Abstract
Objective Although both osteoarthritis (OA) and rheumatoid arthritis (RA) can necessitate total knee arthroplasty (TKA), their mechanisms and radiographic patterns of joint space narrowing (JSN) differ. Deep learning (DL) enables compartment-specific measurement of minimum joint space width (mJSW). This study compared JSN patterns between patients with RA and OA undergoing TKA and evaluated the associations between mJSW and RA disease activity.
Methods In this retrospective study, 409 patients with RA undergoing TKA (2000-2021) were age- and sex-matched with OA patients. A validated DL model quantified medial and lateral mJSW from anteroposterior knee radiographs at 2 timepoints: early (> 2 years before TKA) and late (≤ 2 years before TKA). The rate of JSN was calculated in 302 patients with paired radiographs. Mixed-effects models adjusted for BMI and alignment. RA disease activity, serology, and medications were recorded.
Results RA demonstrated narrower lateral mJSW than OA at the late timepoint (5.6 vs 7.0 mm; P < 0.001), with comparable medial values. RA exhibited uniform bicompartmental JSN, whereas OA showed medial narrowing. RA had a faster mean lateral ΔmJSW (0.11 mm/year; P = 0.01), whereas OA had a faster mean medial ΔmJSW (0.21 mm/year; P < 0.001). Longer RA duration and higher inflammatory markers were negatively associated with lateral mJSW. Seronegative patients showed faster lateral ΔmJSW in the surgical knee than seropositive patients (0.12 vs 0.06 mm/year; P = 0.03).
Conclusion Patients with RA demonstrated diffuse, symmetric cartilage loss that contrasted with medial-predominant narrowing in patients with OA. Automated DL-based measurement enables scalable, compartment-specific assessment of joint degeneration patterns.
Osteoarthritis (OA) and rheumatoid arthritis (RA) are among the most common underlying diagnoses leading to total knee arthroplasty (TKA).1 Despite both culminating in significant joint destruction, the pathogenesis of articular cartilage damage occurs through fundamentally different mechanisms.2-5 In OA, joint degeneration is largely biomechanical and often localized, with preferential involvement of the central region of the tibial plateau.2 These patterns correlate strongly with age, BMI, and limb alignment.2,3 A varus lower extremity tends to preferentially overload the medial compartment, whereas a valgus deformity leads to more lateral compartment degeneration. In contrast to localized joint overload based on the mechanical axis, RA is a systemic autoimmune condition characterized by diffuse synovial inflammation with inflammatory cellular infiltration that promotes articular cartilage degradation in a relatively nondiscriminatory pattern,6 leading to global bicompartmental cartilage loss, bone erosions, and relatively limited osteophyte formation.4 However, previous literature suggests that patients with RA are increasingly presenting with OA-like radiographic features at the time of TKA.4,5 This trend may be attributable to currently available disease-modifying antirheumatic drugs (DMARDs), including biologics, which are believed to more effectively suppress synovitis and mitigate inflammatory joint destruction than previous therapies.7,8 As a result, the radiographic presentation of joint space narrowing (JSN) in RA is becoming more heterogeneous and potentially more biomechanical in character, suggesting an association between RA disease activity and the pattern of articular cartilage loss.
A continuous but indirect measure of articular cartilage thickness is the minimum joint space width (mJSW), which is a validated metric for monitoring degenerative joint disease progression and has been used in trials of DMARDs.9 Unlike previous methods that relied on cadaveric data, 3D imaging, or semiquantitative scoring of JSN, automated mJSW measurement enables a granular and scalable approach for comparative analysis of cartilage loss patterns in OA vs RA.2,3,5
Recent advances in artificial intelligence, and particularly deep learning (DL), have introduced new opportunities to quantify and analyze radiographic features of joint degeneration with improved precision and reproducibility.5,10-13 An open-source DL model that automatically and accurately measures mJSW in both the medial and lateral compartments of the knee is readily available to assist in the exploration of JSN patterns in these patient populations.10
This study aimed to compare preoperative mJSW measurements in patients with OA and RA undergoing TKA. Specifically, the objectives were to evaluate whether (1) lateral and medial mJSW differed between groups at the time of TKA, (2) preoperative JSN patterns differed between groups, and (3) knee mJSW before TKA was associated with RA disease activity and severity. It was hypothesized that there would be a significant difference in mJSW progression between the OA and RA populations, with RA patients exhibiting more bicompartmental disease. Further, it was expected that higher RA disease activity or longer RA duration would be associated with both lower mJSW measurements and greater change in mJSW leading up to TKA compared with patients with lower disease activity or shorter duration.
