Method

Database and participants

The study population was derived from the Osteoarthritis Initiative (OAI) (http://www.oai.ucsf.edu), a prospective, multicenter cohort study designed to include individuals who were either healthy with or without risk factors for knee OA or had symptomatic knee OA. The analytic cohort comprised a mixed OAI population, including participants with mild established knee OA as well as individuals at increased risk of developing knee OA, and therefore was not restricted to participants with established symptomatic or radiographic disease.

The dataset included the WOMAC questionnaire [16] with three subcategories: pain (0–20 range), stiffness (0–8 range), and functional limitation (0–68 range). Physical performance and strength were assessed by performing the Repeated Chair Stand Test (CST) (ability to complete 5 CST - yes/no) at baseline and at the 48-month follow-up.

MRI-derived TMV data were available for 2818 participants at baseline and 48-month follow-up. Participants were excluded if baseline body mass index (BMI) data were missing (n = 7), if they had rheumatoid arthritis at baseline or developed rheumatoid arthritis during the study period (n = 218), or if severe comorbidities (cancer or stroke) were present (n = 161). Participants with inflammatory arthritis were excluded irrespective of the timing of diagnosis to ensure that the study population represented degenerative rather than inflammatory joint disease. In total, 1715 participants were included in the analysis (see Fig. 1).

Fig. 1: Flowchart of study inclusion. V00 = baseline, V06 = follow-up after 48 months, BMI = body mass index, WORMS = Whole-Organ Magnetic Resonance Imaging Score.

Ethical approval

Written informed consent was obtained from all participants, and the study adhered to the regulations outlined in the Health Insurance Portability and Accountability Act (HIPAA). Ethical approval was granted by the institutional review boards of all participating centers.

MR imaging

MRI scans of the right knee and thigh were conducted at baseline and at the 48-month follow-up according to the standardized OAI imaging protocol using four identical 3T scanners (Siemens Magnetom Trio; Siemens Healthcare, Erlangen, Germany) equipped with transmit-receive coils (USA Instruments, Aurora, OH) across four imaging centers. TMV was assessed using T1-weighted axial sequences with the following imaging parameters: repetition time (TR)/echo time (TE) 600/10 ms, slice thickness of 5 mm, and a craniocaudal coverage of approximately 7.5 cm, beginning 10 cm proximal to the right femoral epiphysis [19,20]. Further details regarding the analyzed MRI sequences can be found in the OAI protocol [21].

Deep learning-based automated muscle analysis

At both timepoints, 15 axial slices of the right thigh were acquired, with the central 7 slices selected as the primary unit of observation for analysis to minimize the effect of image quality degradation in the edge slices. The fully automated artificial intelligence-enabled MR muscle segmentation pipeline consisted of a fine-tuned YOLO (You Only Look Once) model for bounding box detection and a fine-tuned box-prompted SAM (Segment Anything Model) segmentation model for the quadriceps, hamstrings, adductors, gracilis, and sartorius muscles [22]. The pipeline was trained and validated using manual segmentations on all slices of 109 thigh MRI volumes from three musculoskeletal radiologists, and tested on a hold out cohort with manual central slice segmentations on 246 from two musculoskeletal radiologists. In the hold out test cohort, the mean Dice similarity coefficient was 0.96 for both the quadriceps and hamstrings. The coefficients of variation were 0.3% for the quadriceps and 2.5% for the hamstrings, with mean biases of −1.7% and +1.2%, respectively, compared with the manual reference standard [22]. Muscle sub-volumes (in cm3) for all included muscle groups were calculated at baseline and at the 48-month follow-up to quantify changes in muscle volume over time.

Structural knee joint analysis

All knee MRI datasets at baseline and at 48-months follow-up were separately anonymized before being evaluated. OA-related features were assessed using the modified semiquantitative WORMS grading system [7,23,24]. WORMS grading was performed using the same standardized reading protocol and readers as previously described [25], for which good-to-excellent intra- and inter-reader reproducibility has been reported. Meniscal defects in the medial and lateral meniscus (anterior, body, posterior) were graded on a scale of 0–4: 0 = intact, 1 = intrasubstance signal abnormalities, 2 = nondisplaced tear, 3 = displaced or complex tear, 4 = maceration. Ligament abnormalities (anterior/posterior cruciate ligaments, medial/lateral collateral ligaments) and tendon pathologies (patellar and popliteal tendons) were graded on a 0–4 scale: 0 = no pathology, 1 = signal abnormalities around the tendon/ligament, 2 = signal abnormalities within the tendon/ligament, 3 = partial tear, and 4 = complete tear. Cartilage defects were graded on a 0–6 scale, and bone marrow edema-like lesions (BMELL) on a 0–3 scale across six distinct knee regions: patella, trochlea, medial/lateral femur, and medial/lateral tibia. Tissue-specific WORMS sum scores for cartilage, meniscal abnormalities, BMELL, tendon abnormalities, and ligament abnormalities were calculated by summing the regional WORMS scores across all evaluated knee subregions. Higher scores indicate a greater overall burden of structural abnormalities within the respective tissue, with all subregions contributing equally to the corresponding tissue-specific sum score. This approach has been used previously to summarize the overall burden of tissue-specific structural abnormalities [26]. Only the right knee and right thigh were analyzed because WORMS assessments were available only for the right knee in the imaging dataset used for this study.

