and digestive system illnesses, broadening the scope of the comorbidity assessment. For functional assessment, this evaluation includes a three-tier mobility test (running/jogging 1 km, 1 km walk, 100-meter walk) along with urinary/bowel incontinence challenges, leading to a more thorough assessment of functional deficits. Generally, the more variables an FI incorporates, the more accurate its assessments are²³.

Furthermore, frailty and pre-frailty conditions tend to be more prevalent in older adults, women, and people with a greater number of comorbidities^42,43. This study focused on a high-risk population of postmenopausal women, providing an in-depth exploration of the association between FI trajectories and SKOA, thereby addressing a gap in longitudinal research for this group. The sharp decline in estrogen during menopause not only accelerates muscle loss and affects joint stability³⁷ but also inhibits osteoblast activity, promotes osteoclast proliferation leading to bone loss, reduces cartilage matrix synthesis capacity, and exacerbates articular cartilage degeneration⁴⁰. These physiological changes collectively form a chain reaction pathway: “estrogen decline → musculoskeletal functional decline → frailty → joint degeneration.” Focusing on this population facilitates a clearer analysis of the dynamic mediating role of FI trajectories in this pathological process.

While our findings suggest that FI trajectories may serve as a potentially useful tool for SKOA risk stratification, external validation in independent cohorts and prospective implementation studies are needed before clinical translation. Until such evidence is available, FI trajectories should be regarded as a research-oriented risk indicator rather than a validated clinical decision-making tool.

The strengths of this study include the utilization of a large, nationally representative sample and a 9-year longitudinal follow-up, which helped minimize selection bias and boost the reliability of the findings. This study thoroughly examined the link between frailty and SKOA by combining static (baseline) and dynamic (trajectory) perspectives. Methodologically, this study used rigorous analytical approaches, including multivariable Cox regression, stratified subgroup analyses, NRI, IDI, and E-value estimations. Additionally, the FI trajectories were refitted and validated using LCGMM. These multidimensional analyses collectively enhance the reliability and clinical reference value of the study results.

This study has a few limitations. First, SKOA diagnoses relied on self-reporting without objective validation from the imaging results. Similarly, the FI assessment was based on self-reported data, which could lead to measurement errors. Second, the observed hazard ratios of up to 3.86 warrant careful consideration of potential residual confounding. Several factors support the robustness of these findings: (i) the magnitude is consistent with prior evidence—frailty has been associated with a 3.6-fold increased risk of prevalent knee OA¹⁴ and approximately doubled the risk of incident knee OA in longitudinal analyses³⁵; (ii) sensitivity analyses excluding baseline knee pain and adjusting for BMI yielded consistent estimates; and (iii) E-value calculations indicate that an unmeasured confounder would need to be associated with both FI trajectories and SKOA by a risk ratio of at least 4.44 to nullify the primary finding—a magnitude exceeding typical effect sizes for known risk factors. Nevertheless, data on physical activity (55.6% missing), menopause duration, and hormone replacement therapy were unavailable, and residual confounding from these factors cannot be entirely excluded. Furthermore, although the longitudinal design and FI trajectory analysis helped clarify temporal relationships, the results cannot directly infer causality. Reverse causation remains a possibility that merits careful consideration. Early undiagnosed knee osteoarthritis—manifesting as subclinical pain, stiffness, or reduced mobility—could theoretically contribute to physical inactivity and accelerating frailty progression, creating a bidirectional association. While our longitudinal design and baseline exclusion of diagnosed SKOA provide temporal support for the directionality from frailty to SKOA, we acknowledge that subclinical knee pathology prior to formal diagnosis could partially explain the observed associations. Finally, missing data management relied on the missing-at-random assumption, with multiple imputations performed via the random forest method. If the data did not follow a completely random missing pattern, this might have influenced the conclusions of the study.

Conclusion

This study is the first to identify three distinct frailty evolution trajectories in a Chinese population cohort, confirming that moderate- and high-baseline FI trajectories are linked to a markedly higher risk of developing SKOA. Relative to a single baseline frailty evaluation, dynamic FI trajectories showed stronger predictive ability for SKOA.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Mengjie Zhao: Methodology, Software, Visualization, Writing – original draft. Xujie Wang: Conceptualization, Software, Writing – original draft. Mengxuan Li: Conceptualization, Validation, Resources. Yufei Wu: Investigation, Visualization, Data curation. Zirong Li: Conceptualization, Software, Formal analysis. Fang Lu: Writing – review & editing, Methodology, Project administration. Qiuyan Li: Writing – review & editing, Validation, Supervision.

Data availability

Research data are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent

The CHARLS survey received approval from Peking University’s Ethical Review Committee. All participants submitted written informed consent.