METHODS
Patient selection. This retrospective case-control study included patients who underwent unilateral primary TKA at a single tertiary referral center (ie, Mayo Clinic) between January 1, 2000, and December 31, 2021. The case group comprised patients with a clinical diagnosis of RA, whereas the control group consisted of patients diagnosed with primary OA. Inclusion criteria were (1) a confirmed diagnosis of RA or OA by a board-certified rheumatologist or orthopedic surgeon, (2) treatment with primary unilateral TKA, and (3) availability of a preoperative anteroposterior (AP) knee radiograph of the operative limb within 2 years prior to TKA. A control group of patients with OA was randomly selected and matched to the patients with RA by age and sex in a 1:1 ratio.
Demographic and clinical variables, including age, sex, laterality, and BMI, were collected for both the RA and control groups through a review of institutional records to allow for group-level comparisons.
RA-specific disease characteristics were manually abstracted from the medical record at the visit closest to the TKA date within the preceding 2 years. Disease activity measures included the Clinical Disease Activity Index (CDAI), Disease Activity Score in 28 joints (DAS28), and Health Assessment Questionnaire (HAQ). The 4-variable DAS28 includes swollen and tender joint counts based on 28 joints, patient global assessment (PtGA), and either C-reactive protein (CRP) level or erythrocyte sedimentation rate (ESR). Alternative 3-variable DAS28 calculations that do not require a PtGA were used when the PtGA was unavailable. The overall DAS28 was obtained using the following hierarchy of scoring formulas: DAS28-CRP4, DAS28-ESR4, DAS28-CRP3, and DAS28-ESR3. Serologic status (ie, rheumatoid factor [RF] and anticyclic citrullinated protein [anti-CCP] antibodies), medication use (including DMARDs and glucocorticoids) at the time of TKA, and RA disease duration at the time of TKA were also extracted. Medications included those used to treat RA prior to any temporary discontinuations for the surgery.
Hip-knee-ankle alignment (HKAA) was calculated using the hip-knee-ankle angle method, defined as the angle formed between the mechanical axis of the femur (line from the femoral head center to the intercondylar notch) and the mechanical axis of the tibia (line from the midtibial interspinous point to the midtibial plafond). This angle was measured using a previously validated DL-based algorithm that calculates mechanical alignment from short-leg AP radiographs. Although full-length radiographs are traditionally used to measure HKAA, this automated method enables alignment estimation from standard knee radiographs and has demonstrated strong agreement (ie, mean difference < 2°) with conventional measurements.
mJSW measurements. Standing preoperative AP radiographs were retrieved from an institutional imaging registry.14 Radiographs were obtained using our institution’s standardized weight-bearing protocol for patients being considered for TKA. All study patients had an AP knee radiograph of the operative limb within 2 years prior to TKA, which was designated as “late.” When multiple radiographs were available during the 2 years before TKA, the radiograph closest to surgery was preferred; if bilateral radiographs were available, those were prioritized. Where available, the earliest available AP radiograph of the operative limb was also assessed. This earlier radiograph was included if it was > 2 years before TKA and was designated as “early.” The following were measured: (1) medial and lateral compartment mJSW at these 2 defined preoperative timepoints, namely early (> 2 years prior to TKA) and late (≤ 2 years prior to TKA); and (2) the rate of JSN or change in mJSW (ΔmJSW) over time. Measurements were obtained for both the surgical knee and the contralateral (nonsurgical) knee when visualized. ΔmJSW was calculated separately for each compartment and knee using the following formula:
mJSW measurements were obtained using a previously validated DL algorithm developed by our laboratory.10 Briefly, the algorithm consists of a segmentation model that identifies key anatomical landmarks, specifically the medial and lateral tibial plateaus and femoral condyles. Based on the segmentation output, a computer vision algorithm locates the most distal point on each femoral condyle and identifies the shortest distance to the corresponding tibial plateau to yield mJSW in millimeters. The algorithm is designed to automatically exclude radiographs of post-TKA knees by identifying the presence of prosthetic components during segmentation. Validation of the algorithm demonstrated a mean (SD) absolute error of 0.85 (1.20) mm compared with manual measurements, with an intra-class correlation coefficient (ICC) of 0.92, indicating excellent agreement. To maintain a fully automated measurement pipeline and avoid manual measurement variability, radiographs for which the DL algorithm failed to generate measurements (< 0.5%) were excluded from further analysis.