Statistical analysis

Statistical analyses were performed using R version 4.5.2 (R Core Team, 2025) using RStudio version 2025.09.2 + 418 (Posit Software PBC, Boston, MA, USA). TMV was analyzed at baseline and 48-month follow-up, with percentage change in muscle volume calculated for each participant. For the primary analyses, participants were classified into three groups based on quartiles of percentage TMV change: muscle loss (−29.8 to < −7.2%), comprising individuals in the lowest quartile (Q1); stable/minimal muscle change (−7.2% to <2.4%), including participants in the second and third quartiles (Q2–Q3); and muscle gain (2.4–23.7%), comprising individuals in the highest quartile (Q4). This categorization was chosen to compare participants with the greatest muscle loss and greatest muscle gain while using the two middle quartiles as a reference group representing relatively small changes in muscle volume. As a sensitivity analysis, percentage TMV change was additionally analyzed as a continuous variable using the same multivariable regression models. To minimize the influence of extreme values, participants with percentage TMV changes exceeding the upper bound (Q3 + 2 × interquartile range (IQR)) or falling below the lower bound (Q1 − 2 × IQR) were classified as outliers and excluded from the analyses. Baseline characteristics, including age, sex, BMI, muscle volume, mean WORMS sum scores and mean WOMAC scores, were summarized using descriptive statistics for the overall study cohort, as well as for each muscle change group individually. To assess the potential for selection bias resulting from missing WORMS scores, baseline characteristics were compared between participants with complete and missing WORMS data. Continuous variables were compared using Welch's t-test and categorical variables using Pearson's χ2 test. Standardized mean differences (SMDs) were additionally calculated to quantify the magnitude of between-group differences independent of sample size. SMDs <0.10 were considered to indicate negligible imbalance, 0.10–0.20 small imbalance, and >0.20 moderate imbalance.

Due to the large number of outcome parameters, analyses were prespecified and grouped into primary and exploratory outcomes [27,28]. Primary outcomes comprised WORMS BMELL, cartilage, and ligament score, WOMAC pain subscore, because these measures represent key structural and symptomatic features of KOA that have been consistently associated with disease progression in previous studies [27,28]. Exploratory outcomes included WORMS menisci and tendon score, WOMAC stiffness, and disability subscores and CST performance, for which associations with TMV are less well established. To investigate the hypotheses that participants with muscle loss experience greater progression in structural joint damage (evaluated by WORMS scores) and in functional outcomes (evaluated by WOMAC scores) over 48 months compared to those with muscle gain, multivariable linear regression models were conducted. These models were adjusted for sex, age, race and BMI (at baseline), as these variables were considered potential confounders based on previous literature describing differences in body composition and knee osteoarthritis across demographic groups [29]. The independent variables included changes in muscle volume/muscle group, while the outcome variables were changes in WORMS sum scores for cartilage defects, meniscal defects, ligamentous pathologies, tendon pathologies and BMELL across the six knee regions (patella, trochlea, medial femur, lateral femur, medial tibia and lateral tibia) as well as changes in the three WOMAC subscores (pain, stiffness and disability) between baseline and 48 months. Model assumptions for the linear regression analyses were assessed using standard diagnostic procedures (values are shown for the BMELL model as a representative example). Linearity was evaluated using residual-versus-fitted plots. Normality of residuals was assessed using normal Q–Q plots and the Shapiro–Wilk test (p < 0.001). Homoscedasticity was assessed using the studentized Breusch–Pagan test (BP = 13.89, p = 0.085) and Levene's test (F = 1.58, p = 0.207). Multicollinearity was evaluated using variance inflation factors (all VIFs ≤1.03), influential observations using Cook's distance and leverage values (maximum Cook's distance = 0.104; maximum leverage = 0.094), and independence of residuals using the Durbin–Watson statistic (1.98, p = 0.322). Although the Shapiro–Wilk test indicated a statistically significant deviation from normality, visual inspection of the Q–Q plots suggested only minor departures from normality, which were considered acceptable given the large sample size.

Because the primary hypotheses focused on the association between muscle gain (relative to the stable muscle group) and four prespecified primary outcomes (WORMS BMELL, cartilage, ligament, and WOMAC pain), p-values for these primary regression coefficients were additionally adjusted using the Benjamini–Hochberg false discovery rate (FDR) procedure. Exploratory outcomes were not adjusted for multiplicity and should be interpreted as hypothesis-generating. As a sensitivity analysis, all multivariable regression models evaluating WORMS outcomes were additionally adjusted for baseline physical activity (PASE score) to assess potential confounding by habitual physical activity.

To assess the association between changes in CST performance and changes in muscle volume, an ordinal regression analysis, adjusted for sex, age, race, and BMI, was conducted (independent variable: muscle group; dependent variable: change in CST, categorized into three ordered levels: deterioration (−1; able to complete the CST at baseline but unable at 48 months), stability (0; no change in completion status, i.e., able at both visits or unable at both visits), or improvement (1; unable to complete the CST at baseline but able at 48 months)).