Data analyses. Medians with IQRs were used to summarize continuous variables, and counts (percentages) were used to summarize categorical variables. Chi-square and rank-sum tests were used to compare baseline characteristics between groups. Spearman correlation coefficients were used to assess the relationship between RA characteristics and mJSW values. Linear mixed-effects models were used to assess group differences in mJSW after adjusting for measurement time period, compartment, BMI, and HKAA, using the lmerTest package in R (R Foundation for Statistical Computing). A 2-sided P < 0.05 was considered statistically significant for all analyses. Analyses were performed using SAS version 9.4 (SAS Institute) and R version 4.4.1.
RESULTS
This study included 409 patients with RA and 409 controls with OA. Demographic characteristics of the study groups are summarized in Table 1. Clinical data specific to patients with RA, including disease activity scores (eg, DAS28 or CDAI), serologic status (RF and anti-CCP), medication use (including DMARDs and glucocorticoids), and disease duration, are reported in Table 2.
Overall demographics of patients with RA and OA controls.
RA characteristics and medications collected at the visit nearest to TKA.
mJSW at the early and late timepoints. The median (IQR) time between late imaging and TKA was similar in the RA and OA groups (34 [3-87] vs 22 [3-81] days; P = 0.17; Table 3). All patients had mJSW measurements for the surgical knee at the late timepoint, but a small number were missing measurements in either the lateral or the medial compartment. More than 80% of patients had mJSW measurements for both knees at the late timepoint. Figure 1 shows back-to-back histograms of mJSW stratified by knee compartment at both timepoints for surgical knees. For RA vs OA surgical knees, RA patients demonstrated narrower lateral joint space (5.6 [4.2-7.4] vs 7.0 [5.6-8.0] mm; P < 0.001; Table 3). There was no difference in medial compartment mJSW of surgical knees between the groups (2.8 [1.6-5.1] vs 2.6 [1.7-4.4] mm; P = 0.16). Similar trends were observed in the nonsurgical knees at the late timepoint and in the measurements taken at the early timepoint for both knees. Consistently, the mixed-effects model performed on the 147 patients with RA and 155 OA controls who had both an early and late radiograph identified a significant 2-way interaction (P < 0.001) between compartment and group (RA vs OA), further highlighting differences between groups in the lateral compartment at both the late timepoint (RA-OA difference −0.99 [95% CI −1.28 to −0.70]) and the early timepoint (−0.56 [95% CI −1.02 to −0.09]). The 3-way interaction between time, compartment, and diagnosis was not statistically significant (P = 0.10), indicating a lack of evidence that the rate of change between early and late timepoints across compartments differed between the RA and OA groups.
Back-to-back histograms of mJSW stratified by knee compartment for the early (left) and late (right) timepoints for surgical knees. mJSW: minimum joint space width; OA: osteoarthritis; RA: rheumatoid arthritis.
Summary of minimum joint space width measurements at early (≥ 2 years before TKA) and late (≤ 2 years before TKA) imaging timepoints in patients with RA (n = 409) and OA (n = 409).
When BMI and HKAA were included in the mixed-effects model, BMI showed a borderline association with overall mJSW (P = 0.05), whereas HKAA was not associated with mJSW (P = 0.56). Adjustment for BMI and HKAA did not affect the estimated differences between the RA and OA groups.
Rate of change in mJSW before TKA. We fit a mixed-effects model on surgical knee data with ΔmJSW as the outcome, and knee compartment (medial or lateral), diagnosis group, and their interaction as predictors. Tests for the RA vs OA difference demonstrated a faster rate of JSN in the lateral compartment in patients with RA compared with patients with OA (0.11 mm/year, 95% CI 0.02 to 0.20; P = 0.01). In contrast, patients with OA experienced marginally faster JSN in the medial compartment (RA-OA difference −0.08 mm/year, 95% CI −0.16 to 0.01; P = 0.09). Figure 2 depicts the difference in JSN patterns between groups, with relatively uniform declines observed in RA, whereas more medial-predominant degeneration occurred in patients with OA. Patients with OA experienced more rapid JSN in the medial compartment compared with the lateral compartment (lateral-medial difference −0.21 mm/year; 95% CI −0.26 to −0.17; P < 0.001), whereas patients with RA experienced similar rates of JSN between compartments (lateral-medial difference −0.03 mm/year, 95% CI −0.07 to 0.02; P = 0.21; Figure 3). Adjustment for BMI (P = 0.07) and HKAA (P = 0.62) did not affect these results.
Mean mJSW in the lateral and medial compartments at the early and late timepoints with 95% CIs in patients with RA and OA. mJSW: minimum joint space width; OA: osteoarthritis; RA: rheumatoid arthritis; TKA: total knee arthroplasty.
Rate of change per year in mJSW in the lateral and medial compartments with 95% CIs in patients with RA and OA. mJSW: minimum joint space width; OA: osteoarthritis; RA: rheumatoid arthritis.
RA disease characteristics and mJSW. The association between RA disease characteristics and mJSW measurements was assessed at the late preoperative timepoint using Spearman correlation coefficients. RA duration was significantly negatively correlated with lateral compartment mJSW at TKA for both the surgical and nonsurgical knees, such that longer RA duration was associated with lower mJSW in both knees (ρ = −0.14, P = 0.01, n = 361; and ρ = −0.19, P < 0.001, n = 330, respectively). Similarly, ESR was negatively correlated with lateral compartment mJSW at TKA in both knees (surgical: ρ = −0.15, P = 0.02, n = 244; nonsurgical: ρ = −0.18, P = 0.01, n = 228). CRP level was negatively correlated with lateral compartment mJSW at TKA in the surgical knee (ρ = −0.15, P = 0.02, n = 253), but that in the nonsurgical knee did not reach statistical significance (ρ = −0.10, P = 0.13, n = 238). Correlations of a similar magnitude were observed for CDAI, but these did not reach statistical significance, likely reflecting the limited sample size (surgical: ρ = −0.19, P = 0.10, n = 76; nonsurgical: ρ = −0.19, P = 0.12, n = 70). There were no other significant correlations between any of the RA characteristics in Table 2 and lateral or medial compartment mJSW of the surgical or nonsurgical knees.
Correlations between RA disease duration and seropositivity and ΔmJSW in both compartments of both knees were examined. There was no correlation between RA disease duration and ΔmJSW in the lateral compartment of either knee or in the medial compartment of the surgical knee (ρ = −0.11, P = 0.20, n = 141). However, a significant negative correlation was observed in the medial compartment of the nonsurgical knee (ρ = −0.19, P = 0.04, n = 121), indicating that longer RA disease duration corresponded to a smaller ΔmJSW. A negative correlation was observed between RF and/or anti-CCP antibody positivity and ΔmJSW in the lateral and medial compartments of the surgical knee (ρ = −0.20, P = 0.03, n = 121; and ρ = −0.17, P = 0.06, n = 131, respectively), whereas no association was observed in either compartment of the nonsurgical knee (ρ = 0.00, P = 0.96, n = 112; and ρ = 0.03, P = 0.74, n = 114, respectively). Consistently, patients who were seronegative demonstrated a faster rate of lateral JSN in the surgical knee compared with seropositive patients (median [IQR] 0.12 [0.03-0.31] vs 0.06 [−0.08 to 0.18], P = 0.03), with a similar trend in the medial compartment of the surgical knee (median [IQR] 0.18 [0.003-0.33] vs 0.10 [−0.08-0.26], P = 0.06).
DISCUSSION
This study used a DL algorithm to examine mJSW degeneration patterns in patients with RA and OA undergoing TKA. Patients with RA demonstrated a more symmetric bicompartmental degeneration pattern, likely reflecting systemic inflammatory joint injury. Conversely, patients with OA demonstrated medial narrowing consistent with biomechanical stress due to alignment-driven preferential loading, resulting in compartment-specific degeneration. At the late preoperative timepoint, longer RA duration and higher ESR in both knees as well as higher CRP in the surgical knee were associated with smaller lateral mJSW. In the longitudinal analysis of the present study, seronegative status was associated with faster JSN in the surgical knee in both the lateral and medial compartments. Additionally, longer RA duration correlated with smaller ΔmJSW in the medial compartment of the nonsurgical knee. The faster rate of JSN observed among patients with seronegative RA may reflect variability in clinical management and/or timing of diagnosis rather than intensity of inflammation alone.15-17 Adjusted analyses suggested only a marginal influence of BMI on JSN, suggesting that body habitus may be modestly associated with cartilage degeneration in both RA and OA.
The finding of symmetric bicompartmental narrowing in RA aligns with data from previous automated imaging studies. Platten et al used DL to identify early, symmetric hand cartilage loss within the metacarpal joint in RA before erosions appeared on radiographs.18 These patterns were absent in OA, where medial narrowing reflected biomechanical stress, a distinction also captured by our model. Higher inflammatory marker levels (ie, ESR and CRP) and longer disease duration were correlated with smaller lateral mJSW. This finding aligns with prior work from Koga et al, which showed that baseline CRP and time-integrated DAS28-ESR over 1 year predicted radiographic progression, particularly in established RA.19 The results of this study reinforce that OA progression is biomechanically mediated, whereas RA progression remains predominantly inflammatory in nature.
However, contemporary understanding of OA recognizes that the disease is not purely biomechanical. Low-grade synovial inflammation, metabolic factors, and systemic inflammatory mediators are increasingly implicated in OA pathogenesis and may contribute to cartilage degeneration even in the absence of severe malalignment.20 These factors may partially explain the variability in compartmental degeneration observed among patients with OA and the overlap in radiographic patterns between advanced OA and RA. Nonetheless, in end-stage disease requiring TKA, mechanical loading patterns and alignment remain dominant drivers of compartment-specific cartilage loss,21 which is consistent with the medial-predominant narrowing observed in the OA cohort. These findings emphasize the importance of diagnosis-specific radiographic evaluations, especially as modern DMARDs increasingly obscure radiographic distinctions between RA and OA. Automated mJSW analysis using DL may facilitate earlier detection of atypical patterns to support more precise prognostication, more efficient specialty referrals, and more appropriate patient selection for invasive procedures such as injections or knee arthroplasty. Further, the association between RA disease activity and JSN suggests that rheumatologic metrics may contribute to orthopedic risk stratification. Collectively, these insights support a more disease-specific approach to imaging, clinical evaluation, and surgical planning for patients undergoing TKA.
This study has several limitations. This retrospective study focused on patients undergoing TKA, which may introduce selection bias and limit generalizability to earlier disease stages. Radiographs were obtained within a limited preoperative window, which restricted the ability to observe long-term progression. Further, fewer than half of the patients had imaging at both early and late timepoints, limiting the power of the ΔmJSW analyses and associated conclusions. Additionally, mJSW measurements were derived from standard standing AP knee radiographs rather than the semiflexed knee positioning protocol typically used for joint space analysis, which may introduce measurement variability and may limit direct comparison with prior JSW studies. The DL algorithm demonstrated a mean absolute error of 0.85 mm relative to manual measurements, which is nontrivial compared with annual JSW changes. However, this level of error is equivalent to 4 pixels on a standard radiograph, which typically has a resolution of approximately 2056 × 2056 and is comparable to interobserver variability reported for manual mJSW measurements.22
In conclusion, this study compared patterns of JSN in patients undergoing TKA for RA vs OA using automated mJSW measurements. Patients with RA exhibited symmetric, bicompartmental cartilage loss, whereas patients with OA demonstrated predominantly medial JSN, largely attributed to biomechanical loading. Further, RA disease activity measures were weakly correlated with mJSW measurements. These findings highlight disease-specific degeneration patterns in RA and OA, and automated mJSW analysis provides an efficient and reproducible method for large-scale radiographic assessment.
Footnotes
CONTRIBUTIONS
MMG, MK, SS: data curation, writing – original draft, writing – review and editing; KLM, CCW, JMD, GAM: conceptualization, data curation, interpretation, writing – review and editing; CAH, EJA: formal analysis, interpretation, writing – review and editing; CSC: conceptualization, formal analysis, interpretation, funding, writing – review and editing.
FUNDING
This work was partially funded by the Mark E. and Mary A. Davis Initiative in Rheumatoid Arthritis Research.
COMPETING INTERESTS
JMD has royalty rights from intellectual property/patent licensed to Rheumasense and receives grant/research support from Pfizer. All other authors have no financial relationships or conflicts of interest to disclose.
ETHICS AND PATIENT CONSENT
This study was approved by the institutional review board (Mayo Clinic IRB no. 23-011538), and all patients provided consent for the use of their clinical data and images for research.
- Accepted for publication April 18, 2026.
- Copyright © 2026 by the Journal of Rheumatology
This is an Open Access article, which permits use, distribution, and reproduction, without modification, provided the original article is correctly cited and is not used for commercial purposes